Académique Documents
Professionnel Documents
Culture Documents
ELECTRICAL IMPEDANCE
TOMOGRAPHY
Methods, History and Applications
Edited by
David S Holder
Department of Medical Physics and Bioengineering
University College London
London
Series Editors:
C G Orton, Karmanos Cancer Institute and Wayne State University, Detroit,
USA
J H Nagel, Institute for Biomedical Engineering, University Stuttgart,
Germany
J G Webster, University of Wisconsin-Madison, USA
Commissioning Editor: John Navas
Editorial Assistant: Leah Fielding
Production Editor: Simon Laurenson
Production Control: Sarah Plenty
Cover Design: Victoria Le Billon
Marketing: Louise Higham, Kerry Hollins and Ben Thomas
Published by Institute of Physics Publishing, wholly owned by The Institute
of Physics, London
Institute of Physics Publishing, Dirac House, Temple Back, Bristol BS1 6BE, UK
US Oce: Institute of Physics Publishing, The Public Ledger Building, Suite
929, 150 South Independence Mall West, Philadelphia, PA 19106, USA
Typeset by Academic+Technical, Bristol
Printed in the UK by MPG Books Ltd, Bodmin, Cornwall
The Series in medical Physics and Biomedical Engineering is the ocial book
series of the International Federation for Medical and Biological Engineering
(IFMBE) and the International Organization for Medical Physics (IOMP).
IFMBE
The International Federation for Medical and Biological Engineering
(IFMBE) was established in 1959 to provide medical and biological engineering with a vehicle for international collaboration in research and practice of
the profession. The Federation has a long history of encouraging and
promoting international co-operation and collaboration in the use of science
and engineering for improving health and quality of life.
The IFMBE is an organization with membership of national and transnational societies and an International Academy. At present there are 48 national
members and two transnational members representing a total membership in
excess of 30 000 world wide. An observer category is provided to give personal
status to groups or organizations considering formal aliation. The International Academy includes individuals who have been recognized by the
IFMBE for their outstanding contributions to biomedical engineering.
Objectives
The objectives of the International Federation for Medical and Biological
Engineering are scientic, technological, literary, and educational. Within
the eld of medical, clinical and biological engineering its aims are to
encourage research and the application of knowledge, and to disseminate
information and promote collaboration.
In pursuit of these aims the Federation engages in the following activities:
sponsorship of national and international meetings, publication of ocial
journals, co-operation with other societies and organizations, appointment
of commissions on special problems, awarding of prizes and distinctions,
establishment of professional standards and ethics within the eld, as well as
other activities which in the opinion of the General Assembly or the Administrative Council would further the cause of medical, clinical or biological
engineering. It promotes the formation of regional, national, international
or specialized societies, groups or boards, the coordination of bibliographic
or informational services and the improvement of standards in terminology,
equipment, methods and safety practices, and the delivery of health care.
The Federation works to promote improved communication and understanding in the world community of engineering, medicine and biology.
Activities
The IFMBE publishes the journal Medical and Biological Engineering and
Computing which includes a special section on Cellular Engineering. The
IFMBE News, published electronically, keeps the members informed of the
developments in the Federation. In cooperation with its regional conferences,
Activities
Ocial publications of the IOMP are Physiological Measurement, Physics
in medicine and Biology and the Series in Medical Physics and Biomedical
Engineering, all published by the Institute of Physics Publishing. The
IOMP publishes a bulletin Medical Physics World twice a year.
Two council meetings and one General Assembly are held every three
years at the ICMP. These conferences are normally held in collaboration
with the IFMBE to for the World Congress on Medical Physics and Biomedical Engineering. The IOMP also sponsors occasional international
conferences, workshops and courses.
Information on the activities of the IOMP are found on its web site at
http://www.iomp.org/.
Contents
LIST OF CONTRIBUTORS
INTRODUCTION
PART 1
ALGORITHMS
1.5.
1.6.
1.7.
1.8.
Contents
1.9.
1.10.
1.11.
1.12.
1.13.
PART 2
39
40
42
42
44
45
47
50
52
52
53
54
56
HARDWARE
65
2. EIT INSTRUMENTATION
Gary J Saulnier
2.1.
2.2.
2.3.
2.4.
2.5.
2.6.
Introduction
EIT system architecture
Signal generation
2.3.1. Waveform synthesis
2.3.2. Current sources
2.3.3. Driving the current source
2.3.4. Multiplexers
2.3.5. Current source and compensation circuits
2.3.6. Cable shielding
2.3.7. Voltage sources
Voltage measurement
2.4.1. Dierential versus single-ended
2.4.2. Common-mode voltage feedback
2.4.3. Synchronous voltage measurement
2.4.4. Noise performance
2.4.5. Sampling requirements
Example EIT systems
2.5.1. Single-source systems
2.5.2. Multiple-source systems
Discussion and conclusion
References
33
36
67
67
67
69
69
70
79
80
80
86
87
88
88
90
90
93
94
95
96
98
101
103
Contents
PART 3
APPLICATIONS
3.3.
3.4.
3.5.
3.6.
General introduction
Equipment
3.2.1. Sheeld mark 1 system
3.2.2. Newer systems
Cardiac imaging
3.3.1. Introduction
3.3.2. Electrode positioning
3.3.3. EIT and stroke volume
3.3.4. Right ventricular diastolic function
3.3.5. Summary
Pulmonary perfusion measurements
3.4.1. Introduction
3.4.2. Pulmonary perfusion defects
3.4.3. Pathological changes of the pulmonary vascular
bed
3.4.4. Summary
Assessment of regional lung function
3.5.1. Introduction
3.5.2. Experimental and clinical studies
3.5.3. Future directions
General summary and future perspectives
References
4.3.
Introduction
Physiological basis of EIT of brain function
4.2.1. Bioimpedance of brain and changes during activity
or pathological conditions
4.2.2. Eect of coverings of the brain when recording
EIT with scalp electrodes
EIT systems developed for brain imaging
4.3.1. Hardware
4.3.2. Reconstruction algorithms for EIT of brain
function
4.3.3. Development of tanks for testing of EIT systems
xi
105
107
107
107
107
109
110
110
110
112
112
113
113
113
114
114
117
117
117
118
122
123
123
127
127
129
129
136
137
137
141
146
xii
Contents
4.4.
4.5.
4.6.
4.7.
4.8.
5.2.
5.3.
6.3.
148
148
149
154
155
156
157
159
160
161
167
167
167
168
169
171
171
172
172
173
173
174
178
181
182
186
186
188
188
188
188
189
190
Contents
6.4.
6.5.
6.7.
6.8.
6.9.
6.10.
PART 4
191
191
191
191
192
193
194
194
195
196
198
198
198
200
201
201
202
203
205
207
Hyperthermia
EIT imaging of intra-pelvic venous congestion
Other possible applications
References
207
208
209
209
NEW DIRECTIONS
211
xiii
Introduction
The MIT signal
Coils and screening
Signal demodulation
Cancellation of the primary signal
213
213
214
215
218
218
xiv
Contents
8.6.
8.7.
8.8.
8.9.
8.10.
8.11.
8.12.
9.4.
9.5.
Introduction
Problem denition
Forward problem and numerical techniques
9.3.1. Forward problem in MREIT using recessed
electrodes
9.3.2. Eects of recessed electrodes and lead wires
9.3.3. Computation of voltage V and current density J
9.3.4. Computation of magnetic ux density B using the
BiotSavart law
9.3.5. Computation of magnetic ux density B using
FEM
9.3.6. Computation of current density J from magnetic
ux density
9.3.7. Numerical examples of 3D forward solver
Measurement techniques in MREIT
9.4.1. Review of MRCDI techniques
9.4.2. How to measure one component of B
9.4.3. Measurements of all three components of B by
subject rotations
9.4.4. Computation of current density image J in
MRCDI
9.4.5. Data processing
9.4.6. Signal-to-noise ratio (SNR) in magnetic ux and
current density image
Image reconstruction algorithms
9.5.1. Requirements in data collection methods for
uniqueness
220
220
222
225
228
230
230
231
232
233
233
239
239
242
244
244
245
246
247
249
249
249
256
256
257
258
258
259
259
260
261
Contents
9.6.
9.7.
9.8.
10.3.
10.4.
10.5.
Introduction
Data acquisition
10.2.1. Electrical resistance tomography
10.2.2. Electrical capacitance tomography (ECT)
10.2.3. Electromagnetic tomography (EMT)
10.2.4. Electrical impedance tomography
10.2.5. Intrinsically safe systems
10.2.6. Summary of data acquisition systems
Data processing
Industrial applications of electrical tomography
10.4.1. Application of electrical resistance tomography
technology to pharmaceutical processes
10.4.2. Imaging the ow prole of molten steel through
a submerged pouring nozzle
10.4.3. The application of electrical resistance tomography
to a large volume production pressure lter
10.4.4. A novel tomographic ow analysis system
10.4.5. Application of electrical capacitance tomography
for measurement of gas/solids ow characteristics
in a pneumatic conveying system
10.4.6. Imaging wet gas separation process by capacitance
tomography
Summary
Acknowledgements
References
xv
262
263
265
266
266
270
273
274
274
280
288
289
291
295
295
298
299
302
303
305
306
307
307
312
312
316
318
326
330
335
338
340
340
xvi
Contents
11.3.
11.4.
11.5.
11.6.
11.7.
11.8.
11.9.
Beginnings
Making images: applied potential tomography
11.2.1. Back-projection
11.2.2. Normalizing the data
Dierential imaging
Collecting data
11.4.1. The Mark 1
11.4.2. The Mark 2
11.4.3. Limitations
Multifrequency images
11.5.1. The Mark 3
11.5.2. Marks 3a and 3b
The third dimension
Clinical studies
What we have learned
11.8.1. High resolution imaging is not possible
11.8.2. Making reliable in vivo measurements is dicult
11.8.3. Humans are 3D
11.8.4. What do we need to do?
11.8.5. Some suggestions
The future of medical EIT
Appendix. The Sheeld algorithm revisited
References
Early developments
Reconstruction algorithms
Hardware
Applied currents
Optimal currents
Static in vivo images with non-circular boundary and
optimal currents
3D
In vivo applications
348
348
349
350
351
352
355
356
356
358
359
359
361
363
364
365
365
366
366
367
367
368
368
371
373
386
388
388
391
395
398
399
400
400
401
Contents
13.9.
13.10.
13.11.
13.12.
Paying for it
People
Meetings
Concluding remarks
Complete Bibliography
Selected Abstracts
A.4.
A.5.
xvii
403
404
405
406
407
410
411
411
416
418
418
419
420
421
422
B.3.
B.4.
B.5.
Historical perspective
EIT instrumentation
B.2.1. Individual impedance measurements
B.2.2. Data collection
B.2.3. Electrodes
B.2.4. Setting up and calibrating measurements
B.2.5. Data collection strategies
EIT image reconstruction
B.3.1. Back-projection
B.3.2. Sensitivity matrix approaches
B.3.3. Other developments in algorithms
Clinical applications
B.4.1. Performance of EIT systems
B.4.2. Potential clinical applications
Current developments
References
423
425
425
428
431
431
432
435
435
435
439
439
439
442
445
446
List of contributors
D C Barber
Medical Imaging and Medical Physics, Royal Hallamshire Hospital, Glossop
Road, Sheeld S10 2JF, UK
A Borsic
School of Mathematics, The University of Manchester, PO Box 88, Manchester
M60 1QD, UK
D F Evans
Centre for Adult and Paediatric Gastroenterology, The Wingate Institute, Barts
and the London School of Medicine and Dentistry, 26 Asheld Street, London
E1 2AJ, UK
H R van Genderingen
Departments of Pulmonary Medicine and Physics and Medical Technology, Vrije
Universiteit Medical Center, PO Box 7057, 1007 MB Amsterdam, The Netherlands
H Griths
Department of Medical Physics and Clinical Engineering, Swansea NHS Trust,
Singleton Hospital, Swansea SA2 8QA, UK
R Halter
Thayer School of Engineering, Dartmouth College, 8000 Cummings Hall,
Hanover, NH 03755-8000R, USA
A Hartov
Thayer School of Engineering, Dartmouth College, 8000 Cummings Hall,
Hanover, NH 03755-8000R, USA
D S Holder
Departments of Clinical Neurophysiology and Medical Physics and Bioengineering,
University College London, Mortimer Street, London W1T 3AA, UK
P W A Kunst
Departments of Pulmonary Medicine and Physics and Medical Technology, Vrije
Universiteit Medical Center, PO Box 7057, 1007 MB Amsterdam, The Netherlands
S Y Lee
Department of Biomedical Engineering, Impedance Imaging Research Center
(IIRC), Kyung Hee University, 1 Seochun, Kiheung, Yongin, Kyungki, South
Korea 449-701
W R B Lionheart
School of Mathematics, The University of Manchester, PO Box 88, Manchester
M60 1QD, UK
C McLeod
School of Technology, Oxford Brookes University, Gipsy Lane, Oxford OX3 0BP,
UK
J C Newell
Jonsson Engineering Center, Rensselaer Polytechnic Institute, 110 8th Street, Troy,
New York 12180, USA
N Polydorides
School of Mathematics, The University of Manchester, PO Box 88, Manchester
M60 1QD, UK
G J Saulnier
Jonsson Engineering Center, Rensselaer Polytechnic Institute, 110 8th Street, Troy,
New York 12180, USA
Jin Keun Seo
Department of Mathematics, Yonsei University, 134 Sinchon-dong,
Seodaemun-gu, Seoul 120-749, South Korea
H J Smit
Departments of Pulmonary Medicine and Physics and Medical Technology, Vrije
Universiteit Medical Center, PO Box 7057, 1007 MB Amsterdam, The Netherlands
N Soni
Thayer School of Engineering, Dartmouth College, 8000 Cummings Hall,
Hanover, NH 03755-8000R, USA
C Soulsby
Centre for Adult and Paediatric Gastroenterology, The Wingate Institute, Barts
and the London School of Medicine and Dentistry, 26 Asheld Street, London
E1 2AJ, UK
T A T Tidswell
Department of Medical Physics and Bioengineering, University College London,
Mortimer Street, London W1T 3AA, UK
A Vonk Noordegraaf
Departments of Pulmonary Medicine and Physics and Medical Technology, Vrije
Universiteit Medical Center, PO Box 7057, 1007 MB Amsterdam, The Netherlands
E J Woo
Department of Biomedical Engineering, Impedance Imaging Research Center
(IIRC), Kyung Hee University, 1 Seochun, Kiheung, Yongin, Kyungki, South
Korea 449-701
E Yazaki
Centre for Adult and Paediatric Gastroenterology, The Wingate Institute, Barts
and the London School of Medicine and Dentistry, 26 Asheld Street, London
E1 2AJ, UK
T A York
School of Electrical Engineering and Electronics, UMIST, PO Box 88, Sackville
Street, Manchester M60 1QD, UK
Introduction
Reconstruction algorithms have improved markedly, with the introduction of algorithms capable of imaging in 3D with realistic models, and the
development of powerful nonlinear approaches (chapter 1). Instrumentation
has improved incrementally, with systems able to image over multiple
frequencies and apply current patterns through multiple electrodes (chapter
2). There have not been any breakthroughs in clinical applications, but there
has been a continuing stream of pilot and proof of principle studies. A
new development is the acceptance of imaging breast cancer and brain
function among the likely leading candidates for eventual clinical take-up.
At the same time, some new potentially powerful possible applications
have been proposed and clinical trials are in progress in screening for
breast cancer, using EIT as an end-point for articial ventilation in intensive
care units, and in acute stroke and epilepsy (chapters 37). Completely new
developments have been magnetic induction tomography (chapter 8) and
Magnetic Resonance (MR-EIT) (chapter 9). Finally, there is a welcome overview of our sister research area, industrial process tomography (chapter 10),
and personal retrospective views from three of the most productive and longstanding groups in EITSheeld and Oxford Brookes Universities, UK,
and the Rensellaer Polytechnic Institute, USA (chapters 1113).
The nature of EIT is interdisciplinary. All the authors have been
encouraged to write in a non-specialist style so that their subject should be
comprehensible to most readers. All chapters should be comprehensible to
readers with a postgraduate or experienced undergraduate level in medical
physics or bioengineering. The clinical sections and much of the other
sections should be accessible to readers with a clinical background. Two
introductory non-technical appendices have been added for readers of any
background who would like a brief simple introduction to bioimpedance
or the methods of EIT. All authors have been encouraged to draw conclusions from their experience and make recommendations, positive or negative,
for future directions in development and research. I hope that the book will
be of use to those wishing to enter the eld of EIT research, and that these
opinions will be of help in setting up new methods and experiments.
Finally, I should also like to thank John Navas and Leah Fielding from
the Institute of Physics Publishing for their initiative in commissioning this
volume and patience and support in getting it published. I would like to
thank all the authors for their excellent contributions and hard work, and
the other researchers in our eld who have contributed so much to the
material in these pages and made up the happy throng at our annual conferences. Biomedical EIT research is not a subject for the faint-hearted. At the
recent conference in Gdansk, I seemed to strike a resonance in saying that the
attraction and drawback of EIT is that it doesnt clearly work, so we can reap
the fruits of its images, or not work, so we can change direction; it usually
almost works, which is an incitement to redouble our eorts. It is particularly
exciting at the time of writing, as we wait for the results of these clinical trials,
PART 1
ALGORITHMS
Chapter 1
The reconstruction problem
William Lionheart, Nicholas Polydorides and
Andrea Borsic
1.1.
In conventional medical imaging modalities, such as x-ray computerized tomography (CT), a collimated beam of radiation passes through the object in a
straight line, and the attenuation of this beam is aected only by the matter
which lies along its path. In this sense x-ray CT is local, and it means that
the pixels or voxels of our image aect only some (in fact, a very small proportion) of the measurements. If the radiation were at lower frequency (softer xrays) the eect of scattering would have to be taken into account and the eect
of a change of material in a voxel would no longer be local. As the frequency
decreases this non-local eect becomes more pronounced until we reach the
case of direct current, in which a change in conductivity would have some
eect on any measurement of surface voltage when any current pattern is
applied. This non-local property of conductivity imaging, which still applies
at the moderate frequencies used in EIT, is one of the principal reasons that
EIT is dicult. It means that to nd the conductivity image one must solve a
system of simultaneous equations relating every voxel to every measurement.
Non-locality in itself is not such a big problem provided we attempt to
recover a modest number of unknown conductivity parameters from a
modest number of measurements. Worse than that is the ill-posed nature
of the problem. According to Hadamard a mathematical model of a physical
problem is well posed if
1. for all admissible data, a solution exists,
2. for all admissible data, the solution is unique, and
3. the solution depends continuously on the data.
The problem of recovering an unknown conductivity from boundary data is
severely ill-posed, and it is the third criterion which gives us the most trouble.
In practice that means for any given measurement precision, there are
arbitrarily large changes in the conductivity distribution which are undetectable by boundary voltage measurements at that precision. This is clearly bad
news for practical low frequency electrical imaging. Before we give up EIT
altogether and take up market gardening, there is a partial answer to this
problemwe need some additional information about the conductivity
distribution. If we know enough a priori (that is in advance) information,
it constrains the solution so that the wild variations causing the instability
are ruled out.
The other two criteria can be phrased in a more practical way for our
problem. Existence of a solution is not really in question. We believe the
body has a conductivity. The issue is more that the data are suciently
accurate to be consistent with a conductivity distribution. Small errors in
measurement can violate consistency conditions, such as reciprocity. One
way around this is to project our infeasible data on to the closest point in
the feasible set. The mathematicians problem of uniqueness of solution is
better understood in experimental terms as suciency of data. In the mathematical literature the conductivity inverse boundary value problem (or
Calderon problem) is to show that a complete knowledge of the relationship
between voltage and current at the boundary determines the conductivity
uniquely. This has been proved under a variety of assumptions about the
smoothness of the conductivity [80]. This is only a partial answer to the
practical problem as we have only nitely many measurements from a xed
system of electrodes; the electrodes typically cover only a portion of the surface
of the body and in many cases voltage are not measured on electrodes driving
currents. In the practical case the number of degrees of freedom of a parameterized conductivity we can recover is limited by the number of independent
measurements made and the accuracy of those measurements.
This introductory section has deliberately avoided mathematical treatment, but a further understanding of why the reconstruction problem of
EIT is dicult, and how it might be done, requires some mathematical
prerequisites. The minimum required for the following is a reasonably
thorough understanding of matrices [145], and a little multi-variable
calculus, such as are generally taught to engineering undergraduates. For
those desirous of a deeper knowledge of EIT reconstruction, for example
those wishing to implement reconstruction software, an undergraduate
course in the nite element method [138] and another in inverse
problems [20, 22, 72] would be advantageous.
1.2.
MATHEMATICAL SETTING
Mathematical setting
Box 1.1.
Maxwells equations
In the main text we have treated essentially the direct current case. The
basic eld quantities in Maxwells equations are the electric eld E and
the magnetic eld H which will be modelled as vector-valued functions
of space and time. We will assume that there is no relative motion in our
system. The elds, when applied to a material or indeed a vacuum,
produce uxeselectric displacement D and magnetic ux B. The
spacial and temporal variations of the elds and uxes are linked by
Faradays law of induction
rE
@B
@t
@D
J
@t
low a frequency current that the magnetic eld can be neglected. We have a
given body , a closed and bounded subset of 3D space with a smooth (or
smooth enough) boundary @. The body has a conductivity which is a
function of the spatial variable x (although we will not always make this
dependence explicit for simplicity of notation). The scalar potential is
and the electric eld is E r. The current density is J r, which
is a continuum version of Ohms law. In the absence of interior current
sources, we have the continuum Kirchos law1
r r 0:
1:1
There is a recurring error in the EIT literature of calling this Poissons equation. However, it is a
natural generalization of Laplaces equation.
Mathematical setting
Box 1.2.
Sobolev spaces
In the mathematical literature you will often see the assumption that
lies in the Sobolev space H 1 , which can look intimidating to the
uninitiated. Actually these spaces are easily understood on an intuitive
level and have a natural physical meaning. For mathematical details
see Folland [53]. A (generalized) function f is in H k for integer k if
the square kth derivative has a nite integral over . For non-integer
and negative powers Sobolev spaces are dened by taking the Fourier
transform, multiplying by a power of frequency and demanding that
the result is square integrable. For the potential we are simply demanding that jrj2 dV < 1 which is equivalent, provided the conductivity
is bounded, to demanding that the ohmic power dissipated is nitean
obviously necessary physical constraint. Sobolev spaces are useful as a
measure of the smoothness of a function, and are also convenient as
they have an inner product (they are Hilbert spaces). To be consistent
with this nite power condition, the Dirichlet boundary data j@
must be in H 1=2 @ and the Neumann data j 2 H 1=2 @. Note that
the current density is one derivative less smooth than the potential on
the boundary as one might expect.
invariant on the curved face (think of electrodes running the full height of a
cylindrical tank).
The forward problem can be solved by separation of variables giving
cos k k
1 2k
cos k
1 2k
1:2
1
X
ak cos k bk sin k
and notice that the Fourier coecients of the current density will be
k1 2k =1 2k ak and similarly bk . The lowest frequency component
is clearly most sensitive to the variation in the conductivity of the anomaly.
This of itself is a useful observation indicating that patterns of voltage (or
current) with large low frequency components are best able to detect an
object near the centre of the domain. This might be achieved, for example,
by covering a large proportion of the surface with driven electrodes and
exciting a voltage or current pattern with low spacial frequency. We will
explore this further in section 1.9.3. We can understand a crucial feature of
the nonlinearity of EIT from this simple examplesaturation. Fixing the
radius of the anomaly and varying the conductivity, we see that for high
contrasts the eect on the voltage of further varying the conductivity is
reduced. A detailed analysis of the circular anomaly was performed by
Seagar [133] using conformal mappings, including oset anomalies. It is
found, of course, that a central anomaly produces the least change in boundary data. This illustrates the positional dependence of the ability of EIT to
detect an object. By analogy to conventional imaging problems one could
say that the point spread function is position dependent.
Our central circular anomaly also demonstrates the ill-posed nature of
the problem. For a given level of measurement precision, we can construct
a circular anomaly undetectable at that precision. We can make the change
in conductivity arbitrarily large and yet by reducing the radius we are still
not able to detect the anomaly. This shows (at least using the rather severe
L1 norm) that Hadamards third condition is violated.
While still on the topic of a single anomaly, it is worth pointing out that
nding the location of a single localized object is comparatively easy, and
with practise one can do it crudely by eye from the voltage data. Box 1.4
describes the disturbance to the voltage caused by a small object and explains
why, to rst order, this is the potential for a dipole source. This idea can be
made rigorous, and Ammari [3] and Seo [135] show how this could be applied
locating the position and depth of a breast tumour using data from a T-scan
measurement system.
inverse (see section 1.4) Z Y is called the transfer impedance. This follows
from uniqueness of solution of the so-called shunt model boundary value
problem, which is (1.1) together with the boundary conditions
@=@n Il
for 0 l L
1:3
El
@=@n 0
on 0
1:4
r n 0
on
1:5
Here Cn is the set of complex column vectors with n rows, whereas Cm n is the set of complex
m n matrices.
2
10
@
Vl
@n
1:6
where the contact impedance zl could vary over El but is generally assumed
constant. This new boundary condition, together with (1.3) and (1.4), form
the complete electrode model (CEM). For experimental validation of this
model see [37], theory [143] and numerical calculations [117, 155]. A nonzero
contact impedance removes the singularity in the current density, although
high current densities still occur at the edges of electrodes (g. 1.1). For
asymptotics of with the CEM see [45].
The singular values (see section 1.4.3) of Z, sometimes called characteristic
impedances, are sensitive to the electrode model used and this was used by [37]
to validate the CEM. With no modelling of electrodes and a rotationally
symmetric conductivity in a cylindrical tank, the characteristic impedances
tend toward a 1=k decay, as expected from (1.2) with sinusoidal singular
vectors of frequency k, as the number of electrodes increases.
1.4.
11
1:7
1:8
(here arg minx means the argument x which minimizes what follows). In
MATLAB3 the backslash (left division) operator can be used to calculate
the least squares solution, for example x Anb.
1.4.1.
Ill-conditioning
kAxk
:
kxk
m X
n
X
jaij j2 trace A A
i1 j 1
which treats the matrix as simply a vector rather than an operator. We also
dene the condition number
A kAk kA1 k
for A invertible. Assuming that A is known accurately, A measures the
amplication of relative error in the solution.
Specically if
Ax b and Ax x b b
then the relative error in solution and data are related by
kxk
kbk
A
kxk
kbk
as can be easily shown from the denition of operator norm. Note that this is
a worst case error boundoften the error is less. With innite precision,
3
MATLAB1 is a matrix-oriented interpreted programming language for numerical calculation
(The MathWorks Inc, Natick, MA, USA). While we write MATLAB for brevity, we include its
free relatives Scilab and Octave.
12
(a) Current density on the boundary for passive and active electrodes
Figure 1.1. The current density on the boundary with the CEM is greatest at the edge of
the electrodes, even for passive electrodes. This eect is reduced as the contact impedance
increases.
13
Figure 1.1.
(Continued)
any nite A shows that A1 is continuous, but in practice error in data
could be amplied so much the solution is useless. Even if the data b were
reasonably accurate, numerical errors mean that, eectively, A has error, and
kxk
kAk
A
:
kxk
kAk
14
Tikhonov regularization
1:9
Here we trade o actually getting a solution to Ax b and not letting kxk get
too big. The number controls this trade-o and is called a regularization
parameter. Notice that as ! 0, x tends to a generalized solution A b. It
is easy to nd an explicit formula for the minimum
x A A 2 I1 A b:
The condition number A A 2 I1 is
1 2 =
n 2 , where
1
2 0. These are normalized so that p
V v1 j v2 j j vn is
a unitary matrix V V1 . We dene i
i and for i 6 0,
ui i 1 Avi 2 Cm . Now notice that A Avi
i vi 2i vi . And
2
A ui 1
i A Aui i ui . Also AA ui i ui , where i are called singular
4
values vi and ui right and left singular vectors respectively.
The use of for singular values is conventional in linear algebra, and should cause no confusion
with the generally accepted use of this symbol for conductivity.
15
We see that the ui are the eigenvectors of the Hermitian matrix AA , so
they too are orthogonal. For a non-square matrix A, there are more eigenvectors of either A A or AA , depending on which is bigger, but only
minm; n singular values. If A < minm; n some of the i will be zero. It
is conventional to organize the singular values in decreasing order
1 2 minm;n 0.
If rankA k < n then the singular vectors vk 1 ; . . . ; vn form an orthonormal basis for null A, whereas u1 ; . . . ; uk form a basis for rangeA. On
the other hand, if k rankA < m, then v1 ; . . . ; vk form a basis for A ,
and uk 1 ; . . . ; um form an orthonormal basis for null A . In summary
Avi i ui
i minm; n
A ui i v i
i minm; n
Avi 0
rankA < i n
A ui 0
rankA < i m
ui uj
ij ;
vi vj ij
1 2 0:
It is clear from the denition
that for any matrix A, kAk 1 , while the
p
P
1
2
Frobenius norm is kAkF
i i . If A is invertible, then kA k 1=n .
The singular value decomposition (SVD) allows us to diagonalize A
using orthogonal transformations. Let U u1 j j um then AV U,
where is the diagonal matrix of singular values padded with zeros to
make an m n matrix. The nearest thing to diagonalization for nonsquare A is
U AV
and
A UV :
Although the SVD is a very important tool for understanding the illconditioning of matrices, it is rather expensive to calculate numerically and
the cost is prohibitive for large matrices.
In MATLAB the command s=svd(A) returns the singular values and
[U,S,V]=svd(A) gives you the whole singular value decomposition. There
are special forms if A is sparse, or if you only want some of the singular
values and vectors.
Once the SVD is known, it can be used to rapidly calculate the Moore
Penrose generalized inverse from
A V U
where is simply T with the nonzero i replaced by 1=i . This formula is
valid whatever the rank of A and gives the minimum norm least squares solution. Similarly the Tikhonov solution is
x VT U b
16
17
Figure 1.2. Singular values plotted on a logarithmic scale for the linearized 3D EIT
problem with 32 electrodes, and cross sections of two singular vectors.
singular vectors vi dier then they will be able to reliably reconstruct dierent
conductivities. To test how easy it is to detect a certain (small as we have
linearized) conductivity change x, we look at the singular spectrum V x. If
most of the large components are near the top of this vector the change is
easy to detect, whereas if they are all below the lth row they are invisible
with relative error worse than l =0 . The singular spectrum U b of a set of
measurements b gives a guide to how useful that set of measurements will
be at a given error level.
1.4.5.
18
Box 1.3.
PbjxPx
:
Pb
1
exp 12 kAx bk2Q exp 12 kx x0 k2P
Pb
Regularizing EIT
19
1.5.
REGULARIZING EIT
20
1:10
Linearized problem
Regularizing EIT
21
1:11
or any of the equivalent forms [149]. While there are many other forms of
regularization possible for a linear ill-conditioned problem, this generalized
Tikhonov regularization has the benet that (see Box 1.3) the a priori information it incorporates is made explicit and that under Gaussian assumptions
it is the statistically defensible MAP estimate. If only a linearized solution is
to be used with a xed initial estimate s0 , the Jacobian J and a factorization
of J J 2 L L can be precalculated o-line. The eciency of this calculation is then immaterial and the regularized solution can be calculated using
the factorization with complexity ON 2 for N degrees of freedom in the
conductivity (which should be smaller than the number of independent
measurements). Although LU factorization would be one alternative,
perhaps a better choice is to use the GSVD [72], which allows the regularized
solution to be calculated eciently for any value of . The GSVD is now a
standard tool for understanding the eect of the choice of the regularization
matrix L in a linear ill-conditioned problem, and has been applied to linearized EIT [16, 152]. The use of a single linearized Tikhonov regularized
solution is widespread in medical industrial and geophysical EIT, the
NOSER algorithm [35] being a well known example.
1.5.2.
Back-projection
22
measured along rays through that point. Although not an inverse to the
Radon transform itself, a smooth image can be obtained by back-projecting
smoothed data, or equivalently by back-projecting then smoothing the
resulting image.
The Tikhonov regularization formula (1.11) can be interpreted in a loose
way as the back-projection operator J , followed by application of the spatial
lter J J 2 L L1 . Although this approach is quite dierent from the
ltered back-projection along equipotential lines of Barber and Brown [9,
130], it is sometimes confused with this in the literature. Kotres back-projection was until recently widely used in the process tomography community for
both resistivity (ERT) and permittivity (ECT) imaging [163], often supported
by fallacious arguments, in particular that it is fast (it is no faster than the
application of any precomputed regularized inverse) and that it is commonly
used (only by those who know no better). In an interesting development the
application of a normalized adjoint to the residual voltage error for the linearized problem was suggested for ECT, and later recognized as yet another reinvention of the well-known Landweber iterative method [162]. Although there
is no good reason to use pure linear iteration schemes directly on problems
with such a small number of parameters, as they can be applied much faster
using the SVD, an interesting variation is to use such a slowly converging
linear solution together with projection on to a constraint set; a method
which has been shown to work well in ECT [30].
1.5.3.
The use of linear approximation is only valid for small deviations from the
reference conductivity. In medical problems conductivity contrasts can be
large, but there is a good case for using the linearized method to calculate
a change in admittivity between two states, measured either at dierent
times or with dierent frequencies. Although this has been called dynamic
imaging in EIT the term dierence imaging is now preferred (dynamic
imaging is better used to describe statistical time series methods such as
[154]). In industrial ECT modest variations of permittivity are commonplace.
In industrial problems and in phantom tanks it is possible to measure a reference data set using a homogeneous tank. This can be used to calibrate the
forward model; in particular the contact impedance can be estimated [74].
In an in vivo measurement there is no such possibility, and it may be that
the mismatch between the measured data and the predictions from the
forward model is dominated by the errors in electrode position, boundary
shape and contact impedance rather than interior conductivity. Until these
problems are overcome it is unlikely, in the authors opinion, to be worth
using iterative nonlinear methods in vivo using individual surface electrodes.
Note, however, that such methods are in routine use in geophysical problems
[95, 96].
23
TV f jrf j:
1:12
24
TV f sup
f div v
1:13
v2V
Figure 1.3. Three possible functions: f1 , f2 , f3 2 F. All of them have the same TV, but only
f2 minimizes the H 1 semi-norm.
25
1:15
26
y:kyk1
1:16
1:17
1:18
max
y:kyk1
AT Axb LT y0
1
2 kAx
bk2 yT Lx:
1:19
The complementarity condition for (1.15) and (1.19) is set by nulling the
primal dual gap
1
2 kAx
1:20
1:21
The PDIPM framework for the TV regularized inverse problem can thus be
written as
kyk 1
T
1:22a
T
A Ax b L y 0
1:22b
Lx kLxky 0:
1:22c
Application to EIT
The PDIPM algorithm in its original form [33] was developed for inverse
problems with linear forward operators. The following section (based
on [14]) describes the numerical implementation for EIT reconstruction.
The implementation is based on the results of the duality theory for inverse
problems with linear forward operators. Nevertheless it was possible to apply
the original algorithm to the EIT inverse problem with minor modications,
27
and to obtain successful reconstructions. The formulation for the EIT inverse
problem is
srec arg mins f s
f s 12 kFs Vm k2 TVs:
1:23
With a similar notation as used in section 1.6.1, the system of nonlinear equations that denes the PDIPM method for (1.23) can be written as
kyk 1
JT Fs Vm LT y 0
1:24
Ls Ey 0
p
2
with E diag jLi sj where Li is ithe row of L, and J the Jacobian of
the forward operator Fs. Newtons method can be applied to solve (1.24)
obtaining the following system for the updates s and y of the primal and
dual variables:
T
T
J J LT s
J Fs Vm LT y
1:25
y
Ls Ey
HL E
with
H I E1 diagyi Li s
1:26
1:27a
and
y y E 1 Ls E 1 HL s:
1:27b
1:28
maxf
:
kyk yk k 1g:
1:29
where
An alternative way is to calculate the exact step length to the boundary,
applying what is called the steplength rule [5]
yk 1 yk min1;
yk
1:30
28
where
maxf
: kyk
yk k 1g:
1:31
1.7.
JACOBIAN CALCULATIONS
In optimization-based methods it is often necessary to calculate the derivative of the voltage measurements with respect to a conductivity parameter.
The complete matrix of partial derivatives of voltages with respect to
conductivity parameters is the Jacobian matrix, sometimes in the medical
and industrial EIT literature called the sensitivity matrix, or the rows are
called sensitivity maps. We will describe here the basic method for calculating this eciently with a minimal number of forward solutions. Let it be
said rst that there are methods where the derivative is calculated only
once, although the forward solution is calculated repeatedly as the conductivity is updated. This is the dierence between NewtonKantorovich
method and Newtons method. There are also quasi-Newton methods in
which the Jacobian is updated approximately from the forward solutions
that have been made. Indeed this has been used in geophysics [96]. It is
also worth pointing out that where the conductivity is parameterized in a
nonlinear way, for example using shapes of an anatomical model, the
Jacobian with respect to those new parameters can be calculated using the
chain rule.
1.7.1.
Perturbation in power
Using the weak form of r r 0 (or Greens identity), for any w
@
dS:
1:32
r rw dV
w
@n
@
Jacobian calculations
29
We use the complete electrode model. For the special case w we have the
power conservation formula
X
@
@
@
2
dS
dS
1:33
jrj dV
V l zl
@n
@n
@n
@
El
l
hence
jrj dV
X
l
El
@ 2 X
Vl Il :
zl
@n
1:34
This simply states that the power input is dissipated either in the domain or
by the contact impedance layer under the electrodes.
In the case of full time harmonic Maxwells equations (Box 1.1) the
H. The complex power crosspower ux is given by the Poynting vector E
ing the boundary is then equal to the complex power dissipated and stored in
the interior (the imaginary part representing the power stored as electric and
magnetic energy)
H n dS E E i!H H dV
1:35
E
@
X
Il Vl jrj2 dV:
1:36
This gives only the total change in power. To get the change in voltage
on a particular electrode Em when a current pattern is driven in some or all of
the other electrodes, we simply solve for the special measurement current
pattern I~lm lm . To emphasize the dependence of the potential on a
vector of electrode currents I I1 ; . . . ; IL we write I. The hypothetical
measurement potential is uIm ; by contrast the potential for the dth drive
pattern is Id . Taking the real case for simplicity and applying the power
perturbation formula (1.36) to Id Im and Id Im and then
subtracting gives the familiar formula
30
1
1:38
Vij 2 ri rj dV:
I
To calculate the Jacobian matrix one must choose a discretization of the
conductivity. The simplest case is to take the conductivity to be piecewise
constant on polyhedral domains such as voxels or tetrahedral elements.
Taking to be the characteristic function of the kth voxel k we have for
a xed current pattern
@Vdm
Jdm k
ruId rIm dV:
1:39
@k
k
Some EIT and capacitance tomography systems use a constant voltage
source and in this case the change in power of an increase in admittivity
will have the opposite sign to the constant current case.
A common variation in the case of real conductivity is to use the resistivity
1= as the primary variable, or more commonly to use log [10, 26, 155],
which has the advantage that it does not need to be constrained to be positive.
With a simple parameterization of conductivity as constant on voxels, g is
constant on voxels as well, for any function g. In this case from the chain rule
we simply use the chain rule, dividing the kth column of Jacobian we have calculated by g0 k . The regularization will also be aected by the change of variables.
Some iterative nonlinear reconstruction algorithms, such as nonlinear
Landweber, or nonlinear conjugate gradient (see section 1.8.3 and [160])
require the evaluation of transpose (or adjoint) of the Jacobian multiplied
by a vector J z. For problems where the Jacobian is very large it may be
undesirable to store the Jacobian and then apply its transpose to z. Instead
the block of zi corresponding to the ith current drive is written as distributed
source on the measurement electrodes. A forward solution is performed with
this as the boundary current pattern so that when this measurement eld is
combined with the eld for the drive pattern as (1.39), this block accumulates
to give J z. For details of this applied to diuse optical tomography see [6],
and for a general theory of adjoint sources see [160].
Jacobian calculations
Box 1.4.
31
Studying the change in voltage from a small localized change in conductivity is a useful illustration of EIT. Suppose we x a current pattern,
and a background conductivity of , which results in a potential . Now
consider a perturbed conductivity which results in a potential,
with the same current drive, . From r r 0
we see that
r r r r r r 0:
The same procedure used to calculate the Jacobian can be used to show
that the last term is O 2 so that to rst order
r r r r:
Now for simplicity take 1 and we have the Poisson equation for :
r2 r r:
If we now take to be a small change, constant on a small ball near some
point p, then the source term in this Poisson equation approximates a
dipole at p whose strength and direction is given by r. Observing at
the boundary we see it as a dipole eld from which a line through p can
be estimated by eye. This goes some way to explain the ease with which
one small object can be located, even with only a small number of current
patterns. It also illustrates the depth dependence of the sensitivity as the
dipole eld decays with distance, even if the electric eld is relatively
uniform. Typically the electric eld strength is also closer to the boundary.
This continuum argument is paralleled in Yorkeys compensation
method in resistor networks [164]. A resistor in a network is changed
and Yorkey observes that to rst order the change in voltage at each
point in the network is equivalent to the voltage which would result if
a current source were applied in parallel with that resistor.
32
For fast calculation of the Jacobian using (1.39) one can precompute
the integrals of products of nite element (FE) basis functions over
elements. If non-constant basis functions are used on elements, or higher
order elements are used, one could calculate the product of gradients of
FE basis functions at quadrature points in each element. As this depends
only on the geometry of the mesh and not the conductivity, this can be
precomputed unless one is using an adaptive meshing strategy. The same
data is used in assembling the FE system matrix eciently when the conductivity has changed but not the geometry. It is these factors particularly
which make current commercial FE method software unsuitable for use in
an ecient EIT solver.
1.8.
To solve the inverse problem one needs to solve the forward problem for some
assumed conductivity so that the predicted voltages can be compared with the
measured data. In addition, the interior electric elds are usually needed for
the calculation of a Jacobian. Only in cases of very simple geometry, and
homogeneous or at least very simple conductivity, can the forward problem
be solved analytically. These can sometimes be used for linear reconstruction
algorithms on highly symmetric domains. Numerical methods for general
geometry and arbitrary conductivity require the discretization of both the
domain and the conductivity. In the nite element method (FEM), the 3D
domain is decomposed in to (possibly irregular) polyhedra (e.g. tetrahedra,
prisms or hexahedra) called elements, and on each element the unknown
potential is represented by a polynomial of xed order. Where the elements
intersect they are required to intersect only in whole faces or edges or at
vertices, and the potential is assumed continuous (or derivatives up to a certain
order continuous), across faces. The FEM converges to the solution (or at least
the weak solution) of the partial dierential equation it represents, as the
elements become more numerous (provided their interior angles remain
bounded) or as the order of the polynomial is increased [146].
The nite dierence method and nite volume method are close relatives
of the FEM, which use regular grids. These have the advantage that more
ecient solvers can be used at the expense of the diculty in accurately representing curved boundaries or smooth interior structures. In the boundary
element method (BEM) only surfaces of regions are discretized, and an
analytical expression for the Green function is used within enclosed volumes
that are assumed to be homogeneous. BEM is useful for EIT forward modelling provided one assumes piecewise constant conductivity on regions with
smooth boundaries (e.g. organs). BEM results in a dense rather than a
sparse linear system to solve, and its computational advantage over FEM
33
diminishes as the number of regions in the model increases. BEM has the
advantage of being able to represent unbounded domains. A hybrid
method where some regions assumed homogeneous are represented by
BEM, and inhomogeneous regions by FEM, may be computationally
ecient for some applications of EIT [134].
In addition to the close integration of the Jacobian calculation and the
FEM forward solver, another factor which leads those working on EIT
reconstruction to write their own FEM programme for the complete
electrode model (CEM) is a non-standard type of boundary condition not
included in commercial FEM software. It is not hard to implement and
there are freely available codes [122, 157], but it is worth covering the basic
theory here for completeness. A good introduction to FEM in electromagnetics is [138], and details of implementation of the CEM can be
found especially in the theses [123, 155].
1.8.1.
N
X
i wi x
1:40
i1
v r r dV 0 in
1:41
and we demand that this vanishes for all functions v in a certain class. Clearly
this is weaker than assuming directly that r r 0.
Using Greens second identity and the vector identity
r v r r rv vr r
1:42
34
r v r dV r rv dV 0:
1:43
r v r dV
v r n dS
1:44
@
gives
r rv dV
r nv dS
r nv dS
1:45
@
S
where l El is the union of the electrodes, and we have used the fact that
the current density is zero o the electrodes. For a given set of test functions
v, (1.45) is the weak formulation of the boundary value problem for (1.1)
with current density specied on the electrodes.
Rearranging the boundary condition (1.6) as
r n
1
V
zl l
L
X
1
r rv dV
Vl v dS:
z
l 1 El l
1:46
1:47
N
L
X
X
1
rwi rwj dV j
wi wj dS j
E l zl
j1
l1
L
X
1
wi dS Vl 0:
El z l
l1
N
X
1
1
1
Vl dS
Vl
wi dS i
Il
El z l
El z l
El z l
i
and if we assume zl is constant on El this reduces to
N
1
1X
wi dS i
Il jEl jVl
zl
zl i
El
1:48
1:49
1:50
where jEl j is the area (or in two dimensions, length) of the lth electrode.
35
which has the advantage that the j can be taken outside of an integral over
each simplex. If a more elaborate choice of basis is used it would be wise to
use a higher-order quadrature rule.
Our FE system equations now take the form
0
AM AZ AW
1:52
T
V
I
AW
AD
where AM is an N N symmetric matrix
K
X
k
AM;ij rwi rwj dV
k1
k
rwi rwj dV
1:53
which is the usual system matrix for (1.1) without boundary conditions, while
L
X
1
wi wj dS
1:54
AZ;ij
z
l 1 El l
1
AW;ij
w dS
1:55
z l El i
and
jEl j
AD diag
1:56
zl
implement the CEM boundary conditions. One additional constraint is
required as potentials are only dened up to an added constant. One elegant
choice is to change the basis used for the vectors V and I to a basis for the
subspace S orthogonal to constants, for example the vectors
T
1
1
1
1
;...;
; 1;
;...;
1:57
L1
L1
L1
L1
while another choice is to ground an arbitrary vertex i by setting i 0. The
resulting solution can then have any constant added to produce a dierent
grounded point.
As the contact impedance decreases the system, (1.52) becomes illconditioned. In this case (1.6), in the CEM can be replaced by the shunt
model, which simply means the potential is constrained to be constant on
each electrode. This constraint can be enforced directly replacing all nodal
voltages on electrode El by one unknown Vl .
36
It is important for EIT to notice that the conductivity only enters in the
system matrix as linear multipliers of
sijk
rwi rwj dV jk jrwi rwj
k
which depend only on the FE mesh and not on . These coecients can be
pre-calculated during the mesh generation, saving considerable time in the
system assembly. An alternative is to dene a discrete gradient operator
D : CN ! C3K , which takes the representation as a vector of vertex values
of a piecewise linear function to the vector of r on each simplex (on
which of course the gradient is constant). On each simplex dene
k k =jk jI3 , where I3 is the 3 3 identity matrix, or for the anisotropic
case simply the conductivity matrix on that simplex divided by its volume,
and diagk IK . We can now use
AM DT D
1:58
We now consider the solution of the system (1.52). The system has the
following special features. The matrix is sparse: the number of nonzeros in
each row of the main block depends on the number of neighbouring verticies
connected to any given vertex by an edge. It is symmetric (for complex
conductivity and contact impedance that means real and imaginary parts
are symmetric), and the real part is positive denite. In addition, we have
multiple right-hand sides for the same conductivity, and we wish to solve
the system repeatedly for similar conductivities.
A simple approach to solving Ax b is LU-factorization [66], where an
upper triangular matrix U and lower triangular matrix L are found such that
A LU. As solving a system with a diagonal matrix is trivial, one can solve
Lu b (forward substitution) and then Ux u (backward substitution). The
factorization process is essentially Gaussian elimination and has a computational cost On3 , while the backward and forward substitute have a cost
On2 k for k right-hand sides. An advantage of a factorization method
such as this is that one can apply the factorization to multiple right-hand
sides, in our case for each current pattern. Although the system matrix is
sparse, the factors are in general less so. Each time a row is used to eliminate
the nonzero elements below the diagonal it can create more nonzeros above
the diagonal. As a general rule it is better to reorder the variables so that rows
with more nonzeros are farther down the matrix. This reduces the ll in of
nonzeros in the factors. For a real symmetric or Hermitian matrix the
symmetric multiple minimum degree algorithm [55] reduces ll in, whereas
the column multiple minimum degree algorithm is designed for the general
Box 1.5.
37
case. For an example see gure 1.4. The renumbering should be calculated
when the mesh is generated so that it is done only once.
For large 3D systems direct methods can be expensive and iterative
methods may prove more ecient. A typical iterative scheme has a cost of
On2 k per iteration and requires fewer than n iterations to converge. In
fact the number of iterations required needs to be less than Cn=k for some
38
Figure 1.4. Top left: the sparsity pattern of a system matrix which is badly ordered for llin. Bottom left: sparsity pattern for the U factor. On the right, the same after reordering
with colmmd.
C depending on the algorithm to win over direct methods. Often the number
of current patterns driven is limited by hardware to be small, while the
number of vertices in a 3D mesh needs to be very large to accurately
model the electric elds, and consequently iterative methods are often
preferred in practical 3D systems. The potential for each current pattern
can be used as a starting value for each iteration. As the adjustments in
the conductivity become smaller this reduces the number of iterations
required for forward solution. Finally it is not necessary to predict the
voltages to full oating point accuracy when the measurements system
itself is far less accurate than this, again reducing the number of iterations
required.
The convergence of iterative algorithms, such as the conjugate gradient
method (see section 1.8.3), can be improved by replacing the original system
39
The conjugate gradient (CG) method [18, 66] is a fast and ecient method for
solving Ax b for real symmetric matrices A or Hermitian complex
matrices. It can also be modied for complex symmetric matrices [29]. The
method generates a sequence xi (iterates) of successive approximations to
the solution and residuals ri b Axi , and search directions pi and
qi Api used to update the iterates and residuals. The update to the iterate
is
xi xi 1 i pi
1:59
1:60
kri 1 k2
:
pi Api
1:61
1:62
where using
i
kri k2
kri 1 k2
1:63
ensures that pi are orthogonal to all Apj and ri are orthogonal to all rj , for
j < i. The iteration can be terminated when the norm of the residual falls
below a predetermined level.
Conjugate gradient least squares (CGLS) method solves the least
squares problem (1.7) AT Ax AT b without forming the product AT A
(also called CGNR or CGNE conjugate gradient normal equations [18, 32])
and is a particular case of the nonlinear conjugate gradient (NCG) algorithm
of Fletcher and Reeves [52] (see also [160, ch 3]). The NCG method seeks a
minimum of cost functions f x 12 kb Fxk2 , which in the case of CGLS
is simply the quadratic 12 kb Axk2 . The direction for the update in (1.59) is
now
pi rf xi Ji b Fxi
1:64
40
1:65
Mesh generation
Figure 1.5.
41
42
1.9.
MEASUREMENT STRATEGY
Linear regression
We will illustrate the ideas mainly using the assumption that the currents are
prescribed and the voltages are measured, although there are systems which do
the opposite. In this approach we regard the matrix of voltage measurements
to be contaminated by noise, while the currents are known accurately. This
should be compared with the familiar problem of linear regression where we
aim to t a straight line to experimental observations. Assuming a relationship
of the form y ax, we will assume an intercept of zero and mean x of zero.
Measurement strategy
43
The abscissae xi are assumed accurate and the yi contaminated with noise.
Assembling the xi and yi into row vectors x and y, we estimate the slope a by
a^ arg mina ky axk2 :
1:66
Of course the solution is a yx , another way of expressing the usual regression formulae. The least squares approach can be justied statistically [112].
Assuming the errors in y have zero correlation, a^ is an unbiased estimator for
a. Under the stronger assumption that the yi are independently normally
distributed with identical variance, a^ is the maximum likelihood estimate
of a, and is normally distributed with mean a. Under these assumptions we
can derive condence intervals and hypothesis testing for a [112, p 14].
Although less well known, linear regression for several independent
variables follows a similar pattern. Now X and Y are matrices and we seek
^ YX has the same
a linear relation of the form Y AX. The estimate A
desirable statistical properties as the single variable case [112, ch 2].
Given a system of K current patterns assembled in a matrix I 2 CL K
(with column sums zero), we measure the corresponding voltages as V ZI.
Assuming the currents are accurate but the voltages contain error, we then
^ VI . If we have two few linearly independent currents
obtain our estimate Z
of rank I < L 1, then this will be an estimate of a projection of Z on to a
subspace, and if we have more than L 1 current patterns then the generalized
^ . Similarly we
inverse averages over the redundancy, reducing the variance of Z
ML
can make redundant measurements. Let M 2 R
be a matrix containing the
measurement patterns used (for simplicity the same for each current pattern),
so that we measure VM MV. For simplicity we will assume that separate
electrodes are used for drive and measurement, so there is no reciprocity in
the data. Our estimate for Z is now M VM I . For a thorough treatment of
the more complicated problem of estimating Z for data with reciprocity see
[46]. In both cases redundant measurements will reduce variance. Of course it
is common practice to take multiple measurements of each voltage, and the
averaging of these may be performed within the data acquisition system
before it reaches the reconstruction programme. In this case the eect is
identical to using the generalized inverse. The benet in using the generalized
inverse is that it automatically averages over redundancy where there are
multiple linearly dependent measurements. If quantization in the analogueto-digital converter (ADC) is the dominant source of error, averaging over
dierent measurements reduces the error, in a similar fashion to dithering
(adding a random signal and averaging) to improve the accuracy of an ADC.
Some EIT systems use variable gain ampliers before voltage measurements
are passed to the ADC. In this case the absolute precision varies between
measurements and a weighting must be introduced in the norms used to
dene the least squares problem.
For the case where the voltage is accurately controlled and the current
measured, an exactly similar argument holds for estimating the transfer
44
admittance matrix. However, where there are errors in both current and
voltage, for example caused by imperfect current sources, a dierent estimation procedure is required. What we need is multiple correlation analysis [112,
p 82] rather than multiple regression.
One widely used class of EIT systems which use voltage drive and
current measurement are ECT systems used in industrial process
monitoring [30]. Here each electrode is excited in turn with a positive voltage
while the others are at ground potential. The current owing to ground
through the non-driven electrode is measured. Once the voltages are adjusted
to have zero mean this is equivalent to using the basis (1.57) for YjS .
We know that feasible transfer impedance matrices are symmetric, and
^ by
so employ the orthogonal projection on to the feasible set and replace Z
T
1
^
sym Z where sym A 2 A A . This is called averaging over reciprocity
error. The skew-symmetric component of the estimated Z gives an indication
of errors in the EIT instrumentation.
1.9.2. Sheeld measurement protocol
The space of contact impedances is a subset of the vector space of symmetric
L L matrices with column and row sums zero, which has dimension
LL 1=2. In addition the real part of ZjS is positive denite, otherwise
there would be direct current patterns which dissipate no power. There are
other conditions on Z, given in the plane case by [42], associated with
being connected, and it is shown in the planar case that the set of feasible
Z is an open subset of the vector space described above. This conrms that
we can measure up to LL 1=2 independent parameters. Some systems,
however, measure fewer than this, primarily to avoid measuring voltage on
actively driven electrodes.
The Sheeld mark I and II systems [12] use a protocol with L 16 electrodes which are typically arranged in a circular pattern on the subject. Adjacent
pairs El and El 1 are excited with equal and opposite currents, for L ranging
from 1 to L 1. These can be assembled into a matrix IP 2 RL L 1 with
lk lk 1 in the lk position. Clearly the columns of IP span S. Measurements
are made similarly between adjacent pairs and IPT gives the measurement
patterns so that the matrix of all possible voltages measured is ZP IPT ZIP , a
symmetric L 1 L 1 matrix of full rank. However, when the lth electrode pair is excited, the measurement pairs l 1, l and l 1 are omitted
(indices are assumed to wrap around when out of range). The subset of ZP
which is actually measured by the Sheeld system is shown in gure 1.6 and
a simple counting argument shows that the number of independent measurements is L 2L 1=2 1 LL 3=2, or 104 for L 16.
In practice a Sheeld mark I or II system aiming at speed rather than
accuracy measures a non-redundant set of exactly 104 measurements. For
the rst two drive patterns all 13 patterns are measured, and for subsequent
Measurement strategy
45
Figure 1.6. Each column corresponds to a drive pair and each row to a measurement pair.
A l indicates a measurement that is taken and a k one which is omitted.
drives one less is measured each time. If reciprocity error is very small this is
an acceptable strategy.
A pair drive system has the advantage that only one current source is
needed, which can then be switched to each electrode pair. With a more
complex switching network other pairs can be driven at the expense of
higher system cost and possibly a loss of accuracy. A study of the dependence
of the SVD of the Jacobian for dierent separations between driven electrodes
can be found in [25].
One feature of the Sheeld protocol is that on a 2D domain the adjacent
voltage measurements are all positive. This follows as the potential itself is
monotonically decreasing from source to sink. The measurements also
have a U-shaped graph for each drive. This provides an additional feasibility
check on the measurements. Indeed if another protocol is used, Sheeld data
ZP can be synthesized to employ this check.
1.9.3.
The problem of optimizing the drive patterns in EIT was rst considered by
Seagar [133], who calculated the optimal placing of a pair of point drive electrodes on a disk to maximize the voltage dierences between the measurement
of a homogeneous background and an oset circular anomaly. Isaacson [78]
and Gisser et al [60] argued that one should choose a single current pattern
to maximize the L2 norm of the voltage dierence between the measured Vm
and calculated Vc voltages constraining the L2 norm of the current patterns
in a multiple-drive system. This is a simple quadratic optimization problem
Iopt arg minI 2 S
kVm Vc Ik
kIk
1:67
46
kVm Vc Ik
kVm Ik
1:68
Numerical examples
47
are not linearly related, i.e. the null hypothesis H0 : Zm Zc 0, which can
be tested using a suitable statistic with an F-distribution [112, p 133]. If only
one current normalized pattern is used the optimal current will give a test
with the greatest power. In the statistical terminology, power is the conditional probability, so we reject the hypothesis H0 given that it is false.
Kaipio et al [82] suggest choosing current patterns that minimize the total
variance of the posterior. In this Bayesian framework the choice of optimal
current patterns depends on the prior and a good choice will result in a tighter
posterior. Demidenko et al [47] consider optimal current patterns in the framework conventional optimal design of experiments, and dene an optimal set of
current patterns as one that minimizes the total variance of Z.
Eyoboglu and Pilkington [51] argued that medical safety legislation
demanded that one restricts the maximum total current entering the body,
and if this constraint was used the distinguishability is maximized by pair
drives. Cheney and Isaacson [38] study a concentric anomaly in a disk,
using the gap model for electrodes. They compare trigonometric, Walsh
and opposite and adjacent pair drives for this case giving the dissipated
power, as well as the L2 and power distinguishabilities. Koksal and
Eyoboglu [85] investigate the concentric and oset anomaly in a disk using
continuum currents. Further study of optimization of current patterns
with respect to constraints can be found in [93].
1.10.
NUMERICAL EXAMPLES
Figure 1.7. Mesh used for potentials in reconstruction. A coarser mesh, of which this is a
subdivision, was used to represent the conductivity.
48
(b)
(c)
Figure 1.8. (a) Original smooth conductivity distribution projected onto the coarser mesh
(Mayavi surface map). (b) Smoothly regularized GaussNewton reconstruction of this
smooth conductivity. (c) TV regularized PDIPM reconstruction of the same smooth
conductivity.
Numerical examples
49
Figure 1.9. Electrodes, mesh and two spheres test object. The test object consisted of
two spheres of conductivity 1 in a background of 3. An unrelated ner mesh was used
to generate the simulated data.
50
(a)
(b)
Figure 1.10. Reconstruction of a two-spheres test object from gure 1.9 using regularized
GaussNewton and TV PDIPM. (a) Regularized GaussNewton reconstruction, shown
using cut-planes. (b) Total variation reconstruction from PDIPM.
1.11.
51
slightly jokingly called an inverse crime [44, p 133] (by analogy with the
variational crimes in FEM perhaps). We list a few guidelines to avoid
being accused of an inverse crime and to lay out what we believe to be best
practice. For slightly more details see [94].
1. Use a dierent mesh. If you do not have access to a data collection system
and phantom tank, or if your reconstruction code is at an early stage of
development, you will want to test with simulated data. To simulate the
data use a ner mesh than is used in the forward solution part of the
reconstruction algorithm. But not a strict renement. The shape of any
conductivity anomalies in the simulated data should not exactly conform
with the reconstruction mesh, unless you can assume the shape is known
a priori.
2. Simulating noise. If you are simulating data you must also simulate the
errors in experimental measurement. At the very least there is quantization error in the analogue-to-digital converter. Other sources of error
include stray capacitance, gain errors, inaccurate electrode position,
inaccurately known boundary shape, and contact impedance errors. To
simulate errors sensibly it is necessary to understand the basics of the
data collection system, especially when the gain on each measurement
channel before the ADC is variable. When the distribution of the voltage
measurement errors is decided this is usually simulated with a pseudorandom number generator.
3. Pseudo-random numbers. A random number generator models a draw
from a population with a given probability density function. To test the
robustness of your reconstruction algorithm with respect to the magnitude
of the errors it is necessary to make repeated draws, or calls to the random
number generator, and to study the distribution of reconstruction errors.
As our inverse problem is nonlinear, even a Gaussian distribution of
error will not produce a (multivariate) Gaussian distribution of reconstruction errors. Even if the errors are small and the linear approximation good,
at least the mean and variance should be considered.
4. Not tweaking. Reconstruction programmes have a number of adjustable
parameters such as Tikhonov factors and stopping criteria for iteration,
as well as levels of smoothing, basis constraints and small variations of
algorithms. There are rational ways of choosing reconstruction parameters based on the data (such as generalized cross validation and Lcurve), and on an estimate of the data error (Morotzovs stopping criterion). In practice one often nds acceptable values empirically which work
for a collection of conductivities one expects to encounter. There will
always be other cases for which those parameter choices do not work
well. What one should avoid is tweaking the reconstruction parameters
for each set of data until one obtains an image which one knows is
close to the real one. By contrast an honest policy is to show examples
52
1.12.
In this review there is not space to describe in any detail many of the exciting
current developments in reconstruction algorithms. Before highlighting some
of these developments it is worth emphasizing that for an ill-posed problem,
a priori information is essential for a stable reconstruction algorithm, and it
is better that this information is incorporated in the algorithm in a systematic
and transparent way. Another general principle of inverse problems is to think
carefully what information is required by the end user. Rather than attempting
to produce an accurate image, what is often required in medical (and indeed
most other) applications is an estimate of a much smaller number of parameters which can be used for diagnosis. For example, we may know that a
patient has two lungs as well as other anatomical features, but we might
want to estimate their water content to diagnose pulminary oedema. A sensible
strategy would be to devise an anatomical model of the thorax and t a few
parameters of shape and conductivity rather than pixel conductivity values.
The disadvantage of this approach is that each application of EIT gives rise
to its own specialized reconstruction method, which must be carefully designed
for the purpose. In the authors opinion the future development of EIT
systems, including electrode arrays and data acquisition systems as well as
reconstruction software, should focus increasingly on specic applications,
although of course such systems will share many common components.
1.12.1.
53
problem as well as the forward problem [98, 108, 109]. In both cases there is
the interesting possibility of exploring the interaction between the meshes
used for forward and inverse solution.
At the extreme end of this spectrum we would like to describe the prior
probability distribution and for a known distribution of measurement noise
to calculate the entire posterior distribution. Rather than giving one image,
such as the MAP estimate, we give a complete description of the probability
of any image. If the probability is bimodal, for example, one could present
the two local maximum probability images. If one needed a diagnosis, say
of a tumour, the posterior probability distribution could be used to calculate
the probability that a tumour-like feature was there. The computational
complexity of calculating the posterior distribution for all but the simplest
distributions is enormous; however, the posterior distribution can be
explored using the Markov Chain Monte Carlo method which has been
applied to 2D EIT [81]. This was applied to simulated EIT data [54], and
more recently to tank data, for example [111]. For this to be a viable
technique for the 3D problem, highly ecient forward solution will be
required.
1.12.2.
54
fast, relies on the resistivity of the body known to be one of two values. It
works equally well in two and three dimensions and is robust in the presence
of noise. The time complexity scales linearly with the number of voxels
(which can be any shape) and scales cubically in the number of electrodes.
It works for purely real or imaginary admittivity (ERT or ECT), and for
magnetic induction tomography for real conductivity. It is not known if it
can be applied to the complex case and it requires the voltage on current
carrying electrodes.
Linear sampling methods [24, 71, 131] have similar time complexity and
advantages as the monotonicity method. While still applied to piecewise
constant conductivities, linear sampling methods can handle any number
of discrete conductivity values provided the anomalies are separated from
each other by the background. The method does not give an indication of
the conductivity level but rather locates the jump discontinuities in conductivity. Both monotonicity and linear sampling methods are likely to nd
application in situations where a small anomaly is to be detected and located,
for example breast tumours.
Finally, a challenge remains to recover anisotropic conductivity which
arises in applications from brous or stratied media (such as muscle),
ow of non-spherical particles (such as red blood cells), or from compression
(e.g. in soil). The inverse anisotropic conductivity problem at low frequency
is known to suer from insuciency of data, but with sucient a priori
knowledge (e.g. [92]) the uniqueness of solution can be restored. One has
to take care that the imposition of a nite element mesh does not predetermine which of the family of consistent solutions is found [119]. Numerical
reconstructions of anisotropic conductivity in a geophysical context
include [116], although there the problem of non-uniqueness of solution
(dieomorphism invariance) has been ignored. Another approach is to
assume piecewise constant conductivity with the discontinuities known, for
example from an MRI image, and seek to recover the constant anisotropic
conductivity in each region [56], [57].
1.13.
PRACTICAL APPLICATIONS
Practical applications
55
method [131] and the scattering transform method [105] have been applied
to tank data. However, there is a paucity of application of nonlinear
reconstruction algorithms to in vivo human data.
Most of the clinical studies in EIT assume circular or other simplied
geometry and regular placement of electrodes. Without the correct modelling
of the boundary shape and electrode positions [91] the forward model cannot
be made to t the data by adjusting an isotropic conductivity. A nonlinear
iterative reconstruction method would therefore not converge, and for this
reason most clinical studies have used a linearization of the forward problem
and reconstruct a dierence image from voltage dierences. This linearization has been regularized in various ways, using both ad hoc methods such
as those used by the Sheeld group [9, 10] and systematic methods such as
the NOSER method [35] of RPI. Studies of EIT on the chest such as [79,
106, 144] assume a 2D circular geometry, although some attempts have
been made to use a realistic chest shape [90] (see also chapter 13, gure
13.9). Similar simplications have been made for EIT studies of the head
and breast. 3D linear reconstruction algorithms have been applied to the
human thorax [21, 101, 114] (see also chapter 13, gure 13.10). However,
3D measurement has not become commonplace in vivo due to the diculty
of applying and accurately positioning large numbers of individual
electrodes. One possible solution for imaging objects close to the surface is
to employ a rigid rectangular array of electrodes. This is exactly the approach
taken by the TransScan device [100], which is designed for the detection of
breast tumour, although reconstructions are essentially what geophysicists
would call surface resistivity mapping, rather than tomographic reconstruction. Reconstruction of 3D EIT images from a rectangular array using
NOSER-like methods has been demonstrated in vitro by Mueller et al [103],
and in vivo on the human chest using individual electrodes [104]. If the array
is suciently small compared with the body, this problem becomes identical
to the geophysical EIT problem [98] using surface (rather than bore-hole)
electrodes.
The EIT problem is inherently nonlinear. There are of course two
aspects of linearity of a mapping: in engineering terminology, that the
output scales linearly with the input, and that the principle of superposition
applies. The lack of scaling invariance manifests itself in EIT as the
phenomenon of saturation, which means the linearity must be taken into
account to get accurate conductivity images. For small contrasts in conductivity, linear reconstruction algorithms will typically nd a few isolated small
objects, but underestimate their contrast. For more complex objects, even
with small contrasts the lack of the superposition property means that
linear algorithms cannot resolve some features. A simple test can be done
in a tank experiment. With two test objects with conductivity 1 and 2
one can test if Z1 Z2 Z1 2 within the accuracy of the
measurement system. If not then it is certainly worth using a nonlinear
56
References
57
58
[36] COMSOL 2000 The FEMLAB Reference Manual (Stockholm: COMSOL AB)
[37] Cheng K, Isaacson D, Newell J C and Gisser D G 1989 Electrode models for electric
current computed tomography IEEE Trans. Biomed. Eng. 36 918924
[38] Cheney M and Isaacson D 1992 Distinguishability in impedance imaging IEEE
Trans. Biomed. Eng. 39 852860
[39] Chung E T, Chan T F and Tai X C 2003 Electrical impedance tomography using
level set representation and total variational regularization, UCLA Computational
and Applied Mathematics Report 03-64
[40] Cook R D, Saulnier G J, Gisser D G, Goble J C, Newell J C and Isaacson D 1994
ACT 3: A high speed high precision electrical impedance tomograph IEEE Trans.
Biomed. Eng. 41 713722
[41] Coleman T F and Li Y 1992 A globally and quadratically convergent ane scaling
method for linear problems SIAM J. Optimization 3 609629
[42] Colin de Verdie`re Y, Gitler I and Vertigan D 1996 Reseaux electriques planaires II
Comment. Math. Helv. 71 144167
[43] Curtis E B and Morrow J A 2000 Inverse Problems for Electrical Networks, Series on
Applied Mathematics, Vol 13 (Singapore: World Scientic)
[44] Colton D and Kress R 1998 Inverse Acoustic and Electromagnetic Scattering Theory,
2nd edition (Berlin: Springer) p 51
[45] Ciulli S, Ispas S, Pidcock M K and Stroian A 2000 On a mixed NeumannRobin
boundary value problem in electrical impedance tomography Z. Angewandte
Math. Mech. 80 681696
[46] Demidenko E, Hartov A and Paulsen K 2004 Statistical estimation of resistance/
conductance by electrical impedance tomography measurements. Submitted to
IEEE Trans. Medical Imaging
[47] Demidenko E, Hartov A, Soni N and Paulsen K 2004 On optimal current patterns
for electrical impedance tomography. Submitted to IEEE Trans. Medical Imaging
[48] Dobson D C and Vogel C R 1997 Convergence of an iterative method for total variation denoising SIAM J. Numerical Analysis 43 17791791
[49] Dorn O, Miller E L and Rappaport C M 2000 A shape reconstruction method for
electromagnetic tomography using adjoint elds and level sets Inverse Problems 16
11191156
[50] Engl H W, Hanke M and Neubauer A 1996 Regularization of Inverse Problems
(Dordrecht: Kluwer)
[51] Eyuboglu B M and Pilkington T C 1993 Comment on distinguishability in electricalimpedance imaging IEEE Trans. Biomed. Eng. 40 13281330
[52] Fletcher R and Reeves C 1964 Function minimization by conjugate gradients
Computer J. 7 149154
[53] Folland G B 1995 Introduction to Partial Dierential Equations, 2nd edition (Princeton University Press)
[54] Fox C and Nicholls G 1997 Sampling conductivity images via MCMC, in The Art
and Science of Bayesian Image Analysis, ed K Mardia, R Ackroyd and C Gill,
Leeds Annual Statistics Research Workshop, University of Leeds, pp 91100
[55] George A and Liu J 1989 The evolution of the minimum degree ordering algorithm
SIAM Review 31 119
[56] Glidewell M E and Ng K T 1997 Anatomically constrained electrical impedance
tomography for three-dimensional anisotropic bodies IEEE Trans. Med. Imaging
16 572580
References
59
[57] Gong L, Zhang K Q and Unbehauen R 1997 3-D anisotropic electrical impedance
imaging IEEE Trans. Magnetics 33 21202122
[58] Gilbert J R, Moler C and Schreiber R 1992 Sparse matrices in MATLAB: design and
implementation SIAM J. Matrix Anal. Appl. 13 333356
[59] Gibson A P, Riley J, Schweiger M, Hebden J C, Arridge S R and Delpy D T 2003 A
method for generating patient-specic nite element meshes for head modelling
Phys. Med. Biol. 48 481495
[60] Gisser D G, Isaacson D and Newell J C 1987 Current topics in impedance imaging
Clin. Phys. Physiol. Meas. 8 Suppl A, 3946
[61] Gisser D G, Isaacson D and Newell J C 1990 Electric current computed tomography
and eigenvalues SIAM J. Appl. Math. 50 16231634
[62] Giusti E 1984 Minimal Surfaces and Functions of Bounded Variation (Birkhauser)
[63] Goble J and Isaacson D 1990 Fast reconstruction algorithms for three-dimensional
electrical impedance tomography Proc. IEEE-EMBS Conf. 12(1) 100101
[64] Goble J 1990 The three-dimensional inverse problem in electric current computed
tomography, PhD thesis, Rensselaer Polytechnic Institute, NY, USA
[65] Dobson D C and Santosa F 1994 An image enhancement technique for electrical
impedance tomography Inverse Problems 10 317334
[66] Golub G H and Van Loan C F 1996 Matrix Computations, 3rd edition (Baltimore,
MD: Johns Hopkins University Press)
[67] Greenleaf A and Uhlmann G 2001 Local uniqueness for the Dirichlet-to-Neumann
map via the two-plane transform Duke Math. J. 108 599617
[68] Haber E and Ascher U M 2001 Preconditioned all-at-once methods for large, sparse
parameter estimation problems Inverse Problems 17 18471864
[69] Hagger W W 2000 Iterative methods for nearly singular linear systems SIAM J. Sci.
Comput. 22 747766
[70] Hanke M 1995 Conjugate Gradient Type Methods for Ill-Posed Problems, Pitman
Research Notes in Mathematics (Harlow: Longman)
[71] Hanke M and Bruhl M 2003 Recent progress in electrical impedance tomography
Inverse Problems 19 S65S90
[72] Hansen P C 1998 Rank-Decient and Discrete Ill-Posed Problems: Numerical Aspects
of Linear Inversion (Philadelphia: SIAM)
[73] Hettlich F and Rundell W 1998 The determination of a discontinuity in a conductivity from a single boundary measurement Inverse Problems 14 6782
[74] Heikkinen L M, Vilhunen T, West R M and Vauhkonen M 2002 Simultaneous
reconstruction of electrode contact impedances and internal electrical properties:
II. Laboratory experiments Meas. Sci. Technol. 13 18551861
[75] Higham N J 1996 Accuracy and Stability of Numerical Algorithms (Philadelphia:
SIAM)
[76] Hoerl A E 1962 Application of ridge analysis to regression problems Chem. Eng.
Progress 58 5459
[77] Ingerman D and Morrow J A 1998 On a characterization of the kernel of the Dirichlet-to-Neumann map for a planar region SIAM J. Math. Anal. 29 106115
[78] Isaacson D 1986 Distinguishability of conductivities by electric-current computedtomography IEEE Trans. Med. Imaging 5 9195
[79] Isaacson, D, Newell J C, Goble J C and Cheney M 1990 Thoracic impedance images
during ventilation Proc. IEEE-EMBS Conf. 12(1) 106107
[80] Isakov V 1997 Inverse Problems for Partial Dierential Equations (Springer)
60
References
61
[103] Mueller J, Isaacson D and Newell J 1999 A reconstruction algorithm for electrical
impedance tomography data collected on rectangular electrode arrays IEEE
Trans. Biomed. Eng. 46 13791386
[104] Mueller J L, Isaacson D and Newell J C 2001 Reconstruction of conductivity
changes due to ventilation and perfusion from EIT data collected on a rectangular
electrode array Physiol. Meas. 22 97106
[105] Mueller J, Siltanen S and Isaacson D 2002 A direct reconstruction algorithm for
electrical impedance tomography IEEE Trans. Med. Imaging 21 555559
[106] McArdle F J, Suggett A J, Brown B H and Barber D C 1988 An assessment of
dynamic images by applied potential tomography for monitoring pulmonary
perfusion Clin. Phys. Physiol. Meas. 9(4A) 8791
[107] McCormick S F and Wade J G 1993 Multigrid solution of a linearized, regularized
least-squares problem in electrical impedance tomography Inverse Problems 9 697
713
[108] Molinari M, Cox S J, Blott B H and Daniell G J 2002 Comparison of algorithms
for non-linear inverse 3D electrical tomography reconstruction Physiol. Meas. 23
95104
[109] Molinari M 2003 High delity imaging in electrical impedance tomography, PhD
thesis, University of Southampton
[110] Marquardt D 1963 An algorithm for least squares estimation of nonlinear
parameters SIAM J. Appl. Math. 11 431441
[111] West R M, Ackroyd R G, Meng S and Williams R A 2004 Markov Chain Monte
Carlo techniques and spatial-temporal modelling for medical EIT Physiol. Meas.
25 181194
[112] Morrison D F 1983 Applied Linear Statistical Methods (Englewood Clis, NJ:
Prentice Hall)
[113] Natterer F 1982 The Mathematics of Computerized Tomography (Wiley)
[114] Newell J C, Blue R S, Isaacson D, Saulnier G J and Ross A S 2002 Phasic threedimensional impedance imaging of cardiac activity Physiol. Meas. 23 203209
[115] Nichols G and Fox C 1998 Prior modelling and posterior sampling in impedance
imaging. In Bayesian Inference for Inverse Problems, ed A Mohammad-Djafari,
Proc. SPIE 3459 116127
[116] Pain C C, Herwanger J V, Saunders J H, Worthington M H and de Oliveira C R E
2003 Anisotropic resistivity inversion Inverse Problems 19 10811111
[117] Paulson K, Breckon W and Pidcock M 1992 Electrode modeling in electricalimpedance tomography SIAM J. Appl. Math. 52 10121022
[118] Paulson K S, Lionheart W R B and Pidcock M K 1995 POMPUSan optimized
EIT reconstruction algorithm Inverse Problems 11 425437
[119] Perez-Juste Abascal J F 2003 The anisotropic inverse conductivity problem, MSc
thesis, University of Manchester
[120] Phillips D L 1962 A technique for the numerical solution of certain integral
equations of the rst kind J. Assoc. Comput. Mach. 9 8497
[121] Player M A, van Weereld J, Allen A R and Collie D A L 1999 Truncated-Newton
algorithm for three-dimensional electrical impedance tomography Electronics Lett.
35 21892191
[122] Polydorides N and Lionheart W R B 2002 A Matlab toolkit for three-dimensional
electrical impedance tomography: a contribution to the Electrical Impedance and
Diuse Optical Reconstruction Software project Meas. Sci. Technol. 13 18711883
62
[123] Polydorides N 2002 Image reconstruction algorithms for soft eld tomography, PhD
thesis, UMIST
[124] Polydorides N, Lionheart W R B and McCann H 2002 Krylov subspace itemacserative techniques: on the detection of brain activity with electrical impedance
tomography IEEE Trans. Med. Imaging 21 596603
[125] Ramachandran P 2004 The MayaVi Data Visualizer, http://mayavi.sourceforge.net
[126] Rondi L and Santosa F, Enhanced electrical impedance tomography via the
MumfordShah functional, preprint
[127] Rudin L I, Osher S and Fatemi E 1992 Nonlinear total variation based-noise
removal algorithms Physica D 60 259268
[128] Saad Y and Schultz M H 1986 GMRES: A generalized minimal residual algorithm
for solving nonsymmetric linear systems SIAM J. Sci. Statist. Comput. 7 856869
[129] Santosa F 1995 A level-set approach for inverse problems involving obstacles
ESAIM Control Optim. Calc. Var. 1 (1995/96) 1733
[130] Santosa F and Vogelius M 1991 A backprojection algorithm for electrical impedance
imaging SIAM J. Appl. Math. 50 216243
[131] Schappel B 2003 Electrical impedance tomography of the half space: locating
obstacles by electrostatic measurements on the boundary, in Proceedings of the
3rd World Congress on Industrial Process Tomography, Ban, Canada, 25 September, 788793
[132] Schoberl J 1997 NETGENAn advancing front 2D/3D-mesh generator based on
abstract rules Comput. Visual. Sci. 1 4152
[133] Seagar A D 1983 Probing with low frequency electric current, PhD thesis, University
of Canterbury, Christchurch, NZ
[134] Sikora J, Arridge S R, Bayford R H and Horesh L 2004 The application of hybrid
BEM/FEM methods to solve electrical impedance tomography forward problem for
the human head. Proc X ICEBI and V EIT, Gdansk, 2024 June 2004, eds Antoni
Nowakowski et al 503506
[135] Seo J K, Kwon O, Ammari H and Woo E J 2004 Mathematical framework and
lesion estimation algorithm for breast cancer detection: electrical impedance technique using TS2000 conguration. Preprint (accepted for IEEE Trans. Biomedical
Engineering)
[136] Siltanen S, Mueller J and Isaacson D 2000 An implementation of the reconstruction
algorithms of Nachman for the 2D inverse conductivity problem Inverse Problems 16
681699
[137] Shimada K and Gossard D C 1995 Bubble mesh: automated triangular meshing of
non-manifold geometry by sphere packing, in ACM Symposium on Solid Modeling
and Applications Archive. Proceedings of the third ACM Symposium on Solid
Modeling and Applications. Table of Contents. Salt Lake City, Utah, USA, 409419
[138] Silvester P P and Ferrari R L 1990 Finite Elements for Electrical Engineers
(Cambridge: Cambridge University Press)
[139] Somersalo E, Cheney M, Isaacson D and Isaacson E 1991 Layer stripping, a direct
numerical method for impedance imaging Inverse Problems 7 899926
[140] Somersalo E, Isaacson D and Cheney M 1992 A linearized inverse boundary value
problem for Maxwells equations J. Comput. Appl. Math. 42 123136
[141] Somersalo E, Kaipio J P, Vauhkonen M and Baroudi D 1997 Impedance imaging
and Markov chain Monte Carlo methods, in Proc. SPIE 42nd Annual Meeting,
175185
References
63
[142] Soleimani M and Powell C 2004 Black-box Algebraic Multigrid for the 3D Forward
Problem arising in Electrical Resistance Tomography, preprint
[143] Somersalo E, Cheney M and Isaacson D 1992 Existence and uniqueness for
electrode models for electric current computed tomography SIAM J. Appl. Math.
52 10231040
[144] Smallwood R D et al 1999 A comparison of neonatal and adult lung impedances
derived from EIT images Physiol. Meas. 20 401413
[145] Strang G 1988 Introduction to Linear Algebra, 3rd edition (WellesleyCambridge Press)
[146] Strang G and Fix G J 1973 An Analysis of the Finite Element Method (New York:
Prentice-Hall)
[147] Sylvester J and Uhlmann G 1986 A uniqueness theorem for an inverse boundary
value problem in electrical prospection Commun. Pure Appl. Math. 39 91112
[148] Tamburrino A and Rubinacci G 2002 A new non-iterative inversion method in
electrical resistance tomography Inverse Problems 18 2002
[149] Tarantola A 1987 Inverse Problem Theory (Elsevier)
[150] Tikhonov A N 1963 Solution of incorrectly formulated problems and the regularization method Soviet Math. Dokl. 4 10351038 (English translation of 1963 Dokl
Akad. Nauk. SSSR 151 501504)
[151] Lassas M, Taylor M and Uhlmann G 2003 The Dirichlet-to-Neumann map for
complete Riemannian manifolds with boundary Comm. Anal. Geom. 11 207222
[152] Vauhkonen M, Vadasz D, Karjalainen P A, Somersalo E and Kaipio J P 1998
Tikhonov regularization and prior information in electrical impedance tomography
IEEE Trans. Med. Imaging 19 285293
[153] Vauhkonen P J, Vauhkonen M, Savolainen T and Kaipio J P 1998 Static three
dimensional electrical impedance tomography, in Proceedings of ICEBI98, Barcelona,
Spain, 41
Vauhkonen P J, Vauhkonen M and Kaipio J P 2000 Errors due to the truncation of
the computational domain in static three-dimensional electrical impedance tomography Physiol. Meas. 21 125135
[154] Vauhkonen M, Karjalainen P A and Kaipio J P 1998 A Kalman lter approach to
track fast impedance changes in electrical impedance tomography IEEE Trans.
Biomed. Eng. 45 486493
[155] Vauhkonen M 1997 Electrical impedance tomography and prior information, PhD
thesis, University of Kuopio
[156] Vauhkonen P J 1999 Second order and innite elements in three-dimensional
electrical impedance tomography, Phil.Lic. thesis, Department of Applied Physics,
University of Kuopio, Finland, report series ISSN 0788-4672 report No. 2/99
[157] Vauhkonen M, Lionheart W R B, Heikkinen L M, Vauhkonen P J and Kaipio J P
2001 A Matlab package for the EIDORS project to reconstruct two-dimensional
EIT images Physiol. Meas. 22 107111
[158] Mitchell S A and Vavasis S A 2000 Quality mesh generation in higher dimensions
SIAM J. Comput. 29 13341370
[159] Mitchell S A and Vavasis S A 2000 Quality mesh generation in higher dimensions
SIAM J. Comput. 29 13341370
[160] Vogel C 2001 Computational Methods for Inverse Problems (Philadelphia: SIAM)
[161] Wade J G, Senior K and Seubert S 1996 Convergence of Derivative Approximations
in the Inverse Conductivity Problem, Bowling Green State University, Technical
Report No. 96-14
64
PART 2
EIT INSTRUMENTATION
Chapter 2
EIT instrumentation
Gary J Saulnier
2.1.
INTRODUCTION
Since the introduction of the rst systems in the early 1980s, EIT instrumentation has continued to evolve in step with advances in analogue and digital
electronics. While early instruments were designed using primarily analogue
techniques, newer instruments are shifting much of the processing to the
digital domain, making extensive use of digital signal processors and
programmable logic devices. Along with advances in technology have
come advances in system performance, particularly in the areas of system
bandwidth and precision. While the original systems used relatively low
frequency excitationgenerally in the 1020 kHz rangenewer systems
can apply waveforms up to the 110 MHz range. The ability to apply excitation signals over a signicant range of frequencies makes it possible to
perform impedance spectroscopy in which the variation of impedance with
frequency can be used as a discriminating factor for imaging. With this in
mind, some newer systems have been designed to acquire data at multiple
frequencies simultaneously.
This chapter discusses some of the general issues involved in the design
and implementation of the major functions required for EIT instrumentation. Some of these issues have also been discussed in several survey papers
[4, 26]. Later, the structure of several particular systems is discussed in detail.
2.2.
While there are many dierent EIT system designs, most systems apply
currents and measure voltages and can be classied according to the
number of current sourceseither as a single source system or a multiple
68
EIT instrumentation
Figure 2.1.
Signal generation
Figure 2.2.
69
2.3.
2.3.1.
SIGNAL GENERATION
Waveform synthesis
70
EIT instrumentation
Figure 2.3.
Signal generation
Figure 2.4.
frequency.
71
delivered to the load being independent of the load voltage, VL . Real current
sources, however, have a nite Z0 impedance that is usually characterized as
the parallel equivalent of a resistance R0 and capacitance C0 . Figure 2.5(a)
shows an ideal current source driving a load, where the load current IL
equals the source current IS . When a real current source drives a load, as
shown in gure 2.5(b), the current owing in Z0 varies with VL ; consequently, the relationship between IL and IS varies with the value of the
load impedance.
The variation in IL with VL that occurs with nite current source output
impedance is made worse by the presence of additional stray or parasitic
capacitances. Though not associated with the current source itself but,
rather, due to capacitance between wire and/or printed circuit board
Figure 2.5.
72
EIT instrumentation
Figure 2.6.
Signal generation
73
The current source in an EIT system must be able to deliver current with a
desired precision over a specied frequency range to load impedances
within an expected range of values. These requirements translate into
specications for the frequency response, output impedance and voltage
compliance of the current source. Both the voltage compliance and the
output impedance requirements are functions of the expected load
impedance. Since the voltage compliance of the source is the range of load
voltages for which the current source continues to behave as a current
source, it must exceed the voltage when the maximum current is sourced to
(or sinked from) the load with the highest impedance. In medical applications
with single sinusoid excitation, maximum peak current values in the range
0.15 mA are common, with smaller current values being used at lower
frequencies due to safety concerns. Load impedances, which are a function
of electrode size, excitation frequency and the tissue being imaged, typically
range from 100
to 10 k
, with the lower values observed at higher frequencies. With these currents and impedances, voltage compliance in the range of
a few volts is generally sucient.
The required output impedance is also a function of the load impedance.
However, there are two ways to look at the problem. In order to maintain a
desired accuracy of the applied current, i.e. keeping IL and IS of gure 2.5(b)
equal to within a given tolerance, it is necessary to consider the maximum
load impedance that the current source will encounter. The error current
equals the current through the output impedance of the source, IZ0 , which
is given by
IZ0
ZL max
I
Z0 ZL max S
where ZL max is the maximum load impedance and Z0 is the current source
output impedance. For the IL to be accurate to within b bits of precision
requires that the current error be less than one least signicant bit (LSB)
or, equivalently, 1=2b . The output impedance requirement then becomes
Z0 2b 1ZL max :
In this case, a system with 16 bit accuracy with a maximum load impedance
of 10 k
requires a current source with an output impedance of over 655 M
.
74
EIT instrumentation
Signal generation
Figure 2.7.
2.3.2.3.
75
76
EIT instrumentation
The common-mode current is the sum of the currents from the N sources.
The ideal current values sum to zero, making the common-mode current equal
to the sum of N independent noise sources. Since they are independent, the
power in the sum is N times the power in each source, i.e.
p
2 N 2
PCM N
:
12
12
From this equation it can be seen that in order to achieve PCM 2 =12, it is
necessary
to make the step size for the individual current sources equal
p
= N . Therefore, in order to achieve b bits of precision with respect to
the common-mode current, it is necessary to have
b0 b 0:5 log2 N
bits of precision for the individual sources. For a 64 electrode system with 16
bits of precision, the precision of each current source must be 19 bits.
2.3.2.4.
Stray capacitance, when in parallel with the output of the current source,
increases the eective output capacitance of the source and, consequently,
reduces the magnitude of the output impedance. Figure 2.7 shows the
required output impedance for a given precision and these values will now
be related to an allowable total capacitance at the current source output.
Figure 2.8 shows the capacitive reactance presented by capacitors of various
values as a function of frequency. To obtain even the modest output
impedance of 1 M
at approximately 20 kHz requires a total capacitance
of less than 10 pF. At 200 kHz, the allowable capacitance drops to 1 pF
and at 2 MHz it drops to 0.1 pF.
Figure 2.8.
Signal generation
2.3.2.5.
77
2pfCRL 2
2pfCRL
j
:
2
1 2pfCRL
1 2pfCRL 2
For the case where 2pfCRL < 1, the normalized imaginary (reactive) part of
the error exceeds the real part. Consider, for example, the case where
C 20 pF and RL 1 k
for which the real and reactive normalized error
voltages are plotted in gure 2.10 as a function of frequency. For 16 bits
of precision, the normalized error should be less than 2 16 15 10 6 .
In considering the real voltage only, the system can operate up to approximately 10 kHz with an error below this level. The error in the reactive voltage
is below this value only at very low frequencies. Note that, for these values of
C and RL , 2pfCRL exceeds unity for frequencies of approximately 8 MHz
and above where, on gure 2.10, the error for the real voltage moves
above that for the reactive voltage.
By measuring only the real part of the load voltage, it is not possible to
make images of the permittivity of the object. In order to achieve high precision while maintaining the ability to image both resistivity and permittivity, it
is necessary to employ techniques to either cancel the stray capacitance or
render it ineective.
Figure 2.9.
78
Figure 2.10.
EIT instrumentation
Signal generation
Figure 2.11.
79
current that is shunted away from the load can be calculated. Increasing the
applied current value to compensate for this current loss will result in the
desired current being applied to the load [27]. While the output impedance
and stray capacitance can be estimated using a calibration procedure, the
current through this impedance is a function of the load voltage, which
varies with the load impedance seen at the electrode as well as the applied
current. Consequently, this approach is necessarily iterative where currents
must be applied to determine the value of the load impedance and then
adjusted to compensate for shunt impedance [20].
2.3.3.
The current sources used for EIT are generally voltage-to-current converters,
producing a current that is proportional to an input voltage. This input must
be scaled appropriately to set the desired current amplitude. In cases where
the excitation waveform is distributed in analogue form, this scaling process
can be performed using a multiplying DAC (MDAC) as shown in gure 2.12.
The selected MDAC must perform 4-quadrant multiplication to enable both
positive and negative amplitude values. A problem with this approach is that
many MDACs, particularly those implemented using MOS technology,
introduce a code-dependent phase shift into the waveform, meaning that
the phase of the output waveform is somewhat dependent on the digital
current amplitude value. Bipolar MDACs, which do not have the same
phase-shift problem, typically perform only 2-quadrant multiplication and,
consequently, are unable to invert the waveform. A technique is described
in [6], which uses two bipolar MDACs and a high resolution audio DAC
to convert a digital waveform and digital amplitude control value into a
scaled analogue waveform without the phase-shift problem.
Another approach to producing the amplitude-scaled waveform is to
use a 4-quadrant analogue multiplier to multiply the analogue excitation
waveform by an analogue amplitude setting [3, 28]. A conventional DAC
can be used to convert a digital amplitude value into a d.c. signal. Analogue
multipliers, however, are often limited in bandwidth and dynamic range and,
also, introduce harmonic distortion into the signal [3].
An all-digital approach can also be used in which a digital excitation waveform is scaled before passing through the DAC. This approach overcomes
80
Figure 2.12.
EIT instrumentation
most of the limitations described above, though a higher resolution DAC may
be desirable in this case due to the larger dynamic range of the digital waveform. In a multiple-source system, however, this approach requires additional
digital processing on the individual channels.
2.3.4. Multiplexers
Multiplexers are required in single current source systems, as well as systems
that share voltmeters between multiple electrodes. These devices have many
non-ideal properties that make them undesirable in EIT systems, including a
nonzero on resistance that is somewhat dependent on the applied voltage,
limited o isolation, with lower values at high frequencies, and charge
injection during switching. The most signicant problem, however, is the
relatively large capacitance of multiplexer devices. Typically the input
capacitance is in the range 3050 pF and the output capacitance on each
line is in the range 510 pF. Multiplexers made using smaller devices will
have lower capacitance values at the cost of higher on resistance.
2.3.5. Current source and compensation circuits
Since they operate at relatively low frequencies, generally below 1 MHz,
EIT systems are able to use current sources that are built using operational
ampliers or transconductance ampliers. Current sources constructed
using these devices generally provide higher output impedance than simpler
sources constructed using discrete transistors, and have the capability to both
source and sink current. Here, a few of the current source circuits commonly
found in EIT instruments will be discussed.
Figure 2.13 shows a schematic diagram for a oating current source that
is commonly used in single source EIT systems. The transformer provides
d.c. isolation between the source and loadan important feature for patient
safety in medical applicationsand allows the load voltage to oat with
respect to ground potential. The voltage compliance and output impedance
Signal generation
Figure 2.13.
81
Vin VL =A
Ri
where is the current transfer ratio of the current mirrors and A is the open
loop gain of the operational amplier. An interesting property of this current
source is that it acts as an impedance multiplier. Assuming that the voltage
source driving the circuit is ideal, the output impedance of the current
82
Figure 2.14.
EIT instrumentation
Figure 2.15.
Signal generation
Figure 2.16.
83
output. High output impedances have been achieved using this current
source for frequencies in excess of 100 kHz.
The three-operational-amplier current source is shown in gure 2.16
[17]. This current source uses an inverting, summing voltage amplier in
the forward path, a current sensing resistor RS and a non-inverting buer
amplier and an inverting amplier in the feedback path. When the resistor
values are properly adjusted, the current in RS and the load is maintained at a
value that is proportional to Vin :
IL Vin RS :
The primary advantage of the three-operational-amplier source is that it
can provide a reasonably high output impedance when properly trimmed.
A primary disadvantage of the source is degraded performance due to
phase shifts in the feedback path at high frequencies. Other disadvantages
are the fact that trimming is required and the high component count in the
current source.
The Howland current source, shown in gure 2.17, is a single op amp
source that oers good performance [8]. The topology of the current
source has a forward path consisting of an inverting amplier (the op amp
along with R1 and R2 ) and positive feedback. An alternative implementation
of the Howland source uses an instrumentation amplier in place of the
inverting amplier in the circuit [6]. For an ideal op amp, the output
impedance of the source is innite when the resistors satisfy the relationship
R4 =R3 R2 =R1 :
At this balance condition the load current can be expressed as
IL Vin =R3 :
The primary advantages of the Howland source are its simplicity and ability
to produce a high output impedance with the appropriate trimming. In
84
Figure 2.17.
EIT instrumentation
Figure 2.18.
Signal generation
85
case where multiple frequencies are used one at a time, broadband compensation is desirable to avoid needing to retrim the source each time a new
frequency is used. However, in practice, the usefulness of the NIC is limited
by its tendency to oscillate. Stability can be improved by adding capacitance
to the resistive feedback network, but only at the cost of reducing the
frequency range over which the negative capacitance is produced.
The second compensation scheme is to create an LC resonant circuit by
introducing a parallel inductance [31]. This inductance can be synthesized
using a generalized impedance converter (GIC) circuit such as that shown
in gure 2.19 [22]. This circuit is one of several implementations of the
GIC. GICs are most commonly used to implement active lter equivalents
of RLC ladder lters.
The impedance seen looking into the GIC circuit is given by
Zin
Z1 Z3 Z5
:
Z2 Z4
R1 R3 R5 C4
sL:
R2
Figure 2.19.
86
EIT instrumentation
The GIC circuit exhibits good stability and component sensitivity properties. However, as described earlier, the eect of the capacitance is removed
only at the LC resonant frequency, meaning that this compensation
approach cannot be used in systems that apply multiple frequencies
simultaneously, and retuning must occur whenever the frequency is changed
in multi-frequency systems that apply a single frequency at a time.
2.3.6.
Cable shielding
In many EIT systems the electrodes are located at some distance from the
electronics and are connected using cables. Exceptions to this are the few
systems where the electrodes are closely coupled to the driving electronics
[14]. Coaxial or triaxial cables are used to connect the electrodes, as opposed
to individual wires, in order to minimize coupling between the signals to/
from each electrode and reduce the noise susceptibility. Due to their much
higher output impedance, current source outputs are much more susceptible
to noise pick-up than voltage source outputs and need protection.
While coaxial cables can provide the desired shielding, they typically
present a signicant distributed capacitance, on the order of 40100 pF/m.
In addition, the capacitance tends to vary, particularly with the exing of
the cable. Grounding the shield results in this capacitance acting as a
shunt to ground, much like the stray capacitance and current source
output capacitance. Instead, the shield is commonly driven with a voltage
that is nearly equal to that on the conductor as shown in gure 2.20. Now,
since the voltage across the capacitance is zero, it does not carry current
and is essentially removed from the circuit.
When triaxial cables are used, a second grounded shield is positioned
around the driven shield, providing added protection. The primary complication of using a driven shield is the potential for instability as the shield driver
amplier provides a positive feedback path. Additionally, the shield driver
amplier is typically presented with a highly capacitive load, making it less
stable. Maintaining the gain of the shield driver somewhat less than unity
minimizes the risk of oscillation due to positive feedback through the
signal conductor at the expense of increasing the residual cable capacitance.
Figure 2.20.
Driven shield.
Signal generation
Figure 2.21.
87
While a number of op amps are available that can drive large capacitive loads
at unity gain, the circuit shown in gure 2.21 is commonly used to enhance
the stability of the shield driver circuits. In this circuit, the combination of
the 100
series resistance and feedback capacitor allows negative feedback
that is less sensitive to the phase shift introduced by the capacitive load [23].
2.3.7.
Voltage sources
88
EIT instrumentation
Figure 2.22.
the measurement of the load current IL . Figure 2.22 shows the presence of
stray capacitance CS in parallel with the load. A load-voltage-dependent
current will ow in this stray capacitance, meaning that the current measured
through RS is not exactly equal to the load current. This problem is equivalent to the output capacitance/stray capacitance problem with a current
source. Once again, techniques for cancelling the capacitance could be
applied, although this would make the circuitry signicantly more complex,
removing one of the advantages of using voltage sources.
2.4.
VOLTAGE MEASUREMENT
Voltage measurement
89
shows an instrumentation amplier and its inputs and outputs. These inputs
can be expressed in terms of a dierential signal, VD V1 V2 , and a
common-mode signal, VCM V1 V2 =2. If the instrumentation amplier
is ideal, the common-mode gain is zero and the output is determined solely by
the dierential gain AD and the dierence between the input voltages
VO AD VD AD V1 V2 :
A real instrumentation amplier, however, will respond to both VD and VCM ,
and its output is given by
VO AD VD ACM VCM
where ACM is the common-mode gain. Figure 2.23(b) is a block diagram that
illustrates the behaviour of the instrumentation amplier. A gure of merit
for an instrumentation amplier is its common-mode rejection ratio
(CMRR) given by
CMRR 20 log10 jAD =ACM j:
While an ideal dierential amplier has a CMRR of innity, real instrumentation ampliers generally have a CMRR that is large at d.c. and drops with
increasing frequency. Typical CMRR values at d.c. are in the range 100
120 dB, while values at 1 MHz that are in the range 060 dB are common.
The common-mode rejection of an instrumentation amplier is
degraded when there is an imbalance between the driving impedances for
each input. Figure 2.24 shows an instrumentation amplier with capacitors
Ci representing its input capacitance. A common-mode voltage is applied
through unequal resistances, R1 and R2 . The impact of the unequal driving
resistances is that the common mode input signal produces a dierential
voltage between the inputs to the instrumentation amplier. This dierential
voltage is then multiplied by the dierential gain of the amplier to produce
and output, even if the common-mode gain of the instrumentation amplier
90
EIT instrumentation
Figure 2.24.
EIT systems that image both the conductivity and permittivity in the body
require phase-sensitive voltage measurements, i.e. measurement of both the
real and reactive voltages on the electrodes. Likewise, systems that assume
that the load is resistive require phase-sensitive voltage measurements in
order to extract the real part of the electrode voltage. As discussed earlier,
measuring the magnitude of the electrode voltage would result in greater
sensitivity to stray capacitance. These phase-sensitive measurements are
generally made using a synchronous voltmeter that uses a coherent reference
obtained from the system waveform generator. While early systems
performed synchronous voltage measurement using analogue circuitry,
Voltage measurement
Figure 2.25.
91
most newer EIT systems take a digital approach. A discussion of both the
analogue and digital approaches to phase-sensitive voltmetering is found
in [18].
An analogue implementation of a phase-sensitive voltmeter is shown in
gure 2.25. A reference square wave having the exact frequency as the input
sinusoidal waveform is used to control a switch that alternately applies
non-inverted and inverted versions of the input signal to a lowpass lter.
Generally, the square wave is supplied by the waveform synthesis block,
which also produces the system excitation waveform, to ensure that the
frequencies of the two signals are the same. The relative phase of the
reference signal determines whether the voltmeter measures the real voltage,
reactive voltage, or a combination of the two. Adjusting the reference phase
to maximize the output with a resistive load can be used to determine the set
of appropriate reference waveform phases to measure the real voltage. The
lowpass lter ideally retains only the d.c. component of the signal, which is
proportional to the sum of the input voltage waveform components that
are at the signal frequency and its odd harmonics.
The analogue synchronous voltmeter of gure 2.25 essentially mixes the
input signal with a square wave of the same frequency and keeps the d.c.
portion of the result. Integrated circuits such as the Analog Devices
AD630 are available to perform this operation. This analogue voltmeter
has several drawbacks, however. First, the output is sensitive to odd harmonics in the input signal, making it necessary to maintain spectral purity
through the system. Second, the lowpass lter provides limited rejection of
the non-d.c. components in its input signal, reducing the overall precision
of the system. A high-order lowpass lter may be required to achieve a
high degree of measurement precision. Finally, the structure is sub-optimal
with regard to additive broadband noise that may be present in the input
signal.
92
Figure 2.26.
EIT instrumentation
The limitations of the voltmeter in gure 2.25 are due to the limitations
of the lowpass lter and the fact that the reference waveform is a square wave
rather than a sinusoid. While a more complex analogue voltmeter with better
performance could be implemented, generally a digital approach is used
instead. Figure 2.26 is a block diagram of a digital implementation of a
phase-sensitive voltmeter that produces both real and reactive measurements. The voltage is sampled and quantized by the ADC, and the samples
are multiplied by sine and cosine reference waveforms of exactly the same
frequency. The products are subsequently accumulated over an integral
number of cycles of the signal frequency. For the system to work properly,
the sampling clock for the ADC must have the necessary relationship to
the signal frequency. This voltmeter structure is equivalent to a matched
lter used in the detection of communication signals, and it can be shown
that the SNR of the measured voltages is optimal for a given ADC precision
and integration period if the noise in the signal after the ADC is white, meaning that it has a at (frequency independent) power spectral density. Real and
reactive outputs in gure 2.26 are labelled, assuming that a real (resistive)
load produces a voltage waveform that is a cosine having a phase angle of
zero.
It is necessary to integrate over an integral number of cycles of the signal
in order to suppress the double-frequency components of the product of the
ADC samples and the reference sine and cosine. Essentially, multiplying two
sinusoids having the same frequency produces a result that consists of a d.c.
signal, having an amplitude that is dependent on the amplitudes of the
individual sinusoids and their relative phase, plus a sinusoid having double
the original frequency. Integrating over an integral number of periods of
the input signal frequency completely suppresses this double frequency and
all other harmonics of the excitation frequency, because the integration
lter has a frequency response with a j sin x=xj shape centred at d.c. and
nulls at frequencies k=T, where T is the integration period and k is any
Voltage measurement
93
integer not equal to zero. When T N=f , where f equals the signal
frequency, the nulls are at kf =N.
2.4.4.
Noise performance
2
12
where is the ADC quantization step size. Increasing the precision of the
ADC by one bit results in a reduction of by a factor of 2 and a corresponding decrease in the quantization noise power by a factor of 4. Using the
assumption that this noise is white, the power is uniformly distributed over
a bandwidth of fS Hz, where fS is the sampling frequency, resulting in a
noise power spectral density of
PSD
2
:
12fS
94
EIT instrumentation
95
of the ADC be suciently wide to pass the excitation frequency, and its
aperture jitter be suciently small to avoid loss of ADC precision due to
timing uncertainty.
2.5.
There are a wide variety of EIT instruments that have been designed and
built with varying degrees of success in solving the basic problemthat of
determining the impedance distribution within a body from measurements
made on its surface. Probably the most important characteristic of each
instrument is whether it is a single-source system or a multiple-source
system. The choice of which type of instrument to build is fundamentally
one of complexity versus performance, with a single-source system having
much simpler hardware and a multiple-source system having, in theory,
better performance. A few systems of each type are described below.
96
EIT instrumentation
Sheeld systems
The most widely used EIT systems are the 16-electrode mark 1 and mark 2
single source systems developed at Sheeld [11, 12]. While the mark 1 and
mark 2 are both single frequency systems, this group has also developed
multiple frequency systems. The mark 3 system can apply eight frequencies
in the range 9.6 kHz to 1.2 MHz, with a single frequency being applied at a
given time. The mark 3.5 system applies 30 frequencies in the range 2 kHz
to 1.6 MHz simultaneously, using an FFT-based digital voltage measurement system [13]. The mark 3 system uses separate drive and receive
electrodes (eight of each), while the mark 3.5 system uses a total of eight electrodes. These systems all provide real-time imaging at roughly 25 images/s.
The mark 2 system [11] operates with a digitally-generated sinusoidal
excitation signal of 20.83 kHz, which is produced using a 12-bit DAC and
a 48-entry ROM look-up table clocked at 1 MHz. The applied current is
produced using a oating-load voltage-to-current converter like that
shown in gure 2.13. Direct measurement of the applied current, performed
using an in-line resistor and an instrumentation amplier, is used to account
for the presence of variations in phase and amplitude of the applied current
with variations in the load impedance at the electrodes. Two 1-to-16 multiplexers (Analog Devices DG506) are used to direct the currents to a single
pair of electrodes at a given time. A current amplitude of 5 mA peak-topeak is used.
Dierential voltage measurements are made between adjacent pairs of
electrodes. The electrode voltages are a.c.-coupled to a set of 16 instrumentation ampliers (Burr-Brown INA110), providing parallel measurement of all
the dierential voltages. The instrumentation amplier outputs are transformer-coupled to programmable-gain ampliers (PGAs), with gains from
1 to 256 in powers of 2. PGA output voltages are processed by synchronous,
phase-sensitive voltmeters. Only the real component of the measured
voltages is used in image reconstruction due to the greater impact of stray
capacitance on the accuracy of the reactive measurements.
A common-mode feedback circuit is used to reduce the common-mode
voltage applied to the instrumentation ampliers in the voltage measurement
circuit. Since all dierential voltages are measured simultaneously, the
common-mode voltage cannot be minimized for all voltage measurements
but, rather, the circuit reduces the common-mode voltage seen by all instrumentation ampliers. The circuit works using a pair of electrodes located
away from the electrodes used to collect image data. One electrode is used
to sense the common-mode voltage and the second electrode is driven with
a compensating voltage which acts to drive the common-mode voltage to
zero. The gain of the feedback loop must be kept suciently low (32 dB) in
order to avoid oscillation problems.
97
The Sheeld APT systems are the most widely used EIT systemsthe
hardware is compact and reliable and capable of producing real-time
images. The instrumentation has been well designed and its performance is
well documented. The systems have been optimized for obtaining the best
data available in the single current source conguration. However, the
system is ultimately limited by the need for multiplexers to switch the current
source between electrode pairs and the signicant shunting capacitance that
they introduce. While the problem is partially mitigated by using only the
measured real voltages, the penalty is an inability to image the reactive
component of the impedance.
2.5.1.2.
98
EIT instrumentation
result, the d.c. potential due to the electrode/patient interface appears at the
amplier input. The system utilizes a compensation system in which a DAC
drives the bias adjustment on the instrumentation amplier to compensate
for the contact potential. This correction is performed for each electrode
prior to the measurement of the a.c. voltage due to the applied current.
The instrumentation amplier output, after lowpass ltering, is sampled
and quantized by a 14-bit ADC, and digital synchronous detection is used
to measure the real part of the electrode voltage.
As a single source system, the system is limited by the stray capacitance
introduced by the multiplexers, ultimately limiting the excitation frequency
to approximately 50 kHz and not allowing measurement of permittivity.
Also, the system trades o real-time performance for a large number of
electrodes that, in theory, should provide improved image resolution.
However, resolution is a function of both the number of electrodes and the
measurement precision, and the limited measurement precision of the instrumentation may make it impossible to realize the resolution improvement
anticipated by using 256 electrodes.
2.5.2.
2.5.2.1.
Multiple-source systems
Oxford Brookes systems
99
Dartmouth systems
This group has developed multiple-source systems for breast cancer detection
that incorporate both current and voltage sources. A recent system,
described in [3], supports 32 electrodes with a continuously selectable excitation frequency in the range from 1 kHz to 1 MHz. The waveform is generated
using a PC-based arbitrary waveform generation board (Datel PC-420) that
generates waveforms using a 12-bit DAC with a maximum sampling rate of
40 MHz. This waveform is distributed, in analogue form, to custom boards
that support eight electrodes each. The system rack can accommodate up
to 16 boards (128 electrodes) and the design has address space for up to
256 boards (2048 electrodes).
The system contains 32 voltage sources and 32 current sources, enabling
it to apply either voltages or currents to the electrodes. The current sources
are implemented using an OTA (Burr-Brown OPA2662), while the voltage
sources are implemented with unity gain operational amplier buers with
a current sensing resistor in the feedback loop. A current sensing resistor is
also used to enable direct measurement of the applied currents when the
current sources are being used.
The amplitude of the sinusoidal voltages feeding the OTAs and voltage
buers determine the amplitude of the applied signals. The analogue reference waveform is scaled at each channel using an analogue multiplier
100
EIT instrumentation
101
2.6.
This chapter has reviewed various approaches for implementing the major
components of an EIT system and discussed some of the advantages and
disadvantages of each approach. A few example systems were presented to
show how these components have been combined to produce EIT instruments. An unresolved question, however, is how should one design the
best EIT system for a given application? The answer is not always clear
and may vary with the constraints presented by the application.
What is clear is that, for a given number of electrodes, the best data for
making images comes from an instrument with the highest possible precision
and multiple sources. Such a system is also the most complex and expensive
102
EIT instrumentation
References
103
REFERENCES
[1] Isaacson D 1986 Distinguishability of conductivities by electric current computed
tomography IEEE Trans. Med. Imaging MI-5 9295
[2] Zhu Z, Lionheart W R B, Lidgey F J, McLeod C N, Paulson K S and Pidcock M K
1993 An adaptive current tomography using voltage sources IEEE Trans. Biomedical
Engineering 40(2) 163168
[3] Hartov A, Mazzarese R A, Reiss F R, Kerner T E, Osterman K S, Williams D B and
Paulsen K D 2000 A multichannel continuously selectable multifrequency electrical
impedance spectroscopy measurement system IEEE Trans. Biomedical Engineering
47(1) 4958
[4] Murphy D and Rolfe P 1988 Aspects of instrumentation design for impedance
imaging Clin. Phys. Physiol. Meas. 9 Suppl. A 514
[5] Cherepenin V, Karpov A, Korjenevsky A, Kornienko V, Kultiasov Y, Mazaletskaya
A and Mazourov D 2002 Preliminary static EIT images of the thorax in health and
disease Physiol. Meas. 23 3341
[6] Cook R D, Saulnier G J, Gisser D G, Goble J C, Newell J C and Isaacson D 1994
ACT3: A high-speed, high-precision electrical impedance tomograph IEEE Trans.
Biomedical Engineering 41(8) 713722
[7] Edic P M, Saulnier G J, Newell J C and Isaacson D 1995 A real-time electrical
impedance tomograph, IEEE Trans. Biomed. Eng. 42(9) 849859
[8] Franco S 1988 Design with Operational Ampliers and Analog Integrated Circuits
(McGraw-Hill) 5864
[9] Van Valkenburg M E 1982 Analog Filter Design (Holt, Rinehart and Winston) 441
442
[10] Gentile K 1999 The eect of DAC resolution on spurious performance, A technical
tutorial on digital signal synthesis (Analog Devices, Inc)
[11] Smith R W M, Freeston I L and Brown B H 1995 A real-time electrical impedance
tomography system for clinical usedesign and preliminary results IEEE Trans.
Biomed. Eng. 42(2) 133140
[12] Brown B H and Seagar A D 1987 The Sheeld data collection system Clin. Phys.
Physiol. Meas. 8 Suppl. A 9197
[13] Wilson A J, Milnes P, Waterworth A R, Smallwood R H and Brown B H 2001 Mk3.5:
a modular, multi-frequency successor to the Mk3a EIS/EIT system Physiol. Meas. 22
4954
[14] Cherepenin V, Karpov A, Korjenevsky A, Kornienko V, Mazaletskaya A, Mazourov
D and Meister D 2001 A 3D electrical impedance tomography (EIT) system for breast
cancer detection Physiol. Meas. 22 918
104
EIT instrumentation
PART 3
APPLICATIONS
Chapter 3
Imaging of the thorax by EIT
H J Smit, A Vonk Noordegraaf, H R van Genderingen
and P W A Kunst
3.1.
GENERAL INTRODUCTION
3.2.
3.2.1.
EQUIPMENT
Sheeld mark 1 system
Most studies have been performed by using the Sheeld Applied Potential
Tomograph mark 1, developed by Barber and Brown in the 1980s [7], and
its successor the Sheeld APT DAS-01P. Sixteen electrodes are placed
108
equidistantly around the thorax and one earth electrode is placed on the
abdomen. Current is injected at 50 kHz sequentially in adjacent electrode
pairs and the potential dierence is measured in the remaining electrode
pairs (gure 3.1).
Eorts to reconstruct images of absolute impedance distribution have
not so far led to satisfactory results. Therefore, dynamic images are produced
showing the distribution of relative impedance changes. This is done by
feeding voltage changes relative to a reference data set into the Sheeld
back-projection algorithm [8]. The reference data must be obtained from
the same subject to produce reliable results.
The spatial resolution of the system was estimated to be approximately
10% of the array diameter [9]. To obtain adequate noise reduction, special
averaging techniques were required. For cardiac and circulatory application
the method involves ECG-triggered averaging [10], yielding a time-series of
EIT images during a single heart beat from a set of at least 100 heart
beats. The temporal resolution is 0.04 s (25 Hz). For ventilatory applications,
a number of acquisition cycles are averaged leading to sample rates around
0.9 Hz. This temporal resolution is insucient to monitor tidal changes
with great accuracy, but enables the measurement of slow variations in
lung volume. By dening one or more regions of interest (ROI) in the EIT
image, local or regional time-series of relative impedance change can be
determined, which can be used to quantify the observed physiological
phenomena (gure 3.2). In addition, a so-called functional EIT (fEIT) can
be created, an image consisting of pixels that represents the time variation
Equipment
109
Figure 3.2. Regional analysis of a sequence of electrical impedance tomograms. The timecourse of the ventral impedance change (upper panel) during stepwise lung ination is
signicantly dierent from the dorsal pattern (lower panel).
of the local impedance change (gure 3.3). The fEIT analysis was not
included in the original Sheeld device, but in a later stage proposed by
Hahn et al [11].
3.2.2.
Newer systems
One of the successors of the Sheeld mark 1 is the Sheeld mark 3.5,
marketed by Maltron Inc. as the Pulmonary Scan mark 3.5. It is a multifrequency, eight-electrode system, specically designed for neonatal use, where
the space available for electrodes is limited. It operates on frequencies in the
range between 2 kHz and 1.6 MHz, which may enable tissue characterization
in future. Data collection speed is 25 frames/s. Signal-to-noise ratio was
markedly reduced in comparison with the mark 1. A number of other experimental EIT devices have been developed over the years. Recently, the
University of Gottingen group has developed the GoeMF II EIT system, a
multifrequency device with an acquisition rate of 1344 Hz. In essence, it
110
Figure 3.3. Functional electrical impedance tomogram (fEIT) recorded during stable
mechanical ventilation. The image is constructed by calculating the standard deviation
over time in each picture element. The two ventilated lungs are clearly visible in white
(large variation); the white spot in the middle is the heart.
3.3.
3.3.1.
CARDIAC IMAGING
Introduction
McArdle et al showed for the rst time that EIT is able to localize the
impedance variations occurring during the cardiac cycle [13]. Imaging of
the heart by means of EIT is based on the principle that measured impedance
changes are caused by changes in blood volume. Since the blood volume
changes in the ventricles and atria are opposite to each other during the
cardiac cycle, this technique makes it possible to visualize ventricular and
atrial impedance related blood volume changes. Data collection can be
synchronized with the R-wave of the electrocardiogram, making it possible
to average more than one cardiac cycle in order to obtain an optimal data
set without respiratory artefacts.
3.3.2.
Electrode positioning
Most of the studies which have been performed in the eld of cardiac imaging
used the Sheeld DAS-01 P EIT system. The problems involved in cardiac
imaging by means of EIT are twofold. First, the volume changes in the
heart during the cardiac cycle are complex, with the heart moving through
a transversal plane. Second, the spatial resolution of the system is poor.
Therefore, the attachment of the electrodes for the EIT measurements is
Cardiac imaging
111
Figure 3.4. Variations of cross-sectional areas in MRI images (upper curves) and
impedance in EIT images (lower curves) for the ventricles (rst column) and atria
(second column) during the cardiac cycle. The value of line A can be used as a value of
stroke volume.
112
3.3.3.
In the ventricular region, impedance increases during systole as a consequence of blood outow, whereas impedance in the atrial regions decreases
due to lling of the atria. Since the electrical current ow is not planar,
these images represent impedance changes several centimetres above and
below the electrode plane [16]. Furthermore, an earlier study showed that
the impedance changes as measured by means of EIT are proportionally
related to blood volume changes [17]. Based on these ndings, a study was
performed to investigate whether the peak systolic impedance change in
the ventricular region, which was dened automatically on the EIT images,
corresponds with stroke volume [18]. In a group of 26 patients scheduled
for right heart catheterization, stroke volume was assessed by means of the
thermodilution method during catheterization and compared with the EIT
measurement made within 2 h after the catheterization. The correlation
coecient between peak systolic impedance changes and stroke volume
was 0.63 in this study, although a much better relationship could be obtained
by taking the time of the cardiac cycle into account (r 0:86). Although this
study showed that EIT measurements at this level of the thorax and stroke
volume are related to each other, the weak correlation and large spread of
the EIT values indicate that EIT cannot replace the invasive techniques for
the measurement of stroke volume. Several arguments can be put forward
to explain this weak correlation. First, MRI studies revealed that, even by
using the long axis plane, ventricular and atrial regions cannot be dened
as a xed anatomical region in the thoracic cavity, since ventricles will replace
the atria and vice versa during the dynamic process of cardiac contraction.
For this reason, impedance changes in the ventricular and atrial region will
inuence each other to a great extent. Furthermore, the inuence of possible
confounding variables such as thoracic wall thickness, dierent positions of
the heart and the inuence of valvular diseases might further disturb the
relationship between EIT measures and stroke volume.
3.3.4.
113
impedance changes of the right atrium over time. Since the diastolic function
of the right ventricle is dened as an index of early and late diastolic lling,
we investigated whether the corresponding impedance changes in the early
and late diastolic phase provide a measure for the right ventricular function.
In a group of COPD patients (characterized by persistent air ow limitation
and destruction of lung parenchyma) and healthy controls the correlation
between MRI and EIT measurements of right ventricular diastolic function
was 0.78 [20]. Since right ventricular diastolic function is closely related
to pulmonary artery pressure, the relationship between right ventricular
diastolic function measured by EIT and pulmonary artery pressure was
investigated in the same study in a group of 27 patients. This showed that
pulmonary artery pressure was closely related to the lling characteristics
of the right ventricle as measured by EIT (r 0:78).
3.3.5.
Summary
3.4.
3.4.1.
The capacity of EIT to detect systolic blood volume changes in the lungs
oers the possibility of studying the pulmonary perfusion. Eyubogu et al
(1987) showed that ECG-gated dynamic EIT images of the thorax could
be performed; these represented thoracic impedance changes related to
cardiac activity [21]. Shortly afterwards, McArdle et al showed that, by
means of cardiac-gated EIT, pulmonary perfusion can be visualized by
means of this technique [22]. However, the quality of those images was
poor as a consequence of the relatively small changes in the resistivity of
the lungs due to pulmonary perfusion, in the presence of noise, and the
larger resistivity changes due to the ventilation [23]. Image quality could be
improved by multiple time averaging of cardiac-gated data, enabling separation of the perfusion-related impedance changes from the ventilation
inuence. The required number of data frames for this type of processing
is at least 100 cardiac cycles [22, 24, 25].
114
Many pulmonary diseases involve the vessels of the pulmonary vascular bed.
Since the small pulmonary vascular bed is mainly responsible for blood
volume and thus impedance changes, EIT might be of value in the diagnosis
of diseases of the small pulmonary blood vessels. The most common disease
involving the pulmonary vascular bed is chronic obstructive pulmonary
115
disease (COPD), especially the lung emphysema type. This disease is not only
accompanied by a loss of the alveolar wall, but also by a signicant reduction
of the small pulmonary blood vessels. The rst clinical study investigating the
possibilities of EIT to detect the pathological changes of the pulmonary
vascular bed of these patients was performed by Vonk Noordegraaf et al
[30]. They found that in emphysematous patients, cardiac-gated lung
impedance changes are signicantly smaller in comparison with healthy
subjects. To test the hypothesis that indeed the small pulmonary vascular
bed is responsible for the EIT signal, the eects of vasoconstriction and
vasodilation of the small pulmonary blood vessels in a group of healthy
subjects and COPD patients were studied. Pulmonary vasoconstriction was
induced in healthy subjects by inhaling hypoxic air (14% oxygen), causing
a reduction of the EIT signal (gure 3.5). Pulmonary vasodilation was
Figure 3.5. Upper image: systolic related impedance changes (Zsys ) when seven healthy
subjects were breathing room air and 100% oxygen (N.S.). Same conditions for six emphysema patients, indicating release of hypoxic pulmonary vasoconstriction (HPV) in these
patients, detected by EIT (P < 0:05). Lower image: systolic related impedance changes
when seven healthy subjects were breathing room air and 14% oxygen. Induction of
HPV can by detected by EIT (P < 0:05).
116
studied in six emphysematous patients with active hypoxic pulmonary vasoconstriction. By inhaling 100% oxygen, release of hypoxic vasoconstriction
could be obtained in the patients. EIT measurements were performed
while breathing room air, and during hyperoxia. There was indeed a signicant increase in impedance changes during 100% oxygen, whereas stroke
volume and heart rate remained unchanged. These experiments indicate
that EIT is a sensitive method for detecting relaxation of hypoxic pulmonary
vasoconstriction [31]. The clinical importance of a non-invasive tool to
measure the presence of hypoxic pulmonary vasoconstriction can be
illustrated by a study conducted by Ashutosh et al [32]. In their study, 28
emphysematous patients received oxygen. They were able to divide those
patients into a responding group and a non-responding group, in which
response was dened as a minimal fall in the mean pulmonary artery pressure
of 5 mm Hg. After catheterization, all subjects were prescribed supplemental
oxygen. The authors reported a strong two-year survival benet and
improvement of quality of life in the responding group. Moreover, there
was no improvement in mortality in the non-responding group in comparison with patients who had not been treated with long-term domiciliary
oxygen therapy. So, it is important to select the COPD patients who are
still in a reversible stage, as only those patients will benet from long-term
oxygen therapy. EIT might be a suitable technique for selecting those
patients in a non-invasive way.
3.4.3.2.
A second disease in which the application of EIT has been studied is pulmonary arterial hypertension (PAH), characterized by elevated blood pressure in
the pulmonary arteries, due to obliteration of small pulmonary arterial
branches, caused by intima thickening, media hypertrophy and thrombosis
in the small vessels. PAH is a rare disease of the pulmonary vascular bed
that mainly aects young adults (mean age at diagnosis is 36 years), with a
preference for women [33, 34]. The earliest symptom in many cases of
PAH is the gradual onset of shortness of breath after physical exertion.
This shortness of breath is non-specic and is frequently ascribed to a lack
of physical tness. Thus, diagnosis of PAH is commonly delayed, sometimes
for more than two years after the onset of symptoms. Early diagnosis makes
it possible to start therapy at an earlier stage, before the pulmonary vessels
have already been irreversibly obliterated. Until now, the diagnosis of
pulmonary hypertension can only be assessed invasively. Recent studies
showed a low sensitivity and specicity of echo Doppler in the diagnosis of
pulmonary hypertension [35, 36]. Since an early diagnosis of pulmonary
hypertension might alter the course of this fatal disease, it is worthwhile to
test the diagnostic value of EIT for the diagnosis of pulmonary hypertension
in a large group of patients at risk of pulmonary hypertension. As the
117
Summary
3.5.
3.5.1.
118
119
Figure 3.6. Pressureimpedance curves with increasing severity of acute lung injury
(ALI). H, in healthy lungs of a pig; L1L3, after respectively one, two and three lung
lavages with saline; A, the anterior part of the lungs (non-dependent); P, the posterior
part of the lungs (dependent). Note that with increasing severity of ALI, higher pressures
are needed to open up the lung.
used. They found high correlation coecients between 0.81 and 0.93,
showing that local impedance changes were closely related to local changes
in air content. In mechanically ventilated critical care patients, Hinz et al
[51] compared end-expiratory lung impedance changes (ELIC), using the
Gottingen tomograph GoeMF to end-expiratory lung volume changes
(EELV) by open-circuit nitrogen washout. They found a linear correlation
according to the equation ELIC 0:98 EELV 0:68 with r2 0:95, and
concluded that EIT can be used as a bedside technique to monitor lung
volume changes during ventilatory manoeuvres.
Van Genderingen et al [52] elaborated further on the regional PV
observations by Kunst, by assessing the impedance change both during
lung ination and deation in lung-injured pigs. Using EIT, they found a
120
121
122
Figure 3.10. Theoretic eects of dierent electrode positioning when the cross-section of
the body has a trapezoid shape (right). Using the standard electrode positioning,
impedance changes are projected over an electrical impedance tomogram (right), causing
deformation of lung areas. The result may be over-representation of the left lower lobe area
LLL in the EIT image. In the test positioning the mid-electrodes 5 and 13 were moved 3 cm
in the ventral direction. Electrodes 15 have a shorter inter-electrode distance than
electrodes 59. The authors [54] hypothesize that this repositioning will decrease the overrepresentation of area LLL.
3.5.3.
Future directions
Only recently, the medical profession has picked up interest in EIT to determine regional lung function, and at present a number of clinical studies are
being undertaken. In the future, EIT requires further validation, preferably
in patients in comparison with CT as the gold standard. The method
should be further optimized and standardized as follows:
1. Electrode positioning, i.e. the level on the thorax and inter-electrode
distance, needs optimizing and standardizing.
2. The role of the reference data set should be further explored. That is, do
we need to acquire the data set in a certain physiological state to obtain
reliable impedance data.
References
123
EIT has now been under investigation for about 20 years, but the nal step to
routine clinical use has still not been made. EIT must still be regarded as a
research technique. Much eort over the past years has been put into
improvements of the technology. Validation studies have been published,
EIT can be used to analyse physiological phenomena in the lungs, and in
recent years more and more patient-related research has been conducted.
The most promising elds for the clinical application of EIT are in our
opinion the measurement of the characteristics of the pulmonary vascular
bed for the diagnosis of pulmonary hypertension and regional lung function,
in order to determine the optimal airway pressures for articial ventilation.
REFERENCES
[1] Kotre C J 1997 Electrical impedance tomography Br. J. Radiol. 70 Spec No: S200S205
[2] Boone K G and Holder D S 1996 Current approaches to analogue instrumentation
design in electrical impedance tomography Physiol. Meas. 17(4) 229247
[3] Morucci J P and Rigaud B 1996 Bioelectrical impedance techniques in medicine. Part
III: Impedance imaging. Third section: medical applications Crit. Rev. Biomed. Eng.
24(46) 655677.
[4] Brown B H 2003 Electrical impedance tomography (EIT): a review J. Med. Eng.
Technol. 27(3) 97108
[5] Frerichs I 2000 Electrical impedance tomography (EIT) in applications related to lung
and ventilation: a review of experimental and clinical activities Physiol. Meas. 21(2)
R1R21
[6] Dijkstra A M, Brown B H, Leathard A D, Harris N D, Barber D C and Edbrooke D L
1993 Clinical applications of electrical impedance tomography J. Med. Eng. Technol.
17(3) 8998
124
[7] Barber D C and Brown B H 1984 Applied potential tomography J. Phys. E: Sci.
Instrum. 17 723733
[8] Barber D C 1989 A review of image reconstruction techniques for electrical
impedance tomography Med. Phys. 16(2) 162169
[9] Brown B H and Barber D C 1987 Electrical impedance tomography; the construction
and application to physiological measurement of electrical impedance images Med.
Prog. Technol. 13(2) 6975
[10] Eyuboglu B M, Brown B H, Barber D C and Seagar A D 1987 Localisation of cardiac
related impedance changes in the thorax Clin. Phys. Physiol. Meas. 8 Suppl A 167
173
[11] Hahn G, Sipinkova I, Baisch F and Hellige G 1995 Changes in the thoracic
impedance distribution under dierent ventilatory conditions Physiol. Meas. 16(3)
Suppl A A161A173
[12] Hahn G et al 2001 Quantitative evaluation of the performance of dierent electrical
tomography devices Biomed. Tech. (Berl.) 46(4) 9195
[13] McArdle F J, Brown B H, Pearse R G and Barber D C 1988 The eect of the skull of
low-birthweight neonates on applied potential tomography imaging of centralised
resistivity changes Clin. Phys. Physiol. Meas. 9 Suppl. A 5560
[14] Patterson R P, Zhang J, Mason L I and Jerosch-Herold M 2001 Variability in the
cardiac EIT image as a function of electrode position, lung volume and body position
Physiol. Meas. 22(1) 159166
[15] Vonk Noordegraaf A et al 1996 Improvement of cardiac imaging in electrical
impedance tomography by means of a new electrode conguration. Physiol. Meas.
17(3) 179188
[16] Rabbani K S and Kabir A M 1991 Studies on the eect of the third dimension on a
two-dimensional electrical impedance tomography system. Clin. Phys. Physiol. Meas.
12(4) 393402
[17] Vonk Noordegraaf A et al 1997 Validity and reproducibility of electrical impedance
tomography for measurement of calf blood ow in healthy subjects. Med. Biol. Eng.
Comput. 35(2) 107112
[18] Vonk-Noordegraaf A et al 2000 Determination of stroke volume by means of electrical impedance tomography Physiol. Meas. 21(2) 285293
[19] Vonk Noordegraaf A et al 1996 Improvement of cardiac imaging in electrical
impedance tomography by means of a new electrode conguration Physiol. Meas.
17(3) 179188
[20] Vonk Noordegraaf A et al 1997 Noninvasive assessment of right ventricular diastolic
function by electrical impedance tomography Chest 111(5) 12221228
[21] Eyuboglu B M, Brown B H, Barber D C and Seagar A D 1987 Localisation of cardiac
related impedance changes in the thorax Clin. Phys. Physiol. Meas. 8 Suppl A 167173
[22] McArdle F J, Suggett A J, Brown B H and Barber D C 1988 An assessment
of dynamic images by applied potential tomography for monitoring pulmonary
perfusion Clin. Phys. Physiol. Meas. 9 Suppl A 8791
[23] Jongschaap H C, Wytch R, Hutchison J M and Kulkarni V 1994 Electrical impedance
tomography: a review of current literature Eur. J. Radiol. 18(3) 165174
[24] Eyuboglu B M and Brown B H 1988 Methods of cardiac gating applied potential
tomography Clin. Phys. Physiol. Meas. 9 Suppl A 4348
[25] Seagar A D, Barber D C and Brown B H 1987 Theoretical limits to sensitivity and
resolution in impedance imaging. Clin. Phys. Physiol. Meas. 8 Suppl A 1331
References
125
126
[46] Kunst P W et al 2000 Regional pressure volume curves by electrical impedance tomography in a model of acute lung injury Crit. Care Med. 28(1) 178183
[47] Gattinoni L, Pesenti A, Avalli L, Rossi F and Bombino M 1987 Pressure-volume
curve of total respiratory system in acute respiratory failure. Computed tomographic
scan study Am. Rev. Respir. Dis. 136(3) 730736
[48] Amato M B et al 1998 Eect of a protective-ventilation strategy on mortality in the
acute respiratory distress syndrome N. Engl. J. Med. 338(6) 347354
[49] Kunst P W, de Vries P M, Postmus P E and Bakker J 1999 Evaluation of electrical
impedance tomography in the measurement of PEEP-induced changes in lung
volume Chest 115(4) 11021106
[50] Frerichs I et al 2002 Detection of local lung air content by electrical impedance tomography compared with electron beam CT J. Appl. Physiol. 93(2) 660666.
[51] Hinz J et al 2003 End-expiratory lung impedance change enables bedside monitoring
of end-expiratory lung volume change Intensive Care Med. 29(1) 3743
[52] van Genderingen H R, van Vught A J and Jansen J R 2003 Estimation of regional
lung volume changes by electrical impedance pressures tomography during a
pressure-volume maneuver Intensive Care Med. 29(2) 233240
[53] van Genderingen H R, van Vught A J and Jansen J R 2004 Regional lung volume
during high-frequency oscillatory ventilation by electrical impedance tomography
Crit. Care Med. 32(3) 787794
[54] Victorino J A et al 2004 Imbalances in regional lung ventilation: a validation study on
electrical impedance tomography Am. J. Respir. Crit. Care Med. 169(7) 791800
Chapter 4
Electrical impedance tomography of
brain function
David Holder and Thomas Tidswell
4.1.
INTRODUCTION
128
understanding of how information is processed in neuroanatomical pathways. The problem is that such activity is widely distributed and occurs
with a timescale of the order of milliseconds. No system yet exists which
could measure such activity non-invasively and with a high temporal
resolution. One avenue of approach is to image changes in blood ow and
metabolic activity events which are related to nervous activity. These are
caused by the accumulation of the eects of many action potentials or
depolarizations. They are therefore easier to image, being large, but
change over several seconds and so can only give an indirect guide to nervous
activity. Such changes may already be imaged by positron emission tomography (Herholz & Heiss, 2004b) or functional MRI (Matthews & Jezzard,
2004). The temporal resolution of these techniques is seconds or tens of
seconds, because this is the timescale over which these changes in the brain
occur. Measurement of nervous activity with a much greater temporal
resolution of tens of milliseconds has been possible for decades with electroencephalography (EEG) (Michel et al, 2001a,b; Momjian et al, 2003) and,
more recently, by magnetoencephalography (MEG) (Wheless et al, 2004),
but these do not provide unique solutions and are of doubtful accuracy,
especially for deep or distributed sources.
If neuroimaging with EIT is successful, then it could be used in several
key clinical areas in which other methods of functional brain imaging are
unsuited. These include adults and infants receiving intensive care, and the
long term imaging of epilepsy on telemetry units, where prolonged periods
of monitoring are required in order to localize seizure activity in the preoperative assessment for epilepsy surgery. EIT may also be suited to provide
images of brain impedance changes brought about by cell swelling in cerebral
energy failure, in such pathological conditions as stroke, ischaemia, hypoxia
or hypoglycaemia. It also has the unique potential to provide a means of
imaging the tiny fast impedance changes due to opening of ion channels
during neuronal depolarization. This would provide a means of imaging
neuronal activity along neuroanatomical pathways with a temporal resolution of milliseconds, which would constitute a revolutionary development
in neuroscience technology.
The development of EIT for imaging brain function is relatively short.
An impedance scanning system for detecting brain tumours was designed
and tested (Benabid et al, 1978), but was not followed up with a practical
EIT device. Shortly after, Holder (Holder, 1987) independently proposed
EIT as a novel means for imaging the fast impedance changes known to
occur during neuronal activity in the brain. Pilot animal studies were then
performed in which simultaneous scalp and intracranial impedance measurements were made of the brain of anaesthetized rats during cerebral ischaemia
(Holder, 1992b). The conclusion was that measurements of brain impedance
could be made, non-invasively, by scalp electrodes, although these changes
were attenuated by the skull. This study indicated the practicality that EIT
129
could be used to image impedance changes in the human brain. At that time,
the only available EIT system was the Sheeld Mark 1 EIT system (Brown &
Seagar, 1987), which was limited in that current could only be applied
through adjacent electrodes. This system was unlikely to be able to image
impedance changes in the brain from scalp electrodes, as most of the applied
current would be shunted through the scalp. As the EIT technology was not
at the stage to inject current with more widely spaced electrodes, the Sheeld
Mark 1 was used, and experiments were designed to eliminate the eect of the
skull. In these, the eect of the skull was excluded by using a ring of electrodes placed on the exposed cortex of anaesthetized rats or rabbits. The
rst EIT study of brain activity was in articially induced stroke (Holder,
1992b), followed by EIT imaging during cortical spreading depression
(Boone et al, 1994), physiologically evoked responses (Holder et al, 1996b)
and during electrically induced seizures (Rao et al, 1997). The impedance
changes varied between a decrease of 2% and 5% during somatosensory
or visual stimulation, a 10% increase during seizures or up to 100%
during stroke, due mainly to cell swelling and blood volume changes.
Taking the evidence that functional activity changed brain impedance in
the rabbit by 25%, and that from rats the skull attenuated peak impedance
changes by a factor of 10, it seemed plausible that scalp impedance changes
of 0.20.5% might be detected non-invasively during functional activity in
humans. As this level of impedance change was within the sensitivity of an
EIT system, these initial studies paved the way for human functional imaging
studies. EIT of brain function has not yet broken through into routine
clinical use, but substantial progress has been made over the past decade
or so, largely in the authors group at University College London. We are
currently undertaking clinical trials in acute stroke and epilepsy.
In this chapter, we initially review the physiological basis for expecting
impedance changes during these conditions. We then review the development
and testing of hardware and reconstruction algorithms specically for
imaging brain function. Finally, we review animal and human studies in
the development of EIT for imaging brain function in the areas of EIT of
normal brain function, epilepsy and stroke.
4.2.
4.2.1.
The bioimpedance of tissues in the head is relevant in two main ways. EIT of
the brain poses an especially dicult, but not insuperable, problem, because
the brain is encased by a conductive covering, the cerebrospinal uid, two
layers with high resistivities, the pia mater and skull, and then the scalp,
130
131
Resistivity of cerebral white and grey matter in vivo. All measurements were
made at body temperature (3738 8C) in vivo.
Table 4.1.
Reference
R cortex
(
:cm) S.D.
R white matter
(
:cm) S.D.
Frequency
Method
Freygang &
229 9
Landau (1955)
344
1 kHz
4 electrodes
Nicholson
(1965)
85800
20 Hz20 kHz
Ranck (1963)
256356
5 Hz5 kHz
Point electrodes on
cortex
Latikka et al
(2001)
351
50 kHz
Monopolar needle
electrode
391
and white matter, which comprises tracts of long nerve bres which connect
dierent regions of the brain. Nerve bres in the mammalian brain are largely
surrounded by an insulating myelin sheath, and so are anisotropic. There was
anisotropy of about 10 :1 in the impedance of cerebral white matter in cats
over 20 Hz to 20 kHz (Nicholson, 1965)for example, 890
:cm for the longitudinal bres compared with 80
:cm for the transverse ones at 20 Hz. Grey
matter is largely isotropic as nerves and their processes run randomly.
However, Ranck (Ranck, Jr., 1963) noted that there is lamination in the
cortex, so this is only true at distances greater than 200 mm. In rabbit cerebral
cortex in vivo, at 5 Hz, the resistivity was 321 45
:cm (mean S.D.), falling
to 230 36:7
:cm at 0.5 kHz. When the shunting eect of the blood vessels
was taken into account, the resistivity values rose to 356
:cm for 5 Hz and
256
:cm at 0.5 kHz. Latikka (Latikka et al, 2001) recorded the impedance
of white and grey matter in situ using a needle electrode in human subjects
undergoing brain surgery for deep brain tumours. The average resistivity at
50 kHz for grey matter was 351
:cm and 391
:cm for white matter from
nine subjects (table 4.1). In summary, brain grey matter impedance at frequencies below 100 kHz is about 300
:cm in vivo, and white matter, depending on
orientation, is about 50% higher.
4.2.1.2.
132
been extended to include the similar events which occur in ischaemia, spreading depression or epilepsy. These events have been mostly studied in the
cerebral cortex, but also occur in other areas of grey matter in the brain.
When measured in the cerebral cortex, the characteristic event is that
spontaneous electrical activity ceases and a sustained negative shift of tens
of millivolts is recorded with an electrode on the cortical surface. These
events are accompanied by a substantial movement of ions and water, as
ionic homeostasis fails. Water follows sodium and chloride into cells, so
that the extracellular space shrinks by about 50% (Hansen & Olsen, 1980).
At frequencies up to 100 kHz, the great majority of current applied to the
brain passes through the extracellular uid. This component of current will
be resistive and so is measured by EIT systems, such as the Sheeld Mark
1 (Brown & Seagar, 1987), which measure the in-phase component of the
impedance. During anoxic depolarization, the impedance of grey matter in
the brain therefore increases, because the extracellular space shrinks.
Changes in temperature, the impedance of neuronal membranes and blood
volume may also contribute, but the eect due to cell swelling is greatly
predominant (gure 4.1). Changes of this type occur to diering degrees in
the pathological conditions of stroke (or cerebral ischaemia), spreading
depression and epilepsy. In each case, the cells run out of energy needed to
maintain the balance of water and solutes between the intracellular and
extracellular spaces. In stroke, this is because blockage of arteries leads to
an insuciency of blood; in spreading depression or epilepsy, it is because
intense neuronal activity exceeds the capacity of the blood to provide
energy supplies.
Large impedance increases of about 20100% occur during cerebral
ischaemia in species such as the rat (Holder, 1992a), cat (Hossmann, 1971)
and monkey (Gamache et al, 1975). Spreading depression is a phenomenon
which can be elicited in the grey matter of experimental animals by applying
potassium chloride solution or mechanical trauma. Intense activity of depolarized cells occurs, so that potassium and excitatory amino acids pass into the
extracellular space. These excite neighbouring cells by diusion. In this way a
concentric ripple of activity moves out from the site of initial disturbance
like a ripple in a pond. It moves at about 3 mm/min, and has been postulated
to be the cause of the migraine aura in humans (see Pearce, 1985). Impedance
increases of about 40% occur in various species (Bures, 1974). During epilepsy
induced in experimental animals, reversible cortical impedance increases of 5
20% have been observed during measurement at 1 kHz with a two-electrode
system in the rabbit or cat (Van-Harreveld & Schade, 1962). The changes
had a duration similar to the period of epileptic EEG activity and were due
to anoxic depolarization-like processes, as a negative d.c. shift occurred. Similar
changes have been observed in cat hippocampus, amygdala and cortex (Elazar
et al, 1966), and cat cortex (Shalit, 1965). Impedance increases of about 3%
have been recorded in humans during seizures (Holder et al, 1993).
133
Figure 4.1. Mechanisms of impedance change within the brain. Left gure: impedance decrease due to increased blood volume. During physiological
activity, a signal is sent to the blood vessels which increases blood ow and blood volume to that cortical area. As blood has a lower resistivity than the
surrounding brain (150 and 350
:cm, respectively), the increase in the lower resistivity volume of blood will allow more current to ow through that
area of tissue and decrease the bulk impedance of that volume of cortex. Right gure: impedance increase due to cell swelling. Cells expand during cell
swelling (bottom). At rest, the size of the conductive extra-cellular space (ECS) is about 20% of the brain volume. During epilepsy, moderate cell
swelling occurs as water and ions enter the glial cells and the neurones, and the volume of the low resistivity ECS is reduced. This increases the
bulk impedance of that volume of cortex. Larger changes of cell swelling and impedance occur during ischaemia and spreading depression.
134
4.2.1.3.
135
1939) when measured directly across the axon. There should therefore be
an impedance change across populations of cells in nervous tissue during
activity. The eect could be due to action potentials in white matter, or to
summated eects of synaptic activity in grey matter, which is the origin of
the EEG.
At the frequencies of measurement with EIT, most current passes in the
highly conductive extracellular space. The amplitude of the impedance
changes across tissue is therefore likely to be small. Klivington and Galambos (1968) measured impedance changes during physiologically evoked
activity in the auditory cortex of anaesthetized cats at 10 kHz. A maximum
decrease of about 0.005% was observed, which had a similar time course to
the evoked cortical response. Similar changes were measured in visual cortex
during visual evoked responses (Klivington & Galambos, 1967) and less
reproducible impedance decreases of up to 0.02% were observed in subcortical
nuclei during auditory or visual evoked responses in unanaesthetized cats
(Velluti et al, 1968). Freygang and Landau (1955) observed a maximum
decrease in impedance of 3.1%, measured with square wave pulses 0.3
0.7 ms long, during the evoked cortical response in the cat. There are therefore
discrepancies in the published data. Biophysical modelling and experimental
measurement, presented in section 4.7 below, suggests that changes are vanishingly small if recorded with a frequency of applied current above 1 kHz, so the
possibility exists that the above ndings were artefactual.
4.2.1.5.
There are two additional factors which may inuence the impedance of the
brain, but for which there is little experimental information. During
increased neuronal and, therefore, metabolic activity, an increased generation of heat may occur which would increase brain temperature. Decreased
brain temperature increases brain resistivity by approximately 23% per
8C (Ochs & Van Harreveld, 1956; Li et al, 1968). Cortical temperature
changes of up to 1 8C during functional activity, with an average 0.2 8C
decrease in temperature after 12 min of visual stimulation, have been
detected with MRI and fMRI, in humans (Yablonskiy et al, 2000). Such
cortical temperature changes could produce changes in impedance, which
could be detected by EIT, but these would occur over minutes, rather than
changes over seconds expected by blood volume change.
The thickness of the cerebro-spinal uid (CSF) which overlies the
activated cortex is another possible cause of apparent impedance change in
recording with scalp electrodes: an expansion of local cerebral blood
volume, such as during epileptic seizures, might shift small amounts of CSF
overlying adjacent supercial cortex to areas of lower volume (VollmerHaase et al, 1998). Changes of CSF pressure, monitored by indwelling
136
137
Tissue
Resistivity
(
:cm)
Reference
Frequency
(kHz)
White matter
Grey matter
Blood
CSF
Skull
344
229
125
69
6500
1
1
50
0.0110
0.110
This has been investigated by applying current to a skull inside a saline lled
tank (Rush & Driscoll, 1968). Closely spaced current injection electrodes
produced negligible current penetration within the skull, but when electrodes
were widely spaced across the skull (in polar positions), 45% of the applied
current entered the skull cavity. The current that does traverse the skull will
tend to shunt through the highly conductive cerebrospinal uid. The eect of
all this will be to decrease the signal-to-noise ratio, in the sense that the
signal will be sensitive to local changes in the scalp, and relatively insensitive
to events in the brain. One of the challenges in attempting brain EIT has been
to try and maximize the current owing into the brain itself.
4.3.
4.3.1.
The rst EIT recordings of brain function were made with the Sheeld Mark
1 system (Brown & Seagar, 1987). This employed 16 electrodes in a ring;
current was applied and voltage was recorded through adjacent pairs of
electrodes; the algorithm employed the assumption that the problem was
2D and that the imaged subject initially had a uniform resistivity. This was
used in specialized circumstances, where the experimental preparation was
designed to match the limitiations of the system. In anaesthetized rats or
rabbits, the entire upper surface of the skull and brain coverings (the dura
mater) were removed, and a ring of 16 spring-mounted electrodes were
placed on the exposed upper brain surface. As most of the activity occurred
in a layer of cerebral cortex about 3 mm thick, and the upper surface of the
brain in these species is almost planar, this was a good approximation to a 2D
uniform problem, and images were successfully obtained during stroke
(Holder, 1992b), epilepsy (Rao et al, 1997), spreading depression (Boone
et al, 1994) and evoked activity (Holder et al, 1996b).
138
139
Figure 4.2. (a) EIT system based on a Hewlett-Packard impedance analyser (opposite),
being used for human evoked response recording. (b) The UCLH Mark 1a employed in
chest imaging. (c) The UCLH Mark 1b. (d) The UCLH Mark 2.
140
The next system, termed the UCLH Mark 1a or 1b, was similar, but was
purpose built and based on a single impedance measuring circuit similar to the
Sheeld Mark 1 system. A constant current was applied to a pair of electrodes
and the impedance was calculated from the in-phase component of voltage
measurement from another pair. It diered from the Sheeld Mark 1 in
that it could record at much lower frequencies, electrodes could be addressed
exibly from software, and it was suitable for ambulatory recording. Recording could be performed at one of 18 frequencies from 77 to 225 kHz; up to 64
electrodes could be addressed (16 in the Mark 1a and 64 in the Mark 1b). It
comprised a headbox about the size of a paperback book into which the electrode leads were inserted, which could be worn in a waistcoat by the subject;
this connected to the base station by a lead 10 m long (gure 4.2(b)) (Yerworth
et al, 2002). It produced acceptable images down to 200 Hz in saline lled tanks
(Holder et al, 1999; Tidswell et al, 2003a) and has been successfully employed
for the rst ever EIT recordings in human subjects during epilepsy and
epileptic seizures (Bagshaw et al, 2003a; Fabrizi et al, 2004).
Although the Mark 1 systems were capable of applying currents of
dierent frequencies, they were not optimized for multi-frequency measurement and have only been used for time dierence imaging. The next generation device, termed the UCLH Mark 2, was designed with the aim of
imaging stroke, where time dierence imaging is not practicablea single
image needs to be acquired in a novel subject who already has brain pathology. We planned to do this by making dierence images across frequency.
The design is based on a single impedance measuring circuit of the Sheeld
multi-frequency Mark 3 system (Hampshire et al, 1995) for use with up to 64
electrodes through the use of cross-point switches (Yerworth et al, 2003). The
system injects currents from 2 kHz to 1.6 MHz. Some compromise is introduced by the use of the cross-point switches, so that the bandwidth for
good image quality is reduced to 800 kHz and the CMRR reduced by
10 dB to 80 dB. However, acceptable and reproducible images of multifrequency objects such as a banana in a saline lled tank could still be
obtained (gure 4.3). Our conclusion was that there were signicant practical
advantages in being able to address up to 64 electrodes in a software selectable way, and the reduction in signal quality appeared to be acceptable, at
least in tank studies (Yerworth et al, 2003). The system at present comprises
a power supply, a base box and a headnet and so is only suitable for sedentary recording. It is currently being used for a clinical trial of EIT frequency
dierence imaging in acute stroke. A smaller system with a headbox similar
to the Mark 1b, intended for ambulatory recording in epilepsy patients, is
being developed and we anticipate completion before the end of 2004.
Other groups have also been interested in EIT of the head. The earliest
attempts to image in the head were undertaken by a group at Oxford Brookes,
who constructed a system similar to the Sheeld Mark 1. It was intended for
imaging of intraventricular haemorrhage in the neonate, but no validated data
141
Figure 4.3. EIT images acquired with the UCLH Mark 2 EITS system. Banana, cucumber and Perspex were placed in 0.2% saline in a cylindrical tank with 16 electrodes in a
single ring. Time dierence imaging was performed at 640 kHz. The frequency dierence
image was collected at 640 kHz and referenced to 8 kHz (Yerworth et al, 2003).
series were produced (Murphy et al, 1987). A group in Amsterdam has recently
become interested in obtaining absolute conductivity estimates of the skull and
intracranial tissues for the purpose of setting model values for inverse source
modelling of the EEG (Goncalves et al, 2003). They employed a single
constant current source at 60 Hz and a conventional EEG machine with 64
electrodes to record voltages. The data were tted to a boundary element
model of the head which was optimized for a single parameter, the ratio of
mean skull resistivity to the brain. This varied from a ratio of 23 to 56,
mean 42 for six subjects. This represents the rst attempt to perform absolute
resistivity estimation in the head. Abboud and colleagues have been interested
in the possible use of EIT to record resistance changes during cryosurgery to
destroy brain tumours and have produced modelling studies which demonstrate the feasibility of the proposal (Radai et al, 1999; Zlochiver et al, 2002).
4.3.2.
When we rst attempted EIT of brain function, the only available hardware
was the Sheeld Mark 1 EIT system, which employed a 2D ltered
142
143
Figure 4.4. A plastic rod or sponge was immersed in 0.9% saline and placed at one of four
dierent positions. Data were collected with the UCLH Mark 1b system at 50 kHz with 16,
32 or 64 electrodes, and reconstructed with back-projection and constrained optimization.
The spatial resolution increased with increasing numbers of electrodes.
144
PET during similar activation, but it was not clear which of several factors
were responsible.
4.3.2.2.2.
The next step was to utilize an anatomically realistic model of the head,
obtained by segmenting MRI or x-ray CT images of the head. A method for
this computationally demanding task has been presented by Bayford et al
(2001), using integrated design engineering analysis software (IDEAS).
145
Figure 4.5. Finite element mesh used for reconstruction algorithm with realistic geometry.
The four layers (brain, CSF, skull and scalp) are shown.
Using this, our group at UCL produced a tSVD algorithm in which the
head was modelled as an FEM with four realistically shaped compartments
for brain, cerebrospinal uid, skull and scalp (gure 4.5). This produced
clear improvements in image quality in selected individual examples
drawn from tank studies, or recordings in humans during evoked activity or
epileptic seizures (Bagshaw et al, 2003a). However, it is possible that the
complexity introduced by additional computation and the ne meshes used
may outweigh the theoretical advantages of more accurate geometry.
Objective validation with respect to this issue is currently in progress in our
group; EIT images collected during evoked responses in adults and neonates
and during epileptic seizures will be evaluated using a tSVD algorithm and
ne FEM of the head, in comparison with an analytical multishelled model.
Realistic head models have also been implemented by Polydorides et al
(2002), who reconstructed images iteratively from simulation of a visual
evoked response using an FEM model with ve compartments and electrodes
arranged in a ring. In another study, the change in transfer impedance was
studied for a 3040% impedance change due to a 10 cm3 central oedema,
as simulated by an FEM model with realistic head geometry, including 13
dierent tissues and using hexahedral elements (Bonovas et al, 2001).
However, no images were presented using this technique.
4.3.2.2.4.
146
147
(a)
(b)
(c)
Figure 4.6. (a) Spherical tank containing a hollow plaster of Paris shell to simulate the
skull. Left: lower half of the tank and simulated skull. Right: the assembled tank with
no simulated skull. (b) and (c) Realistic phantoms, containing a human skull, for simulating the human head. (b) Latex tank with 0.2% saline simulating brain and scalp. Half
the tank is cut away to show the scalp inside. (c) Marrow tank in which the brain is
simulated by 0.2% saline, the scalp by alginate, and the skin by the skin of a marrow or
giant zucchini.
148
4.4.
There are good grounds for expecting that EIT could produce images of
increases in blood ow and volume, and related changes, which occur when
part of the brain is physiologically active. These changes have been the basis
of functional MRI and PET studies for over a decade, and have been reviewed
in section 4.2.1.3. If successful, EIT could provide a low-cost portable system,
which would produce similar images to fMRI and be widely used in cognitive
neuroscience in healthy and neurological or psychiatric subjects.
The local changes in the brain are small (a few per cent) and occur over
seconds or tens of seconds following the onset of activity. As the mechanism
of impedance dierence is probably changes in resistivity due to a changed
proportion of blood to brain, these may be imaged at any suitable frequency
which can distinguish these. In principle, a low frequency is desirable. This is
because the standing resistivity of brain becomes higher at low frequencies,
because applied current is restricted to the extracellular space (Ranck, Jr.,
1963), so the contrast between brain and the conductive blood will be greater.
On the other hand, instrumentation errors due to skin impedance may be
expected to be greater, as skin impedance is higher at low frequencies
(Rosell et al, 1988). An applied frequency of 50 kHz, as used in the Sheeld
Mark 1 system, appeared to be a good compromise.
4.4.1.
The rst EIT images during evoked physiological activity were collected with a
Sheeld Mark 1 system, using a ring of 16 spring mounted electrodes placed
149
Figure 4.7. EIT images of rabbit cortex during visual stimulation. Images displayed were
collected every 30 s. An impedance decrease may be seen over the posterior visual cortex
which persists for about 30 s after cessation of stimulation.
Human studies
150
Figure 4.8. Examples of impedance changes in the raw impedance data. Impedance
changes from single channel four-electrode impedance recordings, during motor (top
row, eight repetitions) or visual stimulation (bottom row, n 12). On the left, data
from a single electrode combination are shown; all repetitions are superimposed. Reproducible impedance changes are seen at selected electrode combinations with the same time
course as the stimulation paradigms. The y axis indicates the percentage change from baseline impedance. Impedance measurements were made every 25 s; the lines between these
measurements are drawn for clarity. Both impedance increases and decreases were
observed. On the right are shown all 258 electrode combinations for the same subjects,
displayed as a sorted waterfall graph. The 812 runs for each electrode combination
were averaged together. The averages were sorted according to the size of the impedance
change during stimulation and stacked on the vertical axis. Measurements with baseline
noise greater than the impedance changes are excluded from these plots so that these
changes are not obscured. Signicant stimulus-related impedance increases and decreases
are seen in approximately 25% of electrode measurements in these subjects.
151
Unfortunately, the reconstructed images from this data were noisy, and
the impedance changes were not consistently localized to the appropriate
areas of cortex. The reconstruction algorithm used a simple analytical
model of the human head in the forward solution, based on a homogeneous
conductivity sphere (see section 4.3.2.2) (Gibson, 2000). It was likely that the
use of this simple model of the human head led to image errors when used on
real human data. The source of such reconstruction errors could have been
due to dierences in shape, absence of the four layers of scalp, the skull,
CSF and the brain, or there may have been errors in electrode position
between the human head and the reconstruction model.
As the actual impedance changes that occur in the human brain during
stimulation are unknown, the 3D reconstruction algorithm, based on the
homogeneous spherical model of the head, was tested in tanks of increasing
degrees of dierence from the head model employedthe spherical tank
(section 4.3.3, gure 4.6(a)), or the latex tank with a realistic head shape
(gure 4.6(b)), with or without the skull (Tidswell et al, 2001c). EIT images
of a sponge, 14 cm3 volume, with a resistivity contrast of 12%, were acquired
in three dierent positions in tanks lled with 0.2% saline. In the hemispherical tank, 19 cm in diameter, the sponge was localized to within 3.410.7% of
the tank diameter. In a head-shaped tank, the errors were between 3.1 and
13.3% without a skull and between 10.3 and 18.7% when a real human
skull was present. This demonstrated that a signicant increase in localization error occurred if an algorithm based on a homogeneous sphere was
used on data acquired from a head-shaped tank. In addition, the localization
error was mainly due to the presence of the skull, as no signicant increase in
error occurred if a head-shaped tank was used without the skull present,
compared with the localization error within the hemispherical tank. The
error due to the skull signicantly shifted the impedance change within the
skull towards the centre of the image by up to 8% of the image diameter.
As soon as an improved reconstruction algorithm became available, in
which the head was modelled with four realistic compartments (section
4.3.2.2, gure 4.5), the data was re-analysed. The images produced using
this reconstruction algorithm showed a clear improvement. Correctly localized impedance changes with the same time course as the stimulus were
found in 38/51 images19 when reconstructed with the algorithm which
employed a homogeneous sphere head model (gure 4.9). Unfortunately,
despite these improvements the EIT images were still noisy and contained
multi-focal impedance changes, even after statistical thresholding.
In summary, the evoked response studies have been encouraging, but
are not yet at a stage where EIT systems could be condently used as a
robust tool for human psychophysical or clinical studies. The reason for
the bottleneck in image quality is not entirely clear. The size of changes
about 1% in human studies with scalp electrodesis close to the noise
from electronic and physiological sources, but the reliable raw impedance
152
Electrical impedance tomography of brain function
Figure 4.9. Impedance changes in four subjects during right motor stimulation (repetitive movements of the ngers of the right hand). These all
show an impedance decrease in the area of the contralateral motor cortex on the left, and are more in keeping with the hypothesis that blood
volume is increased in the area of cortex expected to be stimulated by the motor stimulus.
153
154
4.5.
EIT OF EPILEPSY
Because EIT systems can produce several images a second, and are portable
and safe, they are ideally suited to image blood ow and related changes
that occur during epileptic activity with a high time resolution. EIT could be
employed to localize the part of the brain that produces epileptic seizures, so
that resective surgery can be performed. At present, about 80% of patients
with epilepsy can be satisfactorily treated with medication. Of the remainder,
some can be cured or improved by surgery. In order to perform this, it is
essential that the correct source of epilepsy in the brain is localized. This is
usually performed with a prior stay in hospital of several days. EEG and
video are monitored continuously, so that they are recorded when a number
of seizures occur. The EEG is usually performed with scalp electrodes but, if
the seizure onset zone is unclear, it may be performed with subdural mat or
depth electrodes, inserted at operation. Together with psychometry and
neuroimaging studies, the onset zone is usually localized, and a decision as
to whether to embark on surgery is undertaken.
EIT could be run concurrently with scalp EEG during this pre-surgical
EEG telemetry. EIT images would be recorded about once a second over a
period of days while the patient was observed on the ward. When a seizure
occurred, the EIT images would be retrospectively analysed to see if changes
in impedance occur at the same time as EEG activity. Imaging of this nature,
with a temporal resolution or seconds, is not presently possible by any other
method. In principle, the same information could be obtained if a subject had
a seizure when in an fMRI scanner, but this is not practicable. Recent
advances in neuroimaging have lessened the need for invasive recordings
with depth or subdural mat electrodes, but these still need to be performed
in patients in whom pre-surgical ndings are not congruent. While subdural
electrodes carry a low risk, depth electrodes which penetrate into the cerebral
substance carry a signicant morbidity and mortality. Haemorrhage resulting in permanent neurological damage occurred in 0.8% in one report (879
patients); in another, two patients of a series of 140 died (see Van Buren,
1987). Ictal EIT could be performed safely and non-invasively with EEGtype scalp electrodes, and may become a routine additional method to
EEG during telemetry. If successful, it would reduce further the need for
invasive depth EEG recordings and so have a direct benet for patient
EIT of epilepsy
155
The rst EIT studies in epilepsy were, as for evoked responses, collected with
a Sheeld Mark 1 system using a ring of 16 spring mounted electrodes placed
on the exposed supercial surface of the brain of anaesthetized rabbits.
Epileptic seizures were induced by focal electrical stimulation and were
either localized or spread to involve the entire brain (Rao, 2000; Rao et al,
1997) (gure 4.10). Reproducible predominant impedance increases of
7:1 0:8% (localized) and 5:5 0:8% (generalized) were present in EIT
images in nine animals at the sites where the epilepsy was initiated. As in
the previous animal study in evoked responses, there were smaller adjacent
impedance decreases apparent in the images. In this study (which followed
that of evoked potentials), two probes were placed on the brain near the
site of seizure onset and about 10 mm away, to try to elucidate the physiological mechanisms and establish if the impedance increases and decreases were
physiological or due to reconstruction algorithm artefact. Local impedance
measured at both sites was always an increase. Extracellular potassium,
temperature, d.c. potential and laserDoppler owmetry were all consistent
with the expected mechanism of cell swelling as the explanation for the
increased impedance. The probable increase was about 10%, but was
oset slightly by a concurrent decrease of a few per cent due to increased
temperature and blood volume. The decreases appeared to be due to noise
or to a reconstruction artefact.
In relation to this, some single channel studies were performed in
humans. The previous literature (section 4.2.1.2) demonstrated impedance
increases in animals. The proportion of glial cells in humans is greater, so
the theoretical possibility existed that impedance changes might not occur
156
Figure 4.10. EIT images during a partial epileptic seizure. A ring of 16 electrodes was
placed on the exposed brain of an anaesthetized rabbit. EIT images were collected every
5 s, while a seizure was elicited by electrical stimulation at the site of the small arrow
(near electrode 10). The electrocorticogram was recorded from the same electrodes, and
selected ECoG and EIT images every 30 s are shown. An impedance decrease may be
seen to develop and fade away in concert with the ECoG changes, at the site corresponding
to the electrical onset.
EIT in stroke
157
Figure 4.11. Example of EIT images collected with the UCLH Mark 1b during two
epileptic seizures in a subject undergoing EEG telemetry as assessment prior to surgery
for intractable epilepsy. The EIT headbox is visible in his left breast pocket. Independent
investigations, including MRI and EEG, indicated seizures originated from the left
temporal lobe; blood ow changes occurred in the appropriate region in both seizures
imaged. Only the images at the onset of the seizures are shown, but images recorded
three times a second reveal blood ow changes evolving over tens of seconds. Similar
changes have been observed in four other subjects.
4.6.
EIT IN STROKE
158
drugs is eective for ischaemic stroke due to occlusion of arteries, but needs to
be undertaken within 3 h of the onset of symptoms. A brain scan is required
prior to treatment onset to dierentiate between ischaemic and hemorrhagic
strokes; thrombolytic drugs cannot be used where there is a haemorrhage as
they may extend it. In practice, it is dicult to obtain rapid scans because of
the diculty of obtaining access to a scanner and rapid reporting. There is
therefore a need for a neuroimaging system which could be utilized in casualty
departments or health centres, which is inexpensive, rapid and safe. EIT could
be ideal for this purpose. It could be used with an array of elasticated scalp
electrodes, which may be easily applied by a technician or nurse in a few
minutes. Interpretation could be performed by a trained technician or by a
radiologist using remote reporting over a network or the internet. It could
also be useful for research studies in which new treatments for stroke
needed to be assessed over days as a stroke evolved.
However, unlike the applications above, time dierence imaging could
not be performed as the clinical need is for a single image on presentation.
This could be achieved by absolute imaging, but this has not yet been
shown to be practicable for clinical studies. The possibility exists, however,
for achieving this by multi-frequency imaging in which dierence images
are produced by referencing one frequency against another (Brown et al,
1995). The main principle will be that the impedance spectrum of blood in
the range 1 kHz1 MHz will be dierent from brain and recently ischaemic
brain.
Holder (1992a) performed pilot single channel impedance measurements in a reversible model of cerebral ischaemia in the anaesthetized rat.
With a single applied frequency of 50 kHz, increases of 1560% were
recorded. Scalp recording from the same electrode combinations revealed
changes decreased to 1020%. This suggested that the changes were large
enough to be recorded through the skull, at least in this animal model. The
rst EIT images taken with scalp electrodes were then recorded in the
same animal model. Clear reversible changes of 10% were apparent on
images (Holder, 1992b). However, these were collected with the Sheeld
Mark 1 system and 2D back-projection algorithm. The accuracy of the
images was not clear, as no independent standard was available for
comparison. There were some unexpectedly large posterior changes, so it is
probable that errors occurred, but this work at least qualitatively supported
the principle that this could be possible.
We are not aware of other further physiological studies, but a group has
published a proposal for a reconstruction algorithm for imaging stroke (Clay
& Ferree, 2002). We have developed the UCLH Mark 2 system specically
with this application in mind (Yerworth et al, 2003, 2004). It is capable of
imaging vegetable samples with similar properties to brain and blood in
cylindrical tanks, but a nonlinear algorithm must be used as the large changes
in impedance contrast throughout the tissue, a necessary consequence of
159
multi-frequency referencing, violating the assumptions used in linear algorithms. A clinical trial of this approach in subjects with acute stroke is
currently under way in our group.
4.7.
The novel applications presented above all make use of the low cost and portability of EIT, but similar images can already be obtained with fMRI or PET.
However, EIT could in principle be used to image neuronal activity over
milliseconds (section 4.2.1.4). The proposed application would be to record
EIT images from one or more rings of electrodes, either around the brain
in experimental animals or human surgical subjects, or, ultimately, around
the scalp. Data would be gathered after a repeated stimulus, in the same
way as somatosensory or visually evoked responses. An EIT image would
subsequently be reconstructed for each millisecond or so of the recording
window. In this way, it would be possible to determine the waveform of activity in any selected pathway during evoked responses. This is not currently
possible by any existing method, and, if possible, this would be a substantial
advance. Unfortunately, it poses a formidable technical challenge. The
reconstruction algorithms developed for EIT of the brain (section 4.3.2)
could be employed as they stand, and the small changes would probably
be suitable for linear reconstruction approaches.
The physiological basis is clear, but an important issue is the magnitude
of the likely changes. This has been modelled using cable theory (Boone &
Holder, 1995; Boone, 1995; Liston, 2004). The model was initially for the
ideal case of unmyelinated peripheral nerve. The rst observation was that
the frequency of recording was critical: above about 100 Hz, the resistance
changes during activity fell o steeply. For a four-electrode measurement
for a mean bre diameter of 1 mm, the calculated impedance change was
3.7% at d.c. but 0.009% at 30 kHz (Boone, 1995). Further work and renements, such as the inclusion of incomplete depolarization of the nerve and
the eects of the capacitance of the membrane, were made to the model
(Liston, 2004). At d.c., the new model predicted a resistance decrease of
2.8%. This has been experimentally validated with recordings in crab
nerves (Barbour, 1998), where resistance changes at d.c. of 1:1 0:1%
were recorded.
The modelling has been extended to estimating the resistance changes in
cerebral cortex (Boone, 1995; Liston, 2004). The size of the change depends
critically on the proportion of neurones that depolarize in an active part of
the brain, which is unknown. Assuming this was 10% of available neurons,
the model estimated the resistance change to be 0.6% locally within brain
tissue. For a physiologically reasonable volume of cortex near to the surface,
the resulting peak scalp resistance changes were 0.06%. Ahadzi has modelled
160
this using a realistic nite element model of the head in order to determine
whether more sensitive measurement could be obtained by the use of
magnetoencephalography to detect magnetic elds (Ahadzi et al, 2004). His
conclusion was that peak changes were about 0.03%, and that the signal-tonoise ratio was very similar to those predicted for electrical measurement.
This prediction has not, to the authors knowledge, been fully tested.
Boone (1995) recorded changes of 0.010.03% in preliminary measurements
at low frequency in the cortex of anaesthetized rabbits during physiologically
evoked responses. Published data reviewed in section 4.2.1.4 claimed changes
of about this order in cat brain (Klivington & Galambos, 1968), but these
were at 10 kHz, at which the model predicts vanishingly small changes, so
the validity of these ndings is unclear. Holder (1989) was unable to detect
any reproducible changes larger than 0.002% at 50 kHz during evoked
responses in human subjects.
The application of EIT to imaging these changes is intriguing, but these
estimates of its magnitude place the changes at the extreme limits of detectability. Sensitive impedance recording circuits can detect changes of the order
of 0.01% at low frequencies with prolonged averaging, but this is for peak
changes for relatively large volumes of cortex near the surface. For imaging
to be useful, deeper changes need to be imaged, and recording times for
multiple electrode combinations need to be practicable. At present, it is
not clear if these diculties could be overcome in practice to yield acceptable
EIT images in the half hour or so a subject could be expected to tolerate
recording.
4.8.
At rst sight, EIT of brain function might have been supposed to be too dicult, in view of the resistance barrier of the skull. The substantial preliminary
work presented in this chapter, in tanks and animals, suggests that this is not
the case, and that satisfactory images can indeed be obtained with the use of
specialized reconstruction algorithms and recording equipment. If EIT can
be shown to produce acceptable images, then there is little doubt that the
portability and low cost of EIT could enable it to provide an essential
additional imaging technique when the applied frequency is set up to
image blood ow, cell swelling and related changes. Applications in epilepsy
and stroke are currently the leading areas in this, but there are several others,
such as in monitoring head injury or cryosurgery (Radai et al, 1999). If
imaging of neuronal depolarization were possible, this would be a uniquely
important advance.
However, the critical issue is whether the inherent limitations of EIT
low spatial resolution and sensitivity to noisy measurementcan be
suciently overcome to yield clinically robust data. Preliminary ndings in
References
161
162
Boone K G 1995 The possible use of applied potential tomography for imaging action
potentials in the brain. PhD thesis, University College London
Boone K G and Holder D S 1995 Design considerations and performance of a prototype
system for imaging neuronal depolarization in the brain using direct current
electrical resistance tomography Physiol. Meas. 16 A87A98
Brown, B and Seagar, A 1987 The Sheeld data collection system. Clin. Phys. Physiol.
Meas. 8 9197
Brown B H, Leathard A D, Lu L, Wang W and Hampshire A 1995 Measured and expected
Cole parameters from electrical impedance tomographic spectroscopy images of the
human thorax. Physiol. Meas. 16 A57A67
Bures, J B 1974 The Mechanism and Applications of Leaos Spreading Depression of
Electroencephalographic Activity (New York: Academic Press)
Clay M T and Ferree T C 2002 Weighted regularization in electrical impedance tomography with applications to acute cerebral stroke IEEE Trans. Med. Imaging 21
629637
Cole K S and Curtis H J 1939 Electric impedance of the squid giant axon during activity
J. Gen. Physiol. 649670
Coulter N A and Pappenheimer J R 1948 Development of turbulence in owing blood Am.
J. Physiol. 159 401408
Derdeyn C P, Videen T O, Yundt K D, Fritsch S M, Carpenter D A, Grubb R L and
Powers W J 2002 Variability of cerebral blood volume and oxygen extraction:
stages of cerebral haemodynamic impairment revisited Brain 125 595607
Dietzel I, Heinemann U, Hofmeier G and Lux H D 1982 Stimulus-induced changes in
extracellular Na and Cl concentration in relation to changes in the size of the
extracellular space Exp. Brain Res. 46 7384
Elazar Z, Kado R T and Adey W R 1966 Impedance changes during epileptic seizures
Epilepsia 7 291307
Fabrizi L, Sparkes M, Holder D S, Yerworth R, Binnie C D and Bayford R 2004 Electrical
impedance tomography (EIT) during epileptic seizures: preliminary clinical studies,
in XII International Conference on Bioimpedance and Electrical Impedance Tomography, Gdansk, Poland
Ferree T C, Eriksen K J and Tucker D M 2000 Regional head tissue conductivity estimation for improved EEG analysis IEEE Trans. Biomed. Eng. 47 15841592
Freygang W H and Landau W M 1955 Some relations between resistivity and electrical
activity in the cerebral cortex of the cat J. Cellular Comparative Physiol. 45 377392
Gabor A J, Brooks A G, Scobey R P and Parsons G H 1984 Intracranial pressure during
epileptic seizures Electroencephalogr. Clin. Neurophysiol. 57 497506
Gamache F W Jr., Dold G M and Myers R E 1975 Changes in cortical impedance and
EEG activity induced by profound hypotension Am. J. Physiol. 228 19141920
Geddes L A and Baker L E 1967 The specic resistance of biological material: A compendium of data for the biomedical engineer and physiologist Med. Biol. Eng. 5 271293
Gibson A 2000 Electrical impedance tomography of human brain function. PhD thesis,
University College London
Gibson A, Bayford R and Holder D 2000 Two-dimensional nite element modelling of the
neonatal head Physiol. Meas. 21 4552
Goncalves S, de Munck J C, Heethaar R M, Lopes da Silva F H and van Dijk B W 2000
The application of electrical impedance tomography to reduce systematic errors in
the EEG inverse problema simulation study Physiol. Meas. 21 379393
References
163
164
Klivington K A and Galambos R 1968 Rapid resistance shifts in cat cortex during clickevoked responses J. Neurophysiol. 31 565573
Latikka J, Kuurne T and Eskola H 2001 Conductivity of living intracranial tissues Phys.
Med. Biol. 46 16111616
Law S K 1993 Thickness and resistivity variations over the upper surface of the human
skull Brain Topography 6 99109
Lemieux L, Salek-Haddadi A, Josephs O, Allen P, Tom, N, Scott C, Krakow K, Turner R
and Fish D 2001 Event-related fMRI with simultaneous and continous EEG:
description of the method and initial case report Neuroimage 14 780787
Li C L, Bak A F and Parker L O 1968 Specic resistivity of the cerebral cortex and white
matter Exp. Neurol. 20 544557
Liston A D, Bayford R H, Tidswell A T and Holder D S 2002 A multi-shell algorithm to
reconstruct EIT images of brain function Physiol. Meas. 23 105119
Liston A D 2004 Models and image reconstruction in electrical impedance tomography of
human brain function. PhD thesis, Middlesex University
Liston A D, Bayford R H and Holder D S 2004 The eect of layers in imaging brain
function using electrical impedance tomography Physiol. Meas. 25 143158
Lux H D, Heinemann U and Dietzel I 1986 Ionic changes and alterations in the size of the
extracellular space during epileptic activity Adv. Neurol. 44 619639
Malonek D, Dirnagl U, Lindauer U, Yamada K, Kanno I and Grinvald A 1997 Vascular
imprints of neuronal activity: relationships between the dynamics of cortical blood
ow, oxygenation, and volume changes following sensory stimulation Proc. Natl.
Acad. Sci. USA 94 1482614831
Matthews P M and Jezzard P 2004 Functional magnetic resonance imaging J. Neurol.
Neurosurg. Psychiatry 75 612
Michel C M, Thut G, Morand S, Khateb A, Pegna A J, Grave d P, Gonzalez S, Seeck M
and Landis T 2001 Electric source imaging of human brain functions Brain Res.
Brain Res. Rev. 36 108118
Minns R A and Brown J K 1978 Intracranial pressure changes associated with childhood
seizures Dev. Med. Child Neurol. 20 561569
Momjian S, Seghier M, Seeck M and Michel C M 2003 Mapping of the neuronal networks
of human cortical brain functions Adv. Tech. Stand. Neurosurg. 28 91142
Morucci J P, Granie M, Lei M, Chabert M and Marsili P M 1995 3D reconstruction in
electrical impedance imaging using a direct sensitivity matrix approach Physiol.
Meas. 16 A123A128
Murphy D, Burton P, Coombs R, Tarassenko L and Rolfe P 1987 Impedance imaging in
the newborn Clin. Phys. Physiol. Meas. 8 Suppl A 131140
Nicholson P W 1965 Specic impedance of cerebral white matter Exp. Neurol. 13 386401
Ochs S and Van Harreveld A 1956 Cerebral impedance changes after circulatory arrest
Am. J. Physiol. 187 180192
Oostendorp T, Delbeke J and Stegeman D 2000 The conductivity of the human skull: results
of in vivo and in vitro measurements IEEE Trans. Biomed. Eng. 47 14871492
Palmer J T, de Crespigny A J, Williams S, Busch E and van Bruggen N 1999 Highresolution mapping of discrete representational areas in rat somatosensory cortex
using blood volume-dependent functional MRI Neuroimage 9 383392
Pearce J M 1985 Is migraine explained by Leaos spreading depression? Lancet 2 763766
Pfutzner H 1984 Dielectric analysis of blood by means of a raster-electrode technique Med.
Biol. Eng. Comput. 22 142146
References
165
166
Chapter 5
Breast cancer screening with electrical
impedance tomography
Alex Hartov, Nirmal Soni and Ryan Halter
5.1.
5.1.1.
Approximately one woman in eight will develop breast cancer over a lifetime
in the US [1]. The prognostic for women diagnosed with the disease is greatly
inuenced by the stage at which it is discovered. Long term survival is
signicantly improved for women found with small tumours in the early
stages of development. Periodic mammograms for women over 40 or 50
years of age constitute the principal tool used in screening for breast
cancer and can be credited with saving many lives. However, mammography
has not reached the level of perfection desirable for a mass screening tool.
Exposure to x-rays, although minimal in mammograms, is one objection
that is raised, particularly for women who are advised to have more frequent
exams and to start at an earlier age, due to a family proclivity. It is thought
that the cumulative x-ray exposure, beyond a reasonable lifetime quota, may
itself become a health risk.
More immediately of concern for women who undergo the examination is
the signicant discomfort caused by the need to squeeze the breasts to a thickness of a few centimetres against a detector plate. The procedure is thought to
discourage some women from submitting to regular examinations.
From a public health point of view, the greatest objections to x-ray
mammography is its imprecision as a diagnostic tool. Studies estimate that
a woman with a tumour may remain undiagnosed following a mammogram
(false negative) 1025% of the time [24]. This means a sensitivity of up to
90%. Conversely, women who undergo periodic examinations will have a
high probability of an abnormal nding; nearly a 50% chance after 10
visits according to one study [5]. Such ndings typically call for biopsies to
168
169
170
171
Many more studies can be found that have published data on ex vivo
breast impedance. Most of the results reviewed here seem to concur that
cancer tumours have lower impedance than normal surrounding tissues.
Many fewer studies have published data based on in vivo invasive
measurements. One of the few groups to publish such data, Morimoto et al
[21], used a specially designed probe inserted in breast tumours on anaesthetized patients, and measured impedances between the needle tips and an
abdominal patch electrode, using a three-lead technique. Measurement
data from these studies was presented in the form of equivalent lumped
components Re, Ri and Cm, forming a network in which Re is in parallel
with a series combination of Ri and Cm. This way of presenting the data
makes it dicult to compare with other studies. In this study Re and Ri
were found to be higher in tumours, while Cm decreased in tumours,
compared with normal tissues. Although this study showed that signicant
dierences in the electrical responses of the dierent types of tissue could
be used for dierentiation, it is largely in disagreement with other data
regarding the direction of the changes, presenting an increase in impedance
instead of a drop for cancerous tumours.
A few groups have performed non-invasive two points impedance
measurements on breasts with and without tumours [22, 23]. The reports
based on these experiments indicate again a drop in resistance and an
increase in capacitance [22] for cancerous tumours, or at least that dierentiation is possible [23].
5.2.
Impedance mapping
172
Tomographic imaging
173
At this time it does not appear likely that impedance imaging will unseat
mammography as the primary method of screening for breast cancer. Its
poor spatial resolution, compared with x-ray, represents a barrier to its
being adopted, even if its sensitivity and specicity were to improve.
However, given the current performance of x-ray mammography, it is
quite conceivable that impedance will be adopted as a second step in the
standard examination, when an abnormality is discovered. EIT systems
could be designed to be relatively inexpensive to purchase (<$10 000) and
very inexpensive to use. Examinations could be very rapid (<10 min), and
very safe. They do not involve x-ray exposure and could be repeated as
often as needed.
5.3.
Few groups to date have presented clinical results of breast cancer screening.
Most of the results published were based on planar array instruments such as
the T-Scan (marketed by Siemens as the TS2000), which has received FDA
approval for use as an adjunct to mammography [29] and has been used
by several groups worldwide in clinical trials.
The only clinical experiments we are aware of, using the tomographic
approach based on circular arrays, is under way at Dartmouth [30]. The
clinical trial which is to conclude their ongoing project has not been
concluded yet and so will not be presented here. However, a few preliminary
174
studies and ndings have been published by that group which will be briey
discussed here.
5.3.1.
5.3.1.1.
Planar arrays
Piperno 1990 [31]
This is a large-scale study based on 6000 patients using the Mamoscan device,
an early version of the planar array system now marketed under the TS2000
name. Although this is not the rst such study, it is the largest and most
signicant evaluation of impedance imaging. Of this patient group, 745 have
undergone biopsies to verify a suspicious nding in mammography. Every
patient in the group submitted to mammography, palpation, transillumination, thermography and ultrasound exams. This study set out to compare all
these modalities against each other. The rst nding was that there were
nine cases in which impedance imaging was the only modality to ag as
highly suspicious exams that all other modalities did not detect and which
were conrmed by histopathology. The presentation of the results in that
publication makes it impossible to compute the usual statistics regarding the
rates of true positive (etc.) or the sensitivity and specicity of each modality.
In their tabulated results, it is shown that the Mammoscan was correct (i.e.
true positive true negative) in 454/745 cases and incorrect (i.e. false
positive false negative) in 119/745 cases. The remaining cases were labelled
as borderline cases with no further indication of outcomes. For x-ray
mammography the results are 395/745 correct ndings, and 154/745 incorrect
ndings. The number of correct ndings is greatest for impedance imaging,
compared with all the other exam types, and the number of incorrect ndings
is the lowest for impedance imaging as well. This early study was interpreted as
very encouraging for impedance imaging at the time of its publication.
5.3.1.2.
175
imperfections (lesions, scratches, moles etc.) and air bubbles resulting from
placement constitute a reported practical limitation to the eectiveness of the
TS2000.
5.3.1.3.
The system used here is very similar to the TS2000, consisting of a planar array
of 256 electrodes 12 cm on a side. The image reconstruction is slightly dierent
in that impedances are computed at dierent planes away from the electrode
array in order to reconstruct a 3D map of the volume facing the array. Slices
of the volume between the electrode array and the chest wall were computed
every 8 mm and for up to 6 cm depth. Twenty-one patients with tumours in
sizes ranging from 1.5 to 5 cm in one breast were examined. Both breasts
were imaged, with the contralateral breast used as a normal reference. Imaging
was performed twice, with the patients standing and reclining, resulting in 84
data sets. The data sets were divided into ve groups, including (1) 42
normal breasts, and (2) 42 malignant tumours. Group (2) was subdivided,
based on whether the tumours were visible as white spots, into groups (3)
16 studies without focal abnormalitiesand (4)26 studies with visible
abnormalities. Group (5) contained 13 studies, selected from group (4) for
their high conductivity peaks. Tumours were correctly identied in 14 out of
a total of 21 cases (67%), as evidenced by clearly visible white spots on the
reconstructed images. Four more were identied as anomalies due to the
inhomogeneous aspect of the images, which brings the TP rate to 85.7%.
The analysis was repeated using more sophisticated statistical methods instead
of visual inspection. With this approach, groups (3) and (4) (malignant
tumoursboth types) could be identied in 19 of 21 cases (90.5%), on the
basis of signicant statistical dierences in the property distributions.
Although this study shows that malignant tumours can be identied when
compared with normal breast examinations, it does not tackle the more
important question of whether impedance imaging can be used to discriminate
between malignant and other types of abnormalities, which is where
mammography comes short.
5.3.1.4.
176
In this study, the authors set out to determine whether impedance mapping
duplicates or augments the clinical results obtained with ultrasound and
MRI, as an adjuvant to mammography. One hundred patients were examined with ultrasound, impedance imaging and MRI, following ambiguous
abnormal ndings in their mammograms. In all, 100 abnormalities were
studied. Ultrasound and MRI ndings were categorized by experienced
radiologists using the LOS (level of suspicion) scale, with LOS values corresponding to: 1 normal, 2 most likely benign lesion, 3 probably benign, 4
probably malignant, 5 most likely malignant. Ultrasound ndings with
LOS of 2 or 3 were categorized as non-malignant ndings, while LOS 4
and 5 were categorized as malignant ndings. Impedance imaging images
were categorized as indicative of a malignant nding if a bright spot was
visible, and could not be discarded as artefact due to poor contact or the
presence of the nipple. Sixty-four such lesions were identied on impedance
maps and were categorized as positive ndings. Impedance imaging showed a
sensitivity of 81% and a specicity of 63%, ultrasound had a sensitivity of
77% and a specicity of 89%, and MRI had a sensitivity of 98% and a
specicity of 81% on this group of patients. These ndings correspond to
the individual modalities individual performances. Statistical analysis
further showed that impedance imaging adds clinical information to ultrasound, but that MRI and impedance imaging are mostly similar in the
information they contribute to the diagnosis.
5.3.1.6.
This study uses the same hardware as was presented above for that group, with
one change in the hardware: the planar array consisting of 256 electrodes is
now arranged in a circular pattern which increases the utilization factor for
the electrodes. In the previous square arrangement, electrodes in the corners
tended not to make contact with the patients. Furthermore, the array is
somewhat smaller, increasing the electrode density, and as a result the spatial
resolution is achievable with the device. This study did not seek to evaluate the
performance of impedance mapping in breast cancer screening; rather it sets
out to establish baseline measurements for women in several categories.
Fifty-seven women were examined in ages ranging from 18 to 61. The patients
were selected to t in ve groups: (1) 12 women (18 to 45 years) in the rst
menstrual phase (1 to 10 days); (2) 12 women (18 to 45 years) in the second
menstrual phase (16 to 28 days); (3) 14 women (18 to 39 years) during their
177
In this study, using data collected with the TS2000 impedance imaging system,
the authors implemented an automatic algorithm to identify bright spots which
correspond to conductivity increases and generally to malignant tumours.
They also rened the algorithm to discriminate more reliably between malignant and benign lesions. Their algorithm is based on two main predictors,
the phase at 5 kHz and the crossover frequency (where the imaginary part of
admittance peaks). A learning process was used to adjust the identication
thresholds which were trained on data from 461 examinations, with 83 malignant and 378 benign cases. The designation of every case was based on biopsy
results. With this methodology, they applied their algorithm to a separate
group of 240 examinations (87 malignant, 153 benign). Under these conditions
they reported a sensitivity of 84% and a specicity of 52% in properly identifying malignant and benign impedance images.
5.3.1.8.
178
lobular carcinoma also had positive impedance imaging diagnoses. Only three
of 50 cases of malignant disease (6%) had negative impedance imaging diagnoses. The false positive rate of impedance imaging was 17%, while for this
group of patients the false positive rate for mammography was 17.5%.
5.3.2. Circular arrays
5.3.2.1.
Figure 5.1. Reconstructed conductivity (left) and permittivity (right) image of a normal
subject at 125 kHz using Dartmouth generation 1 EIT system.
179
Figure 5.2. (Top) 125 kHz permittivity images in three dierent planes. The left image is
0.5 cm above the lesion, the right one passes through it, and the bottom one is 2 cm below
it. A 3.5 cm tumour is present at 4 oclock. (Bottom) Diagram of where the lesion is located
relative to the three viewing planes [41].
180
The work presented here describes an EIT system that has multi-frequency
broadband capabilities suitable for use in a clinical setting. It possesses a
fast acquisition rate to minimize exam time and includes patient safety
considerations. Also, because of 3D artefacts present in 2D imaging systems
it incorporates 3D measurement capabilities. Its range of frequency is
10 kHz10 MHz, an order of magnitude higher than its predecessor. The
design of a second generation of electronics, based on a digital signal processing (DSP) architecture, centred around a 66 MHz ADSP-21065L SHARC
processor and a recongurable eld programmable gate array (FPGA) operating at frequencies up to 80 MHz. These two devices are reprogrammable,
giving the design an unprecedented level of exibility both in terms of the
algorithms it can execute and the conguration of the digital circuitry. The
signal-level performance of the system shows very signicant improvements
over previous implementations in accuracy, bandwidth and speed. For
example, signal-to-noise ratio (SNR) is better than 80 dB at high frequency,
compared with 6070 dB previously.
With the development of a high-frequency design based on wedgeshaped circuit boards in close proximity to the electrodes, we realized the
breast interface shown in gure 5.3 which consists of four levels of 16 electrodes, where the electronics are integrated with the electrode-positioning
system. The radial translation stages utilize electrode holding rods arranged
in a sliding pattern under stepper motor control. This results in a very
compact unit consisting of 64 channels, with leads from the electrodes to
the electronic cards not exceeding 4 in (10 cm). We integrated the complete
EIS system with a stereotactic biopsy table by tting it into a sliding assembly
that resides on a custom cart designed to dock against the biopsy table. The
system engages tracks mounted under the table and is locked in position
during an EIS exam. The biopsy table is still fully functional for x-ray
exposures for lesion localization and surface ducial marking prior to an
EIS exam.
181
Figure 5.3. Current instrument attached to a stereotactic biopsy table. The unit ts below
the exam table and above the x-ray tube (top left). Four levels of electrode arrays face the
opening in the table (top right). The patient is prone on the table during exams (bottom).
5.3.3.
The rst observation one makes regarding clinical results using impedance
imaging is that it almost exclusively consists of experiments with planar
array devices. Work on the development of such devices seems to have
started around 1979 [38], and this may explain the predominance of these
types of device. In addition, planar devices are generally simpler and do
not require a complex procedure to reconstruct impedance mapsin some
cases no reconstruction at all is used, the impedances sensed by each electrode simply being displayed in the correct arrangement. When reconstruction computations are used they consist of reconstructing impedance maps
at dierent depths and can be performed very rapidly, allowing real-time
updating of the display.
The only data discussed here that is based on a tomographic impedance
device is still preliminary and does not constitute truly a clinical trial. Such a
trial from the same group is under way, however, which should be concluded
in early 2004.
182
REFERENCES
[1] Boring C C, Squires T S and Tong T 1994 Cancer statistics CA Cancer J. Clin. 44 726
[2] Morrow M, Schmidt R, Cregger B and Hasset C 1994 Preoperative evaluation of
abnormal mammographic ndings to avoid unnecessary breast biopsies Arch. Surg.
129 10911096
References
183
[3] Rosenberg R D et al 1998 Eects of age, breast density, ethnicity, and estrogen replacement therapy on screening mammographic sensitivity and cancer stage at diagnosis:
review of 183,233 screening mammograms in Albuquerque, New Mexico Radiology
209 511518
[4] Kerlikowske K, Grady D, Barclay J, Sickles E A and Ernster V 1996 Eect of age,
breast density, and family history on the sensitivity of rst screening mammography
JAMA 276 3338
[5] Elmore J G, Barton M B, Moceri V M, Polk S, Arena P J and Fletcher S W 1998 Tenyear risk of false positive screening mammograms and clinical breast examinations
New England J. Med. 338 10891096
[6] Schaumloel-Schulze U, Heywang-Kobrunner S H, Alter C, Lampe D and
Buchmann J 1999 Diagnostische Vakuumbiopsie der BrustErgebnisse von 600
Patienten Fortschr. Rontgenstr. S1-170 72
[7] Contact: Linda Pointon, MRI Breast Screening Study, Section of Magnetic
Resonance, Royal Marsden NHS Trust, Sutton, Surrey SM2 5PT, UK
[8] Pogue B W, Poplack S P, McBride T O, Wells W A, O.K. S, Osterberg U L and
Paulsen K D 2001 Quantitative hemoglobin tomography with diuse near-infrared
spectroscopy: pilot results in the breast Radiology 218(1) 261266
[9] Franceschini M A, Moesta K T, Fantini S, Gaida G, Gratton E, Jess H, Mantulin
W W, Seeber M, Schlag P M and Kaschke M 1997 Frequency-domain techniques
enhance optical mammography: initial clinical results Proc. Natn. Acad. Sci. USA
94(12) 64686473
[10] Fear E C, Hagness S C, Meaney P M, Okoniewski M and Stuchly M A 2002
Breast tumor detection with near-eld imaging IEEE Microwave Magazine 3 48
56
[11] Van Houten E E W , Doyley M M , Kennedy F E , Weaver J B and Paulsen K D 2003
Initial in vivo experience with steady-state subzone-based MR elastography of the
human breast J. Magn. Res. Imaging 17 7285
[12] NIH Program Project Grant P01-CA80139, 19982003
[13] Fricke H and Morse S 1926 The electrical capacity of tumors of the breast. J. Cancer
Res. 10 340376
[14] Zou Y and Guo Z 2003 A review of electrical impedance techniques for breast cancer
detection. Med. Eng. Phys. 25 7990
[15] Jossinet J, Lobel A, Michoudet C and Schmitt M 1985 Quantitative technique for
bioelectrical spectroscopy J. Biomed. Eng. 7 289294
[16] Jossinet J 1996 Variability of impedivity in normal and pathological breast tissue
Med. Biol. Eng. Comput. 34 34650
[17] Chaudhary S S, Mishra R K, Swarup A and Thomas J M 1984 Dielectric properties of
normal and malignant human breast tissues at radiowave and microwave frequencies
Indian J. Biochem. Biophys. 21 769
[18] Campbell A M and Land D V 1992 Dielectric properties of female breast tissue
measured in vitro at 3.2 GHz Phys. Med. Biol. 37 193210
[19] Heinitz J and Minet O 1995 Dielectric properties of female breast tumors. Proceedings of 9th International Conference on Electrical Bio-Impedance. Heidelberg:
University of Heidelberg, pp 356359
[20] Stelter J, Wtorek J, Nowakowski A, Kopacz A and Jastrzembski T 1998 Complex
permittivity of breast tumor tissue. Proceedings of 10th International Conference
on Electrical Bio-Impedance, Barcelona, pp 5962
184
References
185
Chapter 6
Applications of electrical impedance
tomography in the gastrointestinal tract
Clare Soulsby, Etsuro Yazaki and David F Evans
6.1.
The gastrointestinal tract (GIT) in man comprises a long hollow viscus with
entry at the mouth and exit at the anus. The physiological role of the GIT is
to process and transport nutrient into the organism to act as fuel to sustain
life; it is an essential organ to life. In man, it is a complex series of biologically
active tubes divided into compartments that function dierentially to convert
ingested nutrient into molecules which can be transported across the epithelium into the blood stream. Via the bloodstream, energy is provided to drive
all other body systems.
We can simplify the physiology into three main processes: digestion,
absorption and transit. The structure of the human GIT is shown in gure
6.1 and can be divided into its main compartments. Sphincters (biological
valves) separate the compartments and control transit within and between
the compartments. The residence time in any one compartment varies
widely depending on the function of that compartment. In the oesophagus
the transit time is about 6 s. In the stomach, the residence time of ingesta
can vary from as little as 510 min up to 68 h, depending on the composition
of the meal. These widely variant periods are essential in that they control the
time required to optimize the processes of assimilation of nutrients.
This large variation in gastric residence time can be understood by
explaining the physiology of normal gastric motility. The stomach has two
main functions: (1) to store food, as we can ingest nutrients faster than we
can digest them; (2) to alter the texture of ingesta using physical and chemical
disruption to produce a viscous uid of nely particulate nutrients known as
chyme. This partly processed ingesta is presented to the small intestine in a
suitable consistency for digestion and absorption. In the fed state, the
Figure 6.1.
187
stomach has three phases of motility: receptive relaxation which allows the
stomach to accommodate a large volume of ingesta; mixing, which consists
of strong contractions that agitate and mix stomach contents with acid
and enzymes; and an emptying phase when the antrum grinds food before
releasing the partially digested chyme into the small intestine. Solid foods
take longer to empty than liquids as it takes time to render solids to a suitable
texture for the small intestine. High energy/high fat foods are also emptied in
a controlled way so that they are presented to the small intestine at a rate that
does not exceed digestive capacity. Thus non-nutritive liquids such as water
empty most quickly (gastric emptying half-time approximately 20 min),
nutritive liquids such as milk empty more slowly (gastric emptying halftime approximately 90 min) and large complex meals such as beefburgers
can take up to 360 min. In some cases, food remnants can be found in the
stomach over 8 h after ingestion.
EIT detects alterations in resistivity within thick slices of body tissue. This
principle is utilized to monitor the movement of luminal materials through
dierent compartments of the GIT in order to study normal physiology,
pathophysiology and the eects of transit modifying substances used in the
treatment of gut transit disorders. The term used to describe this movement
is motility. The area of the GIT that has been most widely studied using
EIT is the stomach, and measurements are made of gastric residence and
emptying times of ingested meals. Transit through other compartments such
as the small and large intestine and the rectum have been attempted without
much success. There are few data to support its use in these areas and this
will not be discussed further in this chapter.
188
6.2.
Manometry
Gamma scintigraphy
189
Chemical
This method measures gastric emptying by assessing the time it takes for
certain drugs or markers that are not absorbed in the stomach, but which
are rapidly absorbed from the small intestine, to appear in circulation or
the breath. What is actually measured is the total time including digestion,
absorption and metabolism, as well as the time taken for gastric emptying.
These methods are often referred to as indirect, and an assumption is
made that gastric emptying is the rate limiting step. Substances that have
been used as chemical markers are paracetamol or acetophamine, where
appearance of the marker in the blood is the surrogate for gastric emptying,
or carbon labelled breath tests (acetate, bicarbonate, octanoin and spirulina),
where appearance of the marker in the breath is the surrogate for gastric
emptying.
Paracetamol absorption: A few years ago there was interest in this
method as its non-invasive nature allowed its use in vulnerable subjects
such as critically ill patients. However, as it is used only to measure gastric
emptying of the non-nutrient liquid phase of the meal it is an unsatisfactory
190
Table 6.1.
Test meal
Marker
Liquid
Chicken liver
Egg white
Chicken liver
Puree potato
Liver pate
Pancake
Nondigestible solid particles
Olive oil
Butter
99m
Tc
Tc
99m
Tc
99m
Tc
99m
111
6.3.
ULTRASONOGRAPHY
6.4.1.
191
The vast majority of work performed in this area has been undertaken using
the single frequency Sheeld Mark 1 system developed by Brown and Barber
(1987). More recently, we and others have attempted to adopt the multifrequency system but with less success. This will be discussed later in the
chapter.
6.4.2.
Experimental method
Sixteen electrodes are placed in a circular array around the trunk, at the level
of the eighth costal margin, and are attached to the data collection unit
Figure 6.2.
192
Figure 6.3.
Figure 6.4.
193
Data set acquired after ingestion of 400 ml beef extract drink at 37 8C.
Windows software WIN7 (Boon & Holder). This enabled greater manipulation of data and statistical comparison could be made. Ultimately, all EIT
systems were supplied with Sheeld-designed Windows-based software
and this is currently used by our Unit.
6.4.5.
Figure 6.5. Region of interest drawn around the stomach. This is identied from
summated frames after the study is completed.
194
Figure 6.6. Impedance gastric emptying curve drawn from serial values of ROIs detected
by the data collection unit. Gastric residence and emptying values can be calculated from
these data sets.
6.5.
6.5.1.
195
6.5.2.
6.5.2.1.
Accuracy of EIT
Position of EIT electrodes
The use of the eighth costal cartilage as a valid position for placing electrodes
to measure gastric resistivity by EIT was investigated using 19 healthy
volunteers (Avill et al 1987). Three electrodes were marked with a 57 Co
marker and a drink of soup containing 100 mCi 99m Tctin colloid was
given. Gastric radioactivity was imaged using a -camera. The electrodes
were shown to be situated at the level of the body or the fundus of the
stomach in all 19 subjects.
Another study (Wright 1995) assessed the eect of electrode placement
using six male volunteers. On the day prior to EIT, the shape and position
of the stomach was imaged in ve of the volunteers using 200 ml of orange
juice labelled with 1 MBq 99m TcDTPA. On the following day, EIT electrodes
were placed at the assumed level of the gastric antrum based on the previous
days scintigraphy. A 57 Co marker was taped to the subjects back at the level
of the EIT to determine the position of the electrodes in relation to the
stomach. Following a pre-meal EIT baseline recording a liquid meal of
500 ml Oxo labelled with 2 MBq 99m TcDTPA, simultaneous EIT and
scintigraphy was recorded (EIT at 1 min intervals and scintigraphy at 8 min
intervals). The study ended when less than 30% of the marker, determined
by scintigraphy, remained in the stomach. Of the six subjects, electrodes
were placed high in one subject, low in two and at the antral level in three.
196
In the subjects with low electrodes there was an apparent delay in emptying
recorded by EIT, possibly due to duodenal lling.
6.5.2.2.
The accuracy of EIT to measure volume changes was assessed in six volunteers
(Avill et al 1987) using intragastric balloons. Each volunteer swallowed a nasogastric tube with a balloon positioned 10 cm distal to the gastro-oesophageal
junction. The balloon was serially inated with 50 ml aliquots of air until
the subject felt discomfort or 250 ml had been inserted. Gastric secretions
were aspirated continuously. Images taken at each volume were compared
with a reference image taken with the tube in position, but the balloon deated.
EIT values were directly proportional to the volume of air in the gastric
balloon (r 0:99).
6.5.2.3.
Baseline variation was assessed in four fasted healthy volunteers 90 min after
taking cimetidine. EIT frames were recorded for 45 min at 1 min intervals.
The subjects then drank 500 ml of Oxo. The fasting variation in the areas
of interest corresponding to the position of the stomach identied after
ingestion of the Oxo drink was plotted as a percentage of the maximum
EIT value, obtained immediately after the drink had been ingested. Baseline
variation usually amounted to no more than 10% of values obtained after
ingestion of the Oxo drink, although occasionally variations up to 20%
could occur.
6.5.3.
6.5.3.1.
197
coecient describing the relationship between half-time for the two methods was 0.83 ( p < 0:001).
6.5.3.2.
6.5.3.3.
198
6.5.4.
6.5.4.1.
Wright (1995) assessed gastric emptying of porridge. EIT was compared with
scintigraphy in eight volunteers on two separate occasions at least 14 days
apart, without acid inhibition (control group) or having taken 400 mg of
cimetidine 2 h prior to the study. Following a pre-meal EIT baseline recording, a semi-solid meal of 500 ml of porridge with salt and labelled with 3 MBq
99m
TcDTPA added to the uid prior to cooking, simultaneous EIT and
scintigraphy was recorded (EIT at 1 min intervals and scintigraphy at
10 min intervals). The study continued for approximately 200 min maximum.
The results from 10 studies were available for analysis. In the EIT
control group, the study ended before half-time for gastric emptying was
reached, and there was a signicant dierence between gastric emptying
measured by scintigraphy and EIT ( p 0:04). There was no signicant
dierence in the acid inhibited group. The administration of cimetidine
increased the half-time in the scintigraphy group ( p 0:04). The correlation
between half-times for EIT and scintigraphy was r 0:87, p 0:05 in the
control group, and r 0:89, p 0:04 in the cimetidine group.
6.5.5.
6.5.5.1.
Gastric acid secretion increases markedly during feeding, and it has been
suggested that it may aect the accuracy of measurements obtained by
EIT (Avill et al 1987).
199
200
6.5.6.4.
In another study looking at a beefburger meal (Mangnall et al 1991), scintographic and EIT measurements were carried out simultaneously using a
single -camera (Pho/Gamma III, model 1201, Nuclear Chicago, Europa,
NY). Subjects were fed a radio-labelled beefburger weighing 160 g (vol
137 ml, 20 g fat, 23 g protein, 8 g carbohydrate, 2.5 g NaCl). The beefburger
was labelled with 100 mCi 99m Tcsulphur colloid, which was beaten into the
egg before incorporation into the raw meat. In eight volunteers scintigraphy
was carried out without inhibition of acid secretion; in 12 volunteers gastric
acid secretion was inhibited by administration of 400 mg of cimetidine 60 min
before the test and a further 800 mg at the start. Images of distribution of
radioactivity were collected for 4 h at 5 min intervals, starting before ingestion of the food. EIT measurements were collected at 2 min intervals for
10 min prior to and 4 h after ingestion of the meal. Images were obtained
in 19/20 subjects; in one subject gastric lling was not detected. There was
signicant correlation between the two methods for half (r 0:713,
p < 0:02) and quarter time (r 0:825, p < 0:01), but the results failed to
reach signicance for the lag period (r 0:585) when gastric acid was
inhibited. When gastric acid secretion was not inhibited, EIT was slower
than scintigraphy in 5/8 studies and there was no correlation for half-time
(r 0:058), lag phase (r 0:376). The half-time was within 10% of the
value obtained by scintigraphy in only 2/8 studies.
6.7.
PAEDIATRIC STUDIES
The use of EIT for paediatric applications is highly desirable for both safety
and operational reasons. As EIT is totally non-invasive and does not require
any exposure to ionizing radiations of any kind, it has been welcomed by
paediatricians as a means of assessment of gastric function in infants and
children with suspected foregut dysfunction. Lamont et al (1988) examined
its role in hypertrophic pyloric stenosis. Milla and Ravelli (1994) detected
both gastric stasis and GOR in children with childhood vomiting and
reux, and Nour et al (1994, 1995) performed extensive studies in these
groups.
In children, the limitations of EIT are similar to those in adults. In addition, there are other problems related to the size and overall compliance of
the subjects:
diculty nding sucient space for 16 electrodes on the abdomen of a
small subject;
. certain electrodes (e.g. ECG electrodes) do not give very good conductivity
in children;
. necessary length of recording period for solid test meals;
.
Summary
201
6.8.
The technique has recently been developed as a method for research and
monitoring enteral feed tolerance, particularly in critically ill patients
(Soulsby et al awaiting publication). In the hospital setting enteral feed is
usually administered as a continuous naso-gastric infusion. Enteral feed
tolerance is monitored by aspirating the stomach contents via the nasogastric tube and measuring the volume aspirated, which is known as the
gastric residual volume. If the gastric residual volume is less than a critical
amount, usually 150200 ml, the patient is considered to be tolerating the
feed. This approach has been criticized for being based on assumptions
that are not physiologically sound (McClave and Snider 2002). In fact
there are no available data patterns of gastric emptying during continuous
infusion, other than those hypothesized using mathematical models
(Lin and Van Citters 1997; Burd and Lentz 2001). We have developed the
technique to investigate continuous infusion of enteral feed (Soulsby et al
2003), and to investigate patterns of gastric emptying during naso-gastric
infusion in critically ill patients (Soulsby et al awaiting publication) and
volunteers.
6.9.
SUMMARY
202
3. Semi-solid and solid meals measured by EIT only correlate with scintigraphy when acid is suppressed.
4. The time to reach gastric emptying t1=2 measured by EIT always takes
longer than when measured by scintigraphy, and when gastric acid is
suppressed, the lag phase measured by EIT is signicantly shorter than
when measured by scintigraphy. Thus EIT is more likely to be measuring
gastric volume, including secretions, whereas scintigraphy only measures
gastric emptying of the radio-labelled portion of the meal.
. Although scintigraphy is the gold standard for measurement of gastric
emptying and has been used in the literature to validate EIT, there are
some aws in this approach:
1. Most studies have use of a single marker, but most solid meals are in
fact complex mixtures or particles and have a solid and a liquid
phase. As the most commonly used marker binds to the protein portion
of the meal, the gastric emptying of the other portions (fat, carbohydrate and liquid) are not monitored.
2. Radionuclide markers may separate from the solid phase of the meal
and empty with the liquid phase, resulting in erroneous results.
3. Gastric secretions provide a signicant contribution to the gastric
volume during meals, and inuence gastric emptying patterns by
progressively diluting both liquid and solid markers. External gamma
counting cannot measure the volume of gastric secretion within or
emptied from the stomach, so this important aspect of gastric emptying
is not monitored.
. Thus while it is necessary to compare EIT with the gold standard, lack of
agreement may in fact reect dierences between the dierent methodologies, particularly the inability of scintigraphy to monitor gastric secretions
(Nour et al 1995).
. The literature has used correlation coecients to compare gastric emptying
measured by EIT and scintigraphy. It is probably better to use other
methods, such as a BlandAltman plot, which may aect some of the
conclusions drawn.
6.10.
GENERAL CONCLUSIONS
References
203
204
Gilja O D et al 1997 Intragastric distribution and gastric emptying assessed by three dimensional ultrasonography Gastroenterology 113 3849
Horowitz M and Dent J 1991 Disordered gastric emptying: mechanical basis, assessment
and treatment Ballieres Clinical Gastroenterology 5(2) 371407
Horowitz M et al 1994 Role and integration of mechanisms controlling gastric emptying
Dig. Dis. Sci. 39(12) 713S
Lamont G L et al 1988 An evaluation of applied potential tomography in the diagnosis of
infantile hypertrophic pyloric stenosis Clin. Phys. Physiol. Meas. 9 Suppl A, 6569
Lee J S et al 2000 Toward oce based measurement of gastric emptying in symptonmatice
diabetics using [13C] octanoic breath test Am. J. Physiology 95(10) 27512761
Lin H C and Van Citters G W 1997 Stopping enteral feeding for arbitary gastric residual
volume may not be physiologically sound: results of a computer simulation model J.
Parenteral and Enteral Nutrition 21(5) 286289
Mangnall Y F et al 1991 Applied potential tomography: noninvasive method for measuring gastric emptying of a solid test meal Dig. Dis. Sci. 36(12) 16801684
McClave S A and Snider H L 2002 Clinical use of gastric residual volumes as a monitor for
patients on enteral tube feeding J. Parenteral and Enteral Nutrition 26(6) S43S48
Mushambi M C et al 1992 A comparison of gastric emptying rate after cimetidine and
ranitedine measured by applied potential tomography British J. Clin. Pharmacol.
34 287280
Nour S et al 1991 Measurement of gastric emptying in infants using applied potential
tomography Gut 32 A1233
Nour S et al 1995 Applied potential tomography in the measurement of gastric emptying in
infants J. Paediatric Gastroenterology and Nutrition 20(1) 6572
Piessevaux H et al 2003 Intragastric distribution of a standardised meal in health and functional dyspepsia: correlation with specic symptoms Neurogastroenterology Motility
15 447455
Ravelli A M and Milla J 1994 Detection of gastroesophageal reux by electrical impedance
tomography J. Paediatric Gastroenterology and Nutrition 18(2) 205213
Soulsby C T et al 2003 Measurement of gastric emptying during continuous nasogastric
infusion of enteral feed Clinical Nutrition 22(1) S59S60
Soulsby C T et al (awaiting publication) Real time measurement of enteral feed tolerance in
critically ill patients: is there a role for electric impedance tomographic spectroscopy?
Proc. Nutrition Society
Vantrappen G 1994 (Supplement) Methods to study gastric emptying Dig. Dis. Sci. 39(12)
91S94S
Wright J W 1995 The eect of intraluminal content on gastrointestinal motility in man.
Nottingham, University of Nottingham 98
APPENDIX
Subjects
Test meal
H2 Blockers
Methodology
Results
Avill
(1987)
Eect of acid
inhibition on
repeatability of GE
measured by EIT
8 normals
No inhibition
versus 800 mg
cimetidine
Mangnall
(1991)
Eect of acid
inhibition on
accuracy of GE
measured by EIT
compared with
scintigraphy
20 normals
160 g beefburger
No inhibition
versus 800 mg
cimetidine
Wright
(1995)
Eect of dierent
types of acid
inhibitors on
repeatability of GE
measured by EIT
16 normals
No inhibition
versus 800 mg
cimetidine versus
40 mg omeprazole
Wright
(1995)
Eect of dierent
types of acid
inhibitors on
repeatability of GE
measured by EIT
16 normals
500 ml of
porridge 4.5 g
salt
No inhibition
versus 800 mg
cimetidine versus
40 mg omeprazole
205
Aim of study
Appendix
Study
206
APPENDIX (Continued)
Aim of study
Subjects
Test meal
H2 Blockers
Methodology
Avill
(1987)
Comparison of GE
measured by EIT
versus scintigraphy
8 normals
300 ml consomme
300 ml 100 mCi
99m
Tctin colloid
water
Avill
(1987)
Comparison of GE
measured by EIT
versus dye dilution
10 normals
750 ml 5%
aqueous sucrose
Wright
(1995)
Comparison of GE
measured by EIT
versus scintigraphy,
and eect of acid
inhibition
11 normals
400 mg versus no
acid inhibition
GE measured simultaneously by
EIT versus scintigraphy on two
occasions in each subject, once
with acid inhibition, once
without
Nour
(1995)
Investigation of the
feasibility of using
EIT in infants
47 infants
Formula milk
(25 ml/kg) or
dioralyte
Avill
(1987)
Comparison of GE
measured by EIT
versus scintigraphy
8 normals
85 g mashed
potato 300 ml
100 mCi 99m Tctin
colloid water
800 mg cimetidine
Wright
(1995)
Comparison of GE
measured by EIT
versus scintigraphy
and eect of acid
inhibition
8 normals
500 ml of
porridge 4.5 g
salt labelled with
2 MBq 99m Tc
DTPA
400 mg versus no
acid inhibition
GE measured simultaneously by
EIT versus scintigraphy on two
occasions in each subject, once
with acid inhibition, once
without
Results
Study
Chapter 7
Other clinical applications of electrical
impedance tomography
David Holder
The principal potential clinical applications for biomedical EIT are imaging of
heart and lung function in the thorax, gastric emptying, screening for breast
cancer and brain function. These are all covered by individual chapters elsewhere in this volume. There are several other possible applications, most of
which are now of historical interestthey were started in the rst ush of
enthusiasm when the Sheeld Mark 1 system became available in the mid
1980s, but then active research was discontinued because of inherent technical
problems, or because other areas within EIT appeared more promising.
However, these ideas may still prove to be practicable and worthwhile if
approached in a dierent light, and are reviewed in this chapter.
7.1.
HYPERTHERMIA
208
Moller et al (1993) compared changes within the EIT image with temperature
determined by thermocouples. The tissue was heated to between 35 and 60 8C
as a result in oscillations in a thermoregulatory feedback system. There was a
qualitative correlation between changes in the EIT image and temperature, but
a substantial impedance drift of uncertain origin occurred. A similar study was
performed in a tank lled with conducting agar, into which small pieces of
foam had been inserted in order to simulate inhomogeneous tissue. Heating
was performed with radiofrequency coils (Conway et al, 1992). A linear
relation was observed between EIT image changes and temperature, but the
slopes varied with position in the phantom.
Temperature calibration experiments have also been performed in vivo. In
three volunteers, 200 ml of conducting solutions at various temperatures were
repeatedly introduced into the stomach, whilst EIT images were made from
electrodes around the abdomen (Conway et al, 1992). Acid production was
suppressed by cimetidine. It was found necessary to compensate for baseline
drifts in the images. After compensation, a linear relationship between the
temperature of the infused uid and region of interest integral was observed,
although the slopes varied between subjects.
Unfortunately, reliable clinical use for hyperthermia monitoring
requires a high degree of both spatial and contrast resolution. Single
images in the thigh (Griths and Ahmed, 1987) and over the shoulder
blade (Conway, 1987) of human subjects, with the Sheeld Mark 1
system, during warming, showed substantial artefacts, and it was also
demonstrated in normal volunteers, without warming, that baseline variability would produce impedance changes which were equivalent to temperature changes of several degrees. More recently, some pilot clinical
measurements with planar arrays at 12.5 kHz showed encouraging average
results, but some estimates of tissue temperature were erroneous by 9 8C
(Moskowitz et al, 1995; Paulsen et al, 1996).
Unfortunately, accurate temperature estimation requires not only
accurate imaging, but also an assumed linear relation between temperature
and conductivity. This latter appears to change in a hysteretic fashion
during tissue heating. Given this uncertainty in calibration a priori, and the
baseline variability in vivo, it unfortunately seems that EIT is unlikely to
be an accurate technique unless there are substantial improvements in
system performance (Blad et al, 1992; Paulsen et al, 1996).
7.2.
209
changes. EIT images were collected with a ring of electrodes around the
pelvis, as the subject was placed in horizontal and vertical positions using
a tilt table. The rationale was that this should produce uid shifts in the
pelvis. A central area of impedance change was observed in both normals
and subjects, with pelvic congestion diagnosed by venography. A signicant
dierence in the ratio of the areas anterior and posterior to the coronal
midline and greater than 10% of the peak impedance change was observed.
No dierence in mean amplitude of impedance changes was observed
between the two groups. Venography is an invasive procedure, so EIT
would provide a welcome alternative. However, there is no direct evidence
concerning the origin of these changes, although it has been shown that
they are at least plausible by comparison with EIT images made in tanks
with saline-lled tubing (Thomas et al, 1994). This is an intriguing and potentially valuable application, but larger prospective studies will be needed
before its use can be established.
7.3.
210
REFERENCES
Blad B, Persson B and Lindstrom K 1992 Quantitative assessment of impedance tomography for temperature measurements in hyperthermia Int. J. Hyperthermia 8 3343
Conway J 1987 Electrical impedance tomography for thermal monitoring of hyperthermia
treatment: an assessment using in vitro and in vivo measurements Clin. Phys. Physiol.
Meas. 8 Suppl A 141146
Conway J, Hawley M, Mangnall Y, Amasha H and van Rhoon G C 1992 Experimental
assessment of electrical impedance imaging for hyperthermia monitoring Clin. Phys.
Physiol. Meas. 13 Suppl A 185189
Griths H and Ahmed A 1987 A dual-frequency applied tomography technique: computer
simulations Clin. Phys. Physiol. Meas. 8 103107
Hughes T A, Liu P, Griths H, Lawrie B W and Wiles C M 1996a Simultaneous electrical
impedance tomography and videouoroscopy in the assessment of swallowing
Physiol. Meas. 17 109119
Hughes T A, Liu P, Griths H and Wiles C M 1996b Repeatability of indices of swallowing in healthy adults: electrical impedance tomography compared with a simple timed
test of swallowing Med. Biol. Eng. Comput. 34 366368
Kulkarni V, Hutchison J M and Mallard J R 1989 The Aberdeen Impedance Imaging
System. Biomed. Sci. Instrum. 25 4758
Kulkarni V, Hutchison J M, Ritchie I K and Mallard J R 1990 Impedance imaging in
upper arm fractures J. Biomed. Eng. 12 219227
Moller P H, Tranberg K G, Blad B, Henriksson P H, Lindberg L, Weber L and Persson B
R R 1993 Electrical impedance tomography for measurement of temperature distribution in laser thermotherapy (laserthermia), in Clinical and Physiological Applications of Electrical Impedance Tomography, ed D S Holder (London: UCL Press)
Moskowitz M J, Ryan T P, Paulsen K D and Mitchell S E 1995 Clinical implementation of
electrical impedance tomography with hyperthermia Int. J. Hyperthermia 11 141149
Paulsen K D, Moskowitz M J, Ryan T P, Mitchell S E and Hoopes P J 1996 Initial in vivo
experience with EIT as a thermal estimator during hyperthermia Int. J. Hyperthermia
12, 573591
Ritchie I K, Chesney R B, Gibson P, Kulkarni V and Hutchison J M 1989 Impedance
osteography: a technique to study the electrical characteristics of fracture healing
Biomed. Sci. Instrum. 25 5977
Sadleir R J and Fox R A 2001 Detection and quantication of intraperitoneal uid using
electrical impedance tomography. IEEE Trans. Biomed. Eng. 48 484491
Thomas D C, McArdle F J, Rogers V E, Beard R W and Brown B H 1991 Local blood
volume changes in women with pelvic congestion measured by applied potential
tomography Clin. Sci. (Lond.) 81 401404
Thomas D C, Siddall-Allum J N, Sutherland I A and Beard R W 1994 Correction of the
non-uniform spatial sensitivity of electrical impedance tomography images Physiol.
Meas. 15 Suppl 2a A147-A152
Vonk N A, Kunst P W, Janse A, Smulders R A, Heethaar R M, Postmus P E, Faes T J and
de Vries P M 1997 Validity and reproducibility of electrical impedance tomography for
measurement of calf blood ow in healthy subjects Med. Biol. Eng. Comput. 35 107112
Woo E J, Hua P, Webster J G and Tompkins W J 1992 Measuring lung resistivity using
electrical impedance tomography IEEE Trans. Biomed. Eng. 39 756760
PART 4
NEW DIRECTIONS
Chapter 8
Magnetic induction tomography
H Griths
8.1.
INTRODUCTION
214
8.2.
There are two contributions to the signal detected by the sensing coil. The
rst is directly induced by the eld from the excitation coil (the primary
signal, B). The second is from the eddy currents induced in the material
which in turn produce their own magnetic eld (the secondary signal, B).
For a sinusoidally-time-varying excitation at angular frequency !, the
skin depth of the electromagnetic eld in the material is given by
215
Figure 8.1. Phasor diagram representing the primary (B) and secondary (B) magnetic
elds detected. The total detected eld (B B) lags the primary eld by an angle .
2=!0 r 1=2 , where and r are the conductivity and relative permeability of the material and 0 is the permeability of free space. If is large
compared with the thickness of the sample, which will normally be so for
biological tissues,
B
P!0 !"0 "r j Qr 1
8:1
B
where "r is the relative permittivity of the material, "0 is the permittivity of
free space and P and Q are geometrical constants (Scharfetter et al 2003).
Thus, the conduction currents induced in the sample give rise to a component
of B, which is proportional to frequency and conductivity and is imaginary
and negative, meaning that it lags the primary signal by 908. Displacement
currents cause a real (in-phase) component proportional to the square of
the frequency. A non-unity relative permeability also gives rise to a real
component, but with a value independent of frequency. The primary and
secondary signals can be represented by the phasor diagram shown in
gure 8.1.
Because for biological tissues B is much smaller in magnitude than B
and is normally dominated by the conductivity term, the phase angle can be
written
B
/ !:
8:2
B
Hence, a higher frequency of excitation will increase the size of the signal.
For a metal sample, where the conductivity is high and the permittivity
negligible, will be much smaller than the thickness of the sample and the
behaviour of B=B departs from the proportionality given in equation
(8.1). Its value will be much larger than for the same volume of biological
tissue, and it will contain not just a negative imaginary part but also a negative real part as the sample tends to act as a screen (Tapp and Peyton 2003).
8.3.
216
Figure 8.2. A practical MIT system, operating at 10 MHz (after Watson et al 2002b). The
16 coils are mounted inside a cylindrical electromagnetic screen of aluminium. The circuit
boards of the transceivers are enclosed in metal boxes xed to the outside of the screen.
217
Figure 8.3. Designs of (a) spiral coil and (b) comb screen for printed circuit board
fabrication (after Peyton et al 2002).
218
the best way of quantifying this error needs to be established. The whole
topic of screening in MIT and the determination of what is optimal deserves
much more study.
8.4.
SIGNAL DEMODULATION
In order to exploit the fact that the conduction signal is in quadrature with
the primary signal, phase-sensitive detection is normally used for demodulation (Yu et al 1994, Griths et al 1999, Scharfetter et al 2001). Commercial
lock-in ampliers have provided an o-the-shelf solution incorporating a
vector voltmeter (phase-sensitive detector), analogue-to-digital conversion
and digital ltering (Riedel et al 2002, 2004, Watson et al 2002b, 2004,
00
Ulker and Gencer 2002, Karbeyaz and Gencer 2003). Phase-sensitive detection can discriminate between the conduction signal and any residual signal
due to capacitive coupling (see section 8.3), as the latter is known to aect
predominantly the real part. Customized circuitry for direct digitization of
the high-frequency signal is likely to become a viable, cost-eective option
with the appearance on the market of new, fast, high-resolution, analogueto-digital converters.
An alternative method of demodulation, advocated by Korzhenevskii
and Cherepenin (1997), is to measure the phase angle directly as it will be
proportional to sample conductivity [equation (8.2)]. The method has been
implemented by passing the signal and a reference waveform through zerocrossing detectors and feeding the resulting signals to an exclusive-OR
gate; the output pulse width will then be proportional to the phase dierence
(Korjenevsky et al 2000, Watson et al 2002a).
Watson et al (2001b) identied three indices of error in MIT demodulators: phase noise, phase drift and phase skew (phase skew being an apparent
change in phase caused by a change in signal amplitude). With exclusive-ORbased, direct-phase measurements, the three indices were compared for
dierent limiter amplier circuits (Watson et al 2002a). In a further study,
direct phase measurement was compared with a vector-voltmeter method
in respect of the same three indices (Watson et al 2003); the two methods
had comparable noise and skew values, but the drift was found to be greater
in the direct-phase system.
8.5.
Because the secondary signal has to be detected against the much larger
primary signal, various methods have been tried for backing o the primary
signal, i.e. for subtracting the phasor B in gure 8.1, such that with no sample
present all recorded signals should be zero. This then allows the gain of the
219
220
8.6.1.
The majority of MIT systems for the process industry have been designed for
detecting metallic or ferromagnetic objects which, having either a high electrical conductivity or a high permeability, can produce large signals with an
excitation frequency of 500 kHz or below.
221
222
Biomedical MIT
223
224
Figure 8.4. MIT images obtained with the 10 MHz system of Watson et al (2002b) for
4 cm diameter cylinder of agar, conductivity 1 S m1 , in a 20 cm diameter bath of saline,
conductivity 0.3 S m1 . (a) Diagram indicating position of agar; the thickness of the air
gap between the saline bath and the coils (white ring) was 3.5 cm. (b) Absolute images
reconstructed relative to empty space, 40 singular values. (c) Dierence images
reconstructed from the dierence in measurements with and without the agar present,
50 singular values. Only positive image values are displayed.
(a)
(b)
Figure 8.5. Human in vivo images obtained with the Moscow 16-coil 20 MHz MIT system
(after Korjenevsky and Sapetsky 2001). (a) Dierence image of the thorax (inhalation
exhalation) reconstructed by weighted back-projection. The authors interpret features 1
and 2 as the left and right lungs, and 3 as chest movement artefact. (b) Absolute image
of the head (referenced to empty space) reconstructed by articial neural network in
which the two bright (high conductivity) features are interpreted as the lateral ventricles
of the brain.
Image reconstruction
225
8.7.
IMAGE RECONSTRUCTION
226
Image reconstruction
227
distribution (e.g. uniformity) and solving the forward problem for all excitor/
sensor combinations. Each voxel is then perturbed by a small amount (e.g.
1%) and the whole computation repeated for all such voxels in turn. As
has been pointed out in the context of EIT, such a method is computationally
very time-consuming and several authors have now described more ecient
methods specically for MIT. Gencer and Tek (1999) derived a method for
computing the sensitivity involving the impressed vector potential and a
derivative of the scalar potential. Two papers have described rapid computation of the sensitivity matrix by what is in eect the Gezelowitz sensitivity
formula extended to take account of changes in conductivity, permittivity
and permeability, and the fact that the electric eld contains magneticallyinduced components as well as that arising from the gradient of the scalar
potential (Lionheart et al 2003, Hollaus et al 2004). The methods require
only two solutions of the forward problem for each coil pair, rst exciting
one coil and then the other.
The articial neural network method used by Korjenevsky and Sapetsky
(2001) to produce in vivo images (see section 8.6.2) is sometimes criticized for
not being based on any underlying physical principles and depending for its
accuracy on the training data. However, the method does not assume linearity,
can be implemented with speed and may well prove valuable for practical MIT
applications.
There is a general consensus that, in order for MIT to advance signicantly, the non-linear inverse problem will need to be solved in three dimensions. In contrast to the linear iterative methods, the NewtonRaphson or
GaussNewton method will be used and the Jacobian (sensitivity matrix)
recomputed at each iteration from the most recent estimate of the conductivity
distribution (Lionheart 2004). Soleimani et al (2003) have illustrated such a
method for a simulated, eight-coil, annular, MIT array and produced
images of a simulated copper bar. The Tikhonov-regularized solution was
used at each linear step. Tamburrino et al (2003) described an interesting
non-iterative, nonlinear algorithm using the concept of a resistance matrix
for ERT and showed how it could be modied for MIT, but no illustrations
of imaging were presented.
Because of the ill-posedness of the inverse problem, several workers
have pointed out the likely advantages in introducing a priori information
to constrain the inverse solution, and this can be done in a number of
ways. A non-negativity constraint and regularization are both common
examples of the use of a priori information, the former because it disallows
physically-impossible, negative values of conductivity and the latter because
it restricts the dierences in conductivity between neighbouring voxels in the
image to a physically acceptable level. A priori information can also be
introduced by conning the solution to a certain class of problems or by
introducing shape information determined by some other method. Bissesseur
and Peyton (2001) described nonlinear, iterative, image reconstruction,
228
8.8.
229
will depend on the phase noise in the system, but that a higher noise level
can be tolerated at higher frequencies. From numerical simulations,
Morris et al (2001) proposed a phase measurement precision of 3 m8 (millidegrees) in order to resolve the internal conductivity features in some
simple models of biological tissues at 10 MHz. This gure reects the very
high measurement precision required of MIT. A phase dierence of this
value amounts to a time dierence of only 1 ps. Light travels less than
1 mm in this time!
In measurements at 10 MHz, Griths et al (1999) reported a maximum
phase shift of 0.02 radian for cylinder of 2 S m1 saline solution. The noise
level was <104 radian <6 m8) for an integrating time of 480 ms. The
random noise gure was not the largest source of phase error as baseline
drifting was sometimes equivalent to 100 m8 over a period of 10 min.
The MIT system with the highest operating frequency yet reported is the
20 MHz system of Korjenevsky et al (2000). The noise level was quoted as
5 103 radian (280 m8), with an integrating time of 4 ms per individual
measurement. The maximum phase shift to be measured was approximately
0.06 radian for a 10 cm diameter cylinder of 3.5 S m1 saline solution in the
centre of the 35 cm diameter coil array. In a more recent publication, an
improvement in the phase noise of the system to 1:5 103 radian (86 m8)
was reported (Korjenevsky and Sapetsky 2001).
Watson et al (2002b) report a gure of 30 m8 combined phase noise and
drift for their 10 MHz multichannel system. The time taken per measurement
was very long, 560 ms, being limited not by the integration time but by the
lock-in time of the amplier.
These noise gures are all signicantly greater than the goal of 3 m8
proposed by Morris et al (2001). In a recent paper, however, Gough (2003)
described a novel method employing two stages of phase-sensitive detection,
and in a single channel operating at 8 MHz, with an integrating time of
100 ms, achieved a noise gure of 1.5 m8 and a drift of only 10 m8 over a
whole day.
Using a planar gradiometer at 150 kHz with an integrating time of
100 ms, Scharfetter et al (2001a) reported a noise level of 8 105 radian
(5 m8) and a typical signal level of 4 103 radian (230 m8). For a single
coil sensor, the signal-to-noise ratio was 20 dB lower than with the gradiometer. When measuring a biological sample, the signal-to-noise ratio was
increased by a further 36 dB by a mechanical chopping of the signal, achieved
by bringing the sample in and out of the eld of view at a frequency of 1 Hz.
Such a technique would not be possible when performing MIT imaging.
It is dicult to judge whether the noise gures achieved by the various
hardware designs so far developed will be adequate for biomedical MIT
imaging. Further detailed modelling studies of the type performed by
Merwa et al (2004) are now needed to determine the required performance
for specic imaging applications.
230
8.9.
8.10.
MULTI-FREQUENCY MEASUREMENTS
231
phase shifts unless care is taken to keep resonances well away from the
frequency band of operation.
8.11.
It was demonstrated some time ago that samples of high permeability, such
as ferrite, could readily be visualized by MIT (see section 8.6.1).
For biological tissues, very little work has so far been carried out in
measuring permittivity and permeability as the signals are so small relative
to the already-small conductivity signal. However, phase-sensitive detection
provides a means of separating the conductivity signal, appearing in the
imaginary part, from the permittivity and pearmeability signals in the real
part, provided that system errors such as electric-eld coupling can be
reduced to a suciently low level. Researchers are now beginning to attempt
such measurements. Because the permittivity signal is proportional to the
square of frequency [equation (8.1)], larger signals might be expected at
high excitation frequencies, but this gain will be oset by the fall in relative
permittivity with increasing frequency exhibited by all biological tissues.
Measuring at 10 MHz, Watson et al (2003a) obtained values for the relative
permittivity of a water sample and an average for a human thigh in vivo.
232
Scharfetter et al (2003) evaluated the stability and sensitivity requirements of an inductive sensor for measuring the magnetic susceptibility
(m r 1) of liver with a view to detecting hepatic iron overload. The
susceptibility of liver tissue is very small and ranges from a normal value
of 9 106 to 5 106 in overload. The authors calculated that a very
narrow receiver bandwidth (<1 Hz) would be needed to achieve the necessary
signal-to-noise ratio for such measurements. They furthermore calculated
that for water (m 10 106 , "r 80), the permittivity contribution to
the signal at 50 kHz would be four orders of magnitude lower than that of
magnetic susceptibility, but that for liver tissue the much higher relative
permittivity (104 ) would reduce this dierence. Single-channel practical
measurements at 28 kHz (not imaging) combined with modelling have now
been reported, and showed that whilst the susceptibility of water could be
measured inductively, for human measurements in vivo, the contribution of
the permittivity outweighed that of the magnetic susceptibility and made
the measurement more dicult (Casanas et al 2004). The authors suggested
the use of multiple frequencies to improve the accuracy of the measurement
by exploiting the dierent frequency-dependencies of the permittivity and
permeability terms in equation (8.1).
Imaging permeability might also be applicable, and with somewhat
larger signals than from the bodys natural magnetism, by using magnetic
material as a tracer. Forsman (2000) introduced particles of iron oxide
(Fe2 O3 ) in a meal and was able to observe gastrointestinal mobility with a
magnetometer detection system (again not imaging). This oers the tantalizing possibility of contrast enhancement in biomedical MIT.
8.12.
CONCLUSIONS
References
233
ACKNOWLEDGEMENTS
I am grateful to A Korjenevsky, A J Peyton and H Scharfetter for their
permission to reproduce gures 8.3, 8.5 and 8.6, and to S Watson for providing gures 8.2 and 8.4. I am further indebted to A J Peyton for constructive
criticism of the rst draft.
REFERENCES
Al-Zeibak S and Saunders N H 1993 A feasibility study of in vivo electromagnetic imaging
Phys. Med. Biol. 38 15160
Binns R, Lyons A R A and Peyton A J 2001 Imaging of molten steel ow proles Meas.
Sci. Technol. 12 11328
234
References
235
Griths H, Stewart W R and Gough W 1999 Magnetic induction tomography: a measuring system for biological tissues Ann. NY Acad. Sci. 873 33545
Griths H 2001 Magnetic induction tomography Meas. Sci. Technol. 12(8) 112631
Hammer E A, Abro E, Cimpan E and Yan G 2001 A high frequency magnetic eld probe
for determination of interface levels in separation tanks Proc. SPIE 4188 2949
Hammer E A and Fossdal G 2002 A new water-in-oil monitor cased on high frequency
magnetic eld excitation Proc. 2nd International Symposium Process Tomography,
Wroclaw, Poland, 1112 Sept, pp 916
Hammer E A, Pettersen F and Ndseth A 2003 Numerical simulation of eddy current
losses in high frequency magnetic eld water fraction meters Proc. 3rd World
Congress on Industrial Process Tomography, Ban, Canada, ISBN 0853162409,
pp 34751
Higson S R, Drake P, Stamp D W, Peyton A, Binns R, Lyons A and Lionheart W 2002
Development of a sensor for visualisation of steel ow in the continuous casting
nozzle Proc. 21st Int. ATS Steelmaking Conf., Paris, 1112 Dec
Holder D (ed) 1993 Clinical and Physiological Applications of Electrical Impedance Tomography (London: UCL Press)
Hollaus K, Magele C, Merwa R and Scharfetter H 2004 Fast calculation of the sensitivity
matrix in magnetic induction tomography by tetrahedral edge nite elements and the
reciprocity theorem Physiol. Meas. 25 15968
Karbeyaz B U and Gencer N G 2003 Electrical conductivity imaging via contactless
measurements: an experimental study IEEE Trans. Med. Imag. 22(5) 62735
Koksal A, Eyuboglu M and Demirbilek M 2002 A quasi-static analysis for a class of
induced-current EIT systems using discrete coils IEEE Trans. Med. Imag. 21(6)
68894
Korjenevsky A, Cherepenin V and Sapetsky S 2000 Magnetic induction tomography:
experimental realisation Physiol. Meas. 21(1) 8994
Korjenevsky A 2001 Solving inverse problems in electrical impedance and magnetic induction tomography by articial neural networks J. Radioelectronics no 12
Korjenevsky AV, Cherepenin VA and Sapetsky SA 2001 Magnetic induction tomographynew imaging method in biomedicine Proc. 2nd World Congress on Industrial
Process Tomography, Hannover, Germany, pp 2406
Korjenevsky A 2003 http://www.cplire.ru/html/tomo/mitimage.html
Korzhenevskii A V and Cherapenin V A 1997 Magnetic induction tomography J. Comm.
Tech. Electronics 42(4) 46974
Korzhenevsky A and Sapetsky S 2001 Visualisation of the internal structure of extended
conducting objects by magnetoinduction tomography Bull. Russian Acad. Sciences:
Physics 65(12) 19459
Levy S, Adam D and Bresler Y 2002 Electromagnetic impedance tomography (EMIT): a
new method for impedance imaging IEEE Trans. Med. Imag. 21(6) 67687
Lionheart B 2001 Reconstruction algorithms for permittivity and conductivity imaging
Proc. 2nd World Congress on Industrial Process Tomography, Hannover, Germany,
pp 411
Lionheart W R B, Soleimani M and Peyton A J 2003 Sensitivity analysis of 3D magnetic
induction tomography (MIT) Proc. 3rd World Congress on Industrial Process Tomography, Ban, Canada, ISBN 0853162409, pp 23944
Lionheart W R B 2004 EIT reconstruction algorithms: pitfalls, challenges and recent
developments Physiol. Meas. 25 12542
236
Ma X, Peyton A J, Binns R and Higson SR 2003 Imaging the ow prole of molten steel
through a submerged pouring nozzle Proc. 3rd World Congress on Industrial Process
Tomography, Ban, Canada, ISBN 0853162409, pp 73642
Matoorian N, Patel BCM and Bowler AM 1995 Dental electromagnetic tomography:
properties of tooth tissues IEE Colloquium Digest 1995/099, pp 3/13/7
Merwa R, Holaus K, Brandtatter B and Scharfetter H 2003 Numerical solution of the
general 3D eddy current problem for magnetic induction tomography (spectroscopy)
Physiol. Meas. 24 54554
Merwa R, Hollaus K, Oszkar B and Scharfetter 2004 Detection of brain oedema using
magnetic induction tomography: a feasibility study of the likely sensitivity and detectability Physiol. Meas. 25 34754
Metherall P, Barber D C, Smallwood R H and Brown B H 1996 Three-dimensional electrical impedance tomography Nature 380 50912
Morris A and Griths H 2001 A comparison of image reconstruction in EIT and MIT by
inversion of the sensitivity matrix Abstract of 3rd EPSRC Engineering Network,
London, 46 April
Morris A, Griths H and Gough W 2001 A numerical model for magnetic induction
tomographic measurements in biological tissues Physiol. Meas. 22 1139
Netz J, Forner E and Haagemann S 1993 Contactless impedance measurement by
magnetic inductiona possible method for investigation of brain impedance Physiol.
Meas. 14 26371
Noel M and Xu B 1991 Archaeological investigation by electrical resistance tomography: a
preliminary study Geophys. J. Int. 107 95102
Peyton, A J, Yu Z Z, Lyon G, Al-Zeibak S, Ferreira J, Velez J, Linhares F, Borges A R,
Xiong H L, Saunders N H and Beck M S 1996 An overview of electromagnetic inductance tomography: description of three dierent systems Meas. Sci. Technol. 7 26171
Peyton A J, Beck M S, Borges A R, de Oliveira J E, Lyon G M, Yu Z Z, Brown M W and
Ferrerra J 1999 Development of electromagnetic tomography (EMT) for industrial
applications. Part 1: Sensor design and instrumentation Proc. 1st World Congress
on Industrial Process Tomography, Buxton, UK, 1417 April, pp 30612
Peyton A J, Mackin R, Goss D, Crescenzo E and Tapp H S 2002 The development of high
frequency electromagnetic inductance tomography for low conductivity materials
Proc. 2nd International Symposium Process Tomography, Wroclaw, Poland, 1112
Sept, pp 2540
Peyton A J, Watson S, Williams R J, Griths H and Gough W 2003 Characterising the
eects of the external electromagnetic shield on a magnetic induction tomography
sensor Proc. 3rd World Congress on Industrial Process Tomography, Ban, Canada,
ISBN 0853162409, pp 3527
Pham M H, Hua Y and Gray N B 1999 Eddy current tomography for metal solidication
imaging Proc. 1st World Congress on Industrial Process Tomography, Buxton, UK,
1417 April, pp 4518
Radai M R, Zlochiver S, Rosenfeld M and Abboud A 2003 Combined injected and induced
current approaches in EITa simulation study. Abstracts of 4th Conference on
Biomedical Applications of Electrical Impedance Tomography, UMIST, Manchester,
2325 April, p 33
Ramli S and Peyton A J 1999 Feasibility study for planar-array electromagnetic inductance
tomography (EMT) Proc. 1st World Congress on Industrial Process Tomography,
Buxton, UK, 1417 April
References
237
238
Tarjan P P and McFee R 1968 Electrodeless measurements of the eective resistivity of the
human torso and head by magnetic induction IEEE Trans. Biomed. Eng. BME-15
26678
Tozer J C, Ireland R H, Barber D C and Barker A T 1998 Magnetic impedance tomography Proceedings of 10th Int. Conf. on Electrical Bioimpedance, Barcelona, Spain, 59
April, pp 36972
00
Ulker B and Gencer NG 2002 Implementation of data acquisition system for contactless
conductivity imaging IEEE Engineering in Medicine and Biology Magazine 21(5)
1525
Watson S, Williams R J, Griths H, Gough W and Morris A 2001a A transceiver for
direct phase measurement magnetic induction tomography Proc. 23rd Ann. Int.
Conf. IEEE EMBS, Istanbul, Turkey, 2528 Oct, Paper 942
Watson S, Williams R J, Gough W, Morris A and Griths H 2001b Phase measurement in
biomedical magnetic induction tomography Proc. 2nd World Congress on Process
Tomography, Hannover, Germany, 2931 Aug, pp 51724
Watson S, Williams RJ, Griths H, Gough W and Morris A 2002a Frequency downconversion and phase noise in MIT Physiol. Meas. 23 18994
Watson S, Williams R J, Morris A, Gough W and Griths H 2002b The Cardi magnetic
induction tomography system Proc. Int. Fed. Med. Biol. Eng. EMBEC02, Vienna,
Austria, 48 Dec, ISBN 3-901351-62-0, vol. 3, Part 1, pp 1167
Watson S, Williams R J, Griths H, Gough W and Morris A 2003a Magnetic induction
tomography: phase vs. vector voltmeter measurement techniques Physiol. Meas. 24
55564
Watson S, Williams R J, Griths H and Gough W 2004 A primary eld compensation
scheme for planar array magnetic induction tomography Physiol. Meas. 25 2719
Williams R A and Beck M S (eds) 1995 Process Tomography Principles, Techniques and
Applications (Oxford: Butterworth Heinmann) 581 pp
Yu Z Z, Peyton A J, Conway W F, Xu LA and Beck M S 1993a Imaging system based on
electromagnetic tomography (EMT) Electron. Lett. 29 6256
Yu Z, Lyon G, Al-Zeibak S, Peyton A J and Beck M S 1993b A review of electromagnetic
tomography at UMIST IEE Colloquium Digest 1995/099, pp 2/12/5
Yu Z Z, Peyton A J and Beck M S 1994 Electromagnetic tomography (EMT), Part I:
Design of a sensor and a system with a parallel excitation eld Proc. European
Concerted Action in Process Tomography, Oporto, Portugal, 2426 March, pp 14754
Yu Z Z, Peyton A J and Beck M S 1995 Optimum excitation eld for non-invasive electrical and magnetic tomography sensors Proc. European Concerted Action in Process
Tomography, Bergen, Norway, ISBN 0-9523165-2-8, pp 31120
Zlochiver S, Radai M M, Abboud S, Rosenfeld M, Dong X-Z, Liu R-G, You F-S, Xiang
H-Y and Shi X-T 2004 Induced current electrical impedance tomography system:
experimental results and numerical simulations Physiol. Meas. 25 23955
Chapter 9
Magnetic resonance electrical impedance
tomography (MREIT)
Eung Je Woo, Jin Keun Seo and Soo Yeol Lee
9.1
INTRODUCTION
240
the magnetic eld inside the subject can be measured by a non-contact method.
The second is how to utilize this internal information in resistivity image
reconstructions. This initiated the research area called magnetic resonance
electrical impedance tomography (MREIT).
Since the late 1980s, measurements of the internal magnetic ux density
due to an injection current have been studied by Joy et al (1989) and Scott
et al (1991, 1992). This requires a magnetic resonance imaging (MRI) scanner
as a tool to capture internal magnetic ux density images. Once we obtain the
magnetic ux density B Bx ; By ; Bz due to an injection current I, we can
produce an image of the corresponding internal current density distribution
J from the Ampe`res law J r B=0 , where 0 is the magnetic permeability of the free space. For this reason, this technique has been called
magnetic resonance current density imaging (MRCDI) and suggested as a
technically feasible way to answer the rst question on the measurement
method.
Combining EIT and MRCDI techniques, the basic concept of MREIT
was proposed by Zhang (1992), Woo et al (1994) and Ider and Birgul
(1998). In MREIT, we measure the induced magnetic ux density B inside a
subject due to an injection current I using an MRI scanner. Then, we may
compute the internal current density J as is done in MRCDI. From B and/
or J, we can perceive the internal current pathways due to the resistivity distribution to be imaged. This is the main reason why MREIT could eliminate the
ill-posedness of EIT, as shown in gure 9.1.
However, if we try to use J r B=0 , there occurs a serious technical
problem in measuring all three components of B. Since any currently
available MRI scanner measures only one component of B that is parallel
to the direction of the main magnetic eld of the MRI scanner, measuring
all three orthogonal components of B Bx ; By ; Bz requires subject rotations. In this chapter, we assume that z-axis is the direction of the main
magnetic eld. Since these subject rotations are impractical and also cause
other problems such as misalignments of pixels, it is highly desirable to
reconstruct resistivity images from only Bz instead of B. Therefore, most
recent MREIT techniques focus on analysing the information embedded in
the measured Bz data to extract any constructive relations between Bz and
the current density or resistivity distribution to be imaged.
Though there are still several technical problems to be solved, MREIT
has the potential to provide cross-sectional resistivity images with better
accuracy and spatial resolution. Reconstructed static resistivity images will
allow us to obtain internal current density images for any arbitrary injection
currents and electrode congurations. Potential clinical applications of
MREIT include functional imaging, neuronal source localization and
mapping, optimization of therapeutic treatments using electromagnetic
energy and so on. Images from MREIT may also be used as a priori information in EIT image reconstructions for better results. The disadvantages of
Introduction
241
(a)
(b)
Figure 9.1. (a) EIT using only boundary measurements. (b) MREIT using both internal
and boundary measurements.
MREIT over EIT may include the lack of portability, potentially long
imaging time and requirement of an expensive MRI scanner.
This chapter addresses the image reconstruction problem in MREIT as
a well-posed inverse problem taking advantage of the information on
internal magnetic ux density distributions. Assuming that the magnetic
ux density B Bx ; By ; Bz or only Bz is available, a mathematical formulation for the MREIT problem is presented to explain the fundamental
concept. As a basic tool in experimental design and verication, as well as
development of image reconstruction algorithms, a 3D forward solver for
MREIT is discussed. Measurement methods in MREIT are explained
based on MRCDI techniques, including data collection and processing
methods. Following the discussion on the uniqueness of a reconstructed
resistivity image, several image reconstruction algorithms are described
including the J-substitution, current constrained voltage scaled reconstruction (CCVSR), harmonic Bz algorithm and others. Practical limitations in
terms of the spatial resolution and accuracy of reconstructed images are
discussed based on the noise analysis of the measured magnetic ux density
distribution. At the end of this chapter, possible applications and future
research directions are summarized.
242
9.2.
Figure 9.2 shows an electrically conducting domain with its boundary @.
We denote two electrodes attached on @ as E 1 and E 2 . Lead wires carrying
an injection current I are denoted as L1 and L2 . Then, we can formulate the
following boundary value problem with the Neumann boundary condition:
8
1
>
>
rVr 0 in
<r
r
9:1
>
1
>
: rV n g on @
where and V are the resistivity and voltage distribution in , respectively, n
is the outward unit normal vector on @ and g is a normal component of the
current density on @ due to the injection current I. A position vector in R3 is
denoted as r. On the current injection electrode E j for j 1 or 2, we have
E j g ds I, where the sign depends on the direction of current and g is
zero on the regions of boundary not contacting with the current injection
electrodes. It is well known that rV 2 L2 is uniquely determined by
and g. Setting a reference voltage Vr0 0 for r0 2 @, we can obtain a
unique solution V of (9.1). Knowing the voltage distribution V, the current
density J is given by
Jr
1
1
rVr
Er
r
r
in
9:2
9:3
r r0
dv0
9:4
B r 0 Jr0
4
jr r0 j3
Figure 9.2. Electrically conducting subject with a pair of electrodes E 1 and E 2 . Lead
wires are denoted as L1 and L2 . Note that more than two electrodes are needed in
MREIT to inject at least two currents, as described in section 9.5.
Problem denition
BE r
0
4
BL r
0 I
4
Jr0
E
243
r r0
dv0
j r r0 j 3
9:5
r r0
dl 0
jr r0 j3
9:6
and
ar0
1
r Br in :
0
9:7
We must have
1
1
rVr
r Br
0
r
and
r Jr 0
in :
9:8
Figure 9.3. MREIT system block diagram. Resistivity, voltage, current density and
magnetic ux density are denoted as , V, J and B, respectively. Quantities from the
imaging subject are shown with superscripts .
244
9.3.
As for the case of EIT, we need a forward solver in MREIT for algorithm
development, experimental design and verication. Since the image reconstruction problem in MREIT is inherently 3D, we describe a 3D forward
solver computing distributions of voltage V, current density J and magnetic
ux density B, all within an electrically conducting domain (Lee et al 2003b).
In real MREIT experiments, it would be desirable to use recessed electrodes
as suggested by Lee et al (2003b) and Oh et al (2003). Therefore, the forward
solver described in this section assumes the use of recessed electrodes.
9.3.1.
(a)
Figure 9.4.
(b)
245
in
9:10
1
r B r r r r; r0 Jr0 dv0
0
Jr rr
Jr r
Jr r
rr0 r; r0 Jr0 dv0
r; r0 Jr0 nr ds0
@
9:11
246
1
r BC r BE r BL r r
r; r0 Jr0 nr ds0 :
0
@
9:12
This means that the current density J within due to BC ; BE and BL is dependent only on the current density or Neumann boundary condition on @.
Therefore, two totally dierent sets of recessed electrodes and lead wires
produce the same current density J in , only if they provide the same
Neumann boundary condition on @. The actual geometrical shape of L
does not aect the computed J, though the shape of C may have some
eect since it can inuence the Neumann boundary condition on @.
Note that the magnetic ux density B in will be dierent depending
on the shape and dimension of recessed electrodes and lead wires. However,
we have
r2 BC r BE r BL r 0
0
for
r2
9:13
We use the nite element method (FEM) to numerically solve (9.9). We rst
construct a 3D model of D and E, assuming that the thickness of each
electrode is negligibly thin. For the discretization of the model into a nite
element mesh, we may use eight-node hexahedral elements with trilinear
interpolation functions i for i 1; . . . ; 8. For the standard hexahedral
element of 1; 13 ,
i
i 1; . . . ; 8
where xi ; yi and zi are the local coordinates of the ith nodal point of the element.
The current density distribution underneath each electrode is not
uniform in most cases. This means that we only know the amount of injection
current I without knowing the Neumann boundary condition g in (9.9).
Therefore, assuming that each electrode is an equipotential surface due to
its high conductivity, we rst solve the following boundary value problem
with mixed boundary conditions:
8
1
>
~
>
rV r 0 in D
r
>
>
>
r
<
9:14
V~ 1 on E 1 and V~ 0 on E 2
>
>
>
1
>
>
: rV~ n 0 on @DnE 1 [ E 2
247
Computation of B and BC
Assuming that J does not change much within each element of the mesh for
, we compute B as
NE
X e
r re
c
Jc
ve
9:15
B r 0
e
4 e 1
jr rc j3
where NE is the number of elements, rce the centre point of the eth element,
e
Jce the current density at re
the volume of the element in the nite
c and v
element mesh of . In order to avoid the singularity where r rce , we
compute B at all nodal points of the mesh. Since we have already computed
J in C from the numerical solution of (9.9) and (9.2), we can calculate BC in
the same way as in (9.15).
9.3.4.2.
Computation of BE
248
(a)
(b)
Figure 9.5. (a) Out-of-plane source and sink currents on the electrode E 1 , and (b) surface
current density within the electrode.
Computation of BL
(a)
Figure 9.6.
(b)
Lead wire geometry. (a) Twisted wires and (b) straight wires.
249
lead wire geometry shown in gure 9.6(b), one might use an analytic solution
for BL .
9.3.5.
9:17
Since J is available from (9.2), we can solve (9.17) for B using FEM if
boundary conditions of B are known on @. We, therefore, compute
B B BC BE BL using the methods described in the previous
section only for r 2 @. Then, we have the Dirichlet boundary condition
on @ and can numerically solve (9.17) for B using FEM. Please note that
it is important to compute all four terms of B on @ to nd the appropriate
Dirichlet boundary condition of B in (9.17). We can also compute (9.17)
in any 3D subdomain of as long as we correctly calculate its boundary
condition.
9.3.6.
250
(a)
(b)
(c)
(d)
(e)
(f )
Lee et al (2003b) used the model in gure 9.7(b) to determine the nite
element mesh with a desirable numerical accuracy. They assumed that the
error in the measured voltage V is larger than 0.1% (Boone et al 1997).
From the sensitivity analysis by Scott et al (1992), the amount of noise in
the measured magnetic ux density B is greater than 0:1 109 Tesla in
most cases. Dividing this by the average value of the computed jBj due to
the injection current of 1 mA, we could get about 1.88% error in the measured
B. Using a mesh with 120 120 120 elements, Lee et al (2003b) showed that
we may obtain less than 0.1% errors in computed V and B. Compromising the
numerical accuracy and computation time, they suggested using a mesh with
80 80 80 elements (total 512 000 elements and 531 441 nodes) for the
domain .
For the homogeneous model with its resistivity of 100 :cm and full-size
recessed electrodes in gure 9.7(c), the computed voltage changes linearly
only along the x-direction, with its values of 28 mV at x 35 mm (on the
left electrode) and 0 V at x 35 mm (on the right electrode). The current
density J in (9.2) can be computed as J 40; 107 ; 108 mA=cm2 , with
a negligibly small error compared with the theoretical value of
J 40; 0; 0 mA=cm2 . For the compatibility test in (9.8), Lee et al (2003b)
251
kJ JB k2
100 [%]
kJk2
kr Jk2 p
100 [%/element]
kJk2
and
"r JB
kr JB k2 p
100 [%/element]
kJB k2
kBz Bm
z k2
100 [%]
kBz k2
where Bz and Bm
z are the computed and measured magnetic ux density,
respectively, they found that "Bz 9:56% for all pixels (or elements) and
"Bz 6:1% excluding the outermost layer of 10 pixels near electrodes.
Comparing the computed and measured magnetic ux density, they observed
mostly random errors and two dierent kinds of systematic error. Random
errors are mainly due to the random noise from the MRI scanner. One of
the systematic errors occurs along the boundary of the cylindrical object.
This is due to the dierence in the resistivity value of the agar object
252
immersed in the saline solution of the phantom, compared with the resistivity
value of the cylindrical object within the model. The other kind of systematic
error occurs near electrodes. This is mainly due to the dierence in lead wire
geometries between the phantom and the model in gure 9.7(e), since it is
dicult to make the lead wires run perfectly straight in real experiments.
To minimize this kind of systematic error, they recommended using a lead
wire guide xed within the MRI scanner. This will be especially important
for image reconstruction algorithms directly using measured B or Bz , without
taking advantage of r2 BL 0 in .
(a)
(b)
Figure 9.8. Typical numerical results for the thorax model in gure 9.7(d) with an injection current of 1 mA. (a) Resistivity distribution of the thorax model. Computed results of
(b) V, (c) Jx , (d) Jy , (e) Jz , (f ) Bx , (g) By and (h) Bz .
(c)
(d)
(e)
Figure 9.8.
(Continued)
253
254
(f)
(g)
(h)
Figure 9.8.
(Continued)
255
(a)
(b)
(c)
Figure 9.9. (a) Measured Bz at z 0 and (b) computed Bz at z 0 from the model in
gure 9.7(e). (c) The dierence between the computed and measured Bz . The amount of
injection current was 28 mA.
256
9.4.
257
Let z be the coordinate that is parallel to the direction of the main magnetic
eld B0 of an MRI scanner. Using a constant current source and a pair of
surface electrodes, we sequentially inject two types of current pulse of I
and I synchronized with the standard spinecho pulse sequence shown in
gure 9.10. The application of the injection current during MR imaging
induces a magnetic ux density B Bx ; By ; Bz . Since the magnetic ux
density B produces inhomogeneity of the main magnetic eld changing B0
to B0 B, it causes phase changes that are proportional to the z-component of B, i.e. Bz . Then, the corresponding MRI signals are
1
SI m; n
Mx; y e jx;y e jBz x;yTc e jxmkx ynky dx dy 9:18
1
and
I
S m; n
9:19
1
Figure 9.10.
258
(a)
(b)
(c)
1
x; y:
2Tc z
9:20
259
Data processing
Due to the main magnetic eld inhomogeneity and the gradient eld nonlinearity, MR images from transversal, sagittal and coronal slices in gure
9.11(a), (b) and (c), respectively, may contain dierent amounts of geometrical distortions. Inhomogeneous susceptibility distribution inside the subject
may also cause geometrical distortions, especially in a high eld MRI
scanner. Since we must compute the internal current density at every pixel
using magnetic ux density images of Bx , By and Bz from the corresponding
imaging slices, we need a geometrical error correction method. This is usually
done by using a grid phantom (Khang et al 2002, Lee et al 2003a).
As described in the next section, MR phase images contain Gaussian
random noise and numerical dierentiations tend to amplify the noise. Therefore, it would be desirable to use a denoising technique before computing (9.21).
Khang et al (2002) suggested a total variation-based denoising technique (Chan
et al 2000). This method eectively removes random noise while preserving
both slow and abrupt changes in magnetic ux density images.
There are two major technical problems in obtaining J. One is the low
signal-to-noise ratio (SNR) in measured magnetic ux density images and
the other is the requirement of subject rotations to obtain B Bx ; By ; Bz .
In the next section, we discuss the SNR problem. The problem of subject
rotations will be treated in section 9.5 on reconstruction algorithms by introducing new techniques based on Bz only.
9.4.6.
1
2Tc SNR
9:22
260
9.5.
261
in :
9:23
The requirement in (9.23) means that the two current densities are not
collinear in . Kim et al (2002) provided a mathematical proof for the 2D
domain, and later Kim et al (2003) proved it for the general 3D domain.
Even with at least two injection currents satisfying (9.23), Kwon et al
(2002a) noted that we can only reconstruct a resistivity image apart from a
multiplicative constant. Therefore, as the second requirement for the uniqueness, they suggested using one boundary voltage measurement to determine
the constant or scaling factor. If we know the true resistivity value at one
point, this scaling factor can also be determined without measuring any
boundary voltage.
In summary, the requirements in data collection methods for the
uniqueness of a reconstructed resistivity image include the following:
1. At least two injection currents satisfying (9.23) almost everywhere in the
imaging slice and the corresponding magnetic ux density images.
2. At least one nonzero boundary voltage measurement or predetermined
resistivity value at least at one point in the imaging slice.
In order to inject two dierent currents, Kwon et al (2002a) suggested using
four electrodes. It is also advantageous to use at least four electrodes since we
can avoid measuring boundary voltage data on current injection electrodes.
262
With four electrodes, we may sequentially inject six dierent currents and
measure the corresponding magnetic ux densities and boundary voltages.
Increasing the amount of measurements beyond the minimal requirement
may be benecial since we can eectively improve the SNR by using all of
them in an appropriate way. Especially, in real experiments, multiple
injection currents with carefully chosen electrode congurations (possibly
more than four electrodes) will be important to minimize the regions
where the induced magnetic ux densities are smaller than a noise level.
Birgul et al (2003) and Oh et al (2003) suggested dierent ways of utilizing
multiple measurements beyond the minimal requirement, and these techniques are described in sections 9.5.4 and 9.5.6.
9.5.2.
Early algorithms
263
J-substitution algorithm
jrV1 rj jrV2 rj
9:24
in
Ji
rVi n gi
jrVi j
Vi r0 0 and Vi r1 Vi
on @;
where V1 and V2 are non-zero voltage dierences between two points r0
and r1 on @. The second coupling identity connecting V1 and V2 stems
from the fact that the change of due to dierent injection currents is
negligible.
The J-substitution algorithm is a natural iterative scheme of the
coupled system (9.24). Since the conductivity should be given by
J1 =jrV1 j J2 =jrV2 j, we can easily design the following iterative
scheme updating V1 ; V2 and .
1. Initial guess 0 1.
2. For each n 0; 1; . . . ; solve
r n rV1n 0 in
n rV1n n g1
on @;
V1n r0 0:
J1 V1n r1
:
jrV1n j V1
4. Solve
(
n 1=2
r n 1=2 rV2
n 1=2
n 1=2
rV2
0 in
n g2
on @;
n 1=2
n 1=2
V2
r0 0:
264
n 1=2
V2
J2
n 1=2
jrV2
j
V2
r1
: jJ r rE rj2 dr
9:25
N
1
X
k0
k
jJ r k E rj2 dr
9:26
where k is the kth element or pixel of the model and k is the conductivity in
k that is assumed to be a constant on each element. P
Note that, in this case,
the conductivity distribution is expressed by r kN01 k k r, where
k r denotes the indicator function of k . In (9.26), E r is also a function
of 0 ; . . . ; N 1 . To update the conductivity from the zero gradient argument for the minimization of the squared residual sum, we dierentiate
(9.26) with respect to k for k 0; . . . ; N 1 to get
@R
2
E r k E r J r dr
@k
k
N
1
X
@E r
2
m
9:27
m E r J r dr:
@k
m 0 m
265
N
1
X
m
m0
@E rk
m E rk J rk
@k
9:28
J rk
E rk
for k 0; . . . ; N 1
9:29
The CCVSR algorithm proposed by Birgul et al (2003) shares the basic idea
with the J-substitution algorithm. They suggested using more than four
electrodes, e.g. 16, uniformly spaced on the boundary of a subject with
its conductivity distribution . Current is injected between a pair of
electrodes facing each other. This current injection method is called the
opposite-drive strategy. Using 16 electrodes, there could be eight injection
currents. For each injection current Ij for j 1; . . . ; 8, they assume that the
corresponding internal current density Jj is available from the measurement
of internal magnetic ux density Bj . They also assume that a set of boundary
voltage data V j@ from electrodes not injecting current are available.
The CCVSR algorithm is iterative, starting with an initial guess n with
n 0 of the true conductivity image . It numerically solves (9.1) for each
injection current Ij with n in place of 1=. From the computed voltage
Vjn , the electric eld intensity is obtained as Enj rVjn . The conductivity
nk of the kth pixel is updated as
P8
n
j 1 Ej r Jj r ds
n1
for k 0; . . . ; N 1:
9:30
k P8 k
n
n
j 1 k Ej Ej ds
The scaling factor is found by minimizing the function
F k1=V n j@ V j@ k with respect to , where V n j@ indicates a set of
computed boundary voltages from (9.1) with n . Then, each kn 1 is multiplied by .
266
9.5.5.
jJrj
jrVrj
for r 2 :
9:31
However, this direct method requires accurate voltage data on the entire
boundary of the subject, which is technically quite dicult in real applications.
Ider et al (2003) also developed a direct method for reconstructing
conductivity images carefully utilizing the vector eld J. They started from
rln J r J in a subject under static conditions. They rewrite
the vector identity as
~j rln r Jj ; j 1; 2; 3
J
~1 0; Jz ; Jy ; J
~2 Jz ; 0; Jx ; J
~3 Jy ; Jx ; 0. We can conwhere J
0
~
struct a characteristic curve with r1 s J1 r1 s passing through a point
r1 s0 at which is assigned. Then,
d
~1 rln r J1
ln r1 s J
ds
and therefore we can determine ln along the characteristic curve r1 s. Next,
~2 r2 s,
we construct a family of characteristic curves r2 s satisfying r02 s J
passing any point on the curve r1 s. We repeat this process until the curves
cover all of and determine in . Here, two injection currents will be
enough to determine in with a known value of at only one point. At
least one boundary voltage measurement can replace the role of the known
conductivity value at one point.
9.5.6.
Harmonic Bz algorithm
All of the previous algorithms utilize the measured data of B and computed
J r B=0 . This means that we must measure all three components of
B Bx ; By ; Bz , and it requires mechanical rotations of the subject within
an MRI scanner as described in section 9.4. In order to eliminate this
impractical subject rotation procedure, Seo et al (2003b) developed a new
algorithm utilizing only one component of the measured magnetic ux
density, such as Bz .
We place a subject inside an MRI scanner and attach surface
electrodes. We assume that the conductivity distribution of the subject is
267
9:34
where
2 @V
6 @y
6 .
.
U6
6 .
4 @V
@y
@V1
@x
..
.
7
7
7;
7
@VN 5
@x
2 @ 3
6 @x 7
7
s6
4 @ 5;
@y
3
r2 B1z
1 6 . 7
b 4 .. 5:
0
r2 BN
z
For the case where two injection currents are used (N 2), we can
obtain s provided that two voltages V1 and V2 corresponding to two injection
currents I1 and I2 satisfy
6 0:
@y @x
@x @y
9:35
268
We can argue that (9.35) holds for almost all positions within the subject, since
two current densities J1 and J2 due to appropriately chosen I1 and I2 will not
have the same direction (Kim et al 2002, Ider et al 2003, Kim et al 2003).
We use N injection currents to better handle measurement noise in Bz
and improve the condition number of UT U, where UT is the transpose of
U. Using the weighted regularized least square method suggested by Oh
et al (2003), we can get s as
~ TU
~ I1 U
~ T ~b
s U
9:36
SNRj
N
X
9:37
SNRj
j1
where SNRj is the signal-to-noise ratio of the measured Bjz . Note that SNRj
should be determined for each position or pixel. In practice, however, it is
dicult to know SNRj for each position. Oh et al (2003) discuss how to
estimate SNRj from measured Bjz data. Computing (9.36) for each position
or pixel, we obtain a distribution of
@ @ T
s
@x @y
inside the subject.
We now tentatively assume that the imaging slice S is lying in the plane
fz 0g and the conductivity value at a xed position r0 x0 ; y0 ; 0 on its
boundary @S is 1. For a moment, we denote r x; y, r0 x 0 ; y0 and
x; y; 0 r. In order to compute from r @=@x; @=@y, Seo
et al (2003b) suggested a method using line integrals. However, since the
line integral technique tends to accumulate errors, it is not suitable for
noisy Bz data. Oh et al (2003), therefore, employed a layer potential technique in two dimension. Then,
r r2 r r0 r0 dr0
S
@S
9:38
269
where
r r0 1=2 log jr r0 j and rr0 r r0 1=2r r0 =jr r0 j2 :
It is well known (Folland 1976) that for r 2 @S,
lim
nr0 rr0 r tnr r0 r0 dlr0
t ! 0 @S
r
@S
@S r 1
r r0 nr0
1
r r0 rr0 0
0
0
r
dl
dr
@S
r
2
2 @S jr r0 j2
2 S
jr r0 j2
9:39
where @S denotes the conductivity restricted at the boundary @S. It is also
well known that the solvability of the integral equation (9.39) for @S is
guaranteed for a given right-hand side of (9.39) (Folland 1976). Since r
is known in S, so does the right-hand side of (9.39). This enables us to
obtain the value @S by solving the integral equation (9.39). Now, we can
compute the conductivity in S by substituting the boundary conductivity
@S into (9.38) as
0
0
0
r rr0 r r rr dr
nr0 rr0 r r0 @S r0 dlr0 :
S
@S
9:40
The process of solving (9.36) for each pixel and (9.39) and (9.40) for each
imaging slice can be repeated for all imaging slices of interest within the
subject, as long as the measured data Bz are available for the slices. Furthermore, we can apply the method described in this section to any imaging slice
of axial, coronal and sagittal direction.
As expressed in (9.32), voltages Vj depend on the unknown true
conductivity and, therefore, we do not know the matrix U corresponding
to . This requires us to use the iterative algorithm described below. For
j 1; . . . ; N, we sequentially inject current Ij through a chosen pair of electrodes and measure the z-component of the induced magnetic ux density
Bjz . For each injection current Ij , we also measure boundary voltages Vj j@S
on electrodes not injecting the current Ij . Then, the r2 Bz algorithm is as
follows:
1. Let n 0 and assume an initial conductivity distribution 0 .
2. Compute Vjn by solving the following Neumann boundary value
problems for j 1; . . . ; N:
r n rVjn 0 in
9:41
n rVjn n gj on @:
270
Partial Bz algorithm
in
on @:
9:42
9:43
0
r r0
J
B r
Jr0
dr0 :
4
jr r0 j3
Similarly, the injection current I along lead wires produces a magnetic ux
density BI . Hence, the total magnetic ux density due to the internal current
density J and external current I is B BJ BI : Using an MRI scanner in
which the subject is located, we can measure the z-component Bz of B.
Let Dt denote a cut of the subject
by the xy-plane fz tg. A thin slice
S
s to be imaged could be s < t < Dt for a small > 0. Since a human
271
body is locally cylindrical in its shape, Dt D0 for < t < and therefore
s D0 ; . If the conductivity of the subject does not change much
in the z-direction, we could produce approximately a transversal internal
current density J, i.e. J Jx ; Jy ; 0 in the cylindrical chop s using
longitudinal electrodes. Note that J could have nonzero z-components in
the exterior ns of the thin chop s . The transversal current density
J Jx ; Jy ; 0 in s satises the following mixed boundary value problem:
8
@
@
>
>
J J 0 in s
>
> @x x @y y
>
<
9:44
J n g on @D0 ;
>
>
>
>
>
: @ J 0 @ J on D [ D :
@z x
@z y
Here, D and D indicate the top and bottom surface of s , respectively. We
assume that the current density g under the electrodes is independent of z
along the lateral boundary @D0 ; of s .
From the BiotSavart law,
Bz r BJz r BIz r
y y0 Jx r0 x x0 Jy r0 0
0
dr BIz r;
4
jr r0 j3
r 2 s
9:45
y y0 Jx r0 x x0 Jy r0 0
0
dr
Bz r
4 s
jr r0 j3
Gr BIz r;
r 2 s :
9:46
y y0 Jx r0 x x0 Jy r0 0
dr ; r 2 s :
Gr : 0
4 ns
jr r0 j3
Since the lead wires are located outside of , we have r2 BIz 0 in s .
Similarly, G also satises r2 G 0 in s . These can be proved using
r2 1=jr r0 j 0 when r 6 r0 .
Since r Jy ; Jx ; 0 0 in s , there is a function w in s such that
rwr Jy r; Jx r; 0
in s :
272
r r0 rwr0 0
Bz r 0
dr Gr BIz r;
4 s
jr r0 j3
r 2 s :
9:47
0
4
r r0
r0
wr0 dSr0
@s
jr r0 j3
Gr BIz r;
r 2 s :
9:48
Since the integral in the right hand side of (9.48) satises the Laplace equation, we obtain
r2 Bz r 0 r2 wr;
r 2 s :
Now we dene H as
Hr : wr
1
B r;
0 z
r 2 s :
x; y 2 D0 :
9:49
wr wr0
g dl; r 2 @D0
Cr
1
B
0 z
g dl;
Cx;y;0
in @D0 ;
r x; y; z 2 @D0 ; :
9:50
9:51
273
@
1 @
H
B;
@y
0 @y z
Jy
@
1 @
H
B:
@x
0 @x z
9:52
Other algorithms
274
9.6.
All MREIT images published until now were obtained from computer simulations or saline phantom experiments. This section presents some of these
results.
9.6.1. Images using the J-substitution algorithm
Lee et al (2003a) reconstructed current density and resistivity images of a
saline phantom using a 0.3 Tesla experimental MRI scanner and the J-substitution algorithm. The main magnet of the MRI scanner was a permanent
magnet with gap size of 500 mm and the main magnetic eld pointed to
the z-direction. They used a cubic phantom (50 mm 50 mm 50 mm,
acrylic plastic) shown in gure 9.12(a). It was lled with a solution containing
12.5 g/l NaCl and 2 g/l CuSO4 :5H2 O. Inside the phantom, a cylindrical
sausage object is located around its centre. The diameter and height of the
sausage object were 30 and 50 mm, respectively. The resistivity values of
the solution and sausage were 50.5 and 123.7 :cm, respectively.
They installed four copper electrodes (5 mm 50 mm) to inject two
dierent currents I1 and I2 . Using a constant current source, the injection
current I1 was applied between the vertical pair of electrodes. After collecting
all image data for I1 , they switched it to the other electrode pair in the
horizontal direction for the injection current I2 . Injection current pulses of
I 28 mA with Tc =2 24 ms were synchronized with the pulse sequence
in gure 9.10. The pulse repetition time was 300 ms and the echo time was
60 ms. The slice thickness was 10 mm and the eld of view was 77 mm. In
obtaining 128 128 MR images, the number of averaging was 16 and the
phase encoding step was 128. The voxel size was 0:6 mm 0:6 mm 10 mm.
They rotated the phantom twice, as shown in gure 9.11, to obtain phase
images for Bx ; By and Bz from three slices of Su ; Sc and Sl , shown in
gure 9.12(b). Since Bz should be dierentiated with respect to x and y in
(a)
(b)
(c)
Figure 9.12. (a) Saline phantom including a cylindrical sausage object, (b) three imaging
slices of Su , Sc and Sl , and (c) MR magnitude image at the centre slice Sc .
MREIT images
(a)
275
(b)
Figure 9.13. Phase image for Bz at the centre slice Sc of the phantom for the vertical injection current I1 : (a) before and (b) after phase unwrapping.
computing the current density J r B=0 , they acquired one phase image
for Bz from the centre slice Sc . We must dierentiate Bx and By with respect
to z, as well as y and x, respectively. Therefore, they obtained three phase
images from three slices of Su ; Sc and Sl for each of Bx and By . For each
injection current of I1 and I2 , they acquired seven phase images from the
three slices.
Figure 9.12(c) shows the MR magnitude image of the phantom at the
centre slice Sc . The artefacts near electrodes are due to the RF shielding
eect of copper electrodes. The SNR of the magnitude image was 27.2 in
the solution and 6.86 in the sausage, assuming that both solution and sausage
are homogeneous. As shown in gure 9.12(c), the phantom occupies a region
of 83 83 pixels in the 128 128 MR image. Since there are artefacts near
electrodes, they extracted magnetic ux density images of 66 66 pixels
from the region of 83 83 pixels. They applied the total variation-based
denoising method by Chan et al (2000) to images of Bx ; By and Bz .
Figure 9.13 shows the wrapped and unwrapped phase image for Bz at
the centre slice Sc . Figure 9.14(a), (b) and (c) are images of Bx , By and Bz ,
respectively, at the same slice of Sc before denoising for the vertical injection
current I1 . Figure 9.14(d), (e) and (f ) are the corresponding images after
denoising. The noise standard deviations in the magnetic ux density
image were estimated using (9.22) as B 1:43 109 Tesla in the solution
and 5:68 109 Tesla in the sausage. Figure 9.15(a) shows horizontal
proles at the centre of two Bz images in gure 9.14(c) and (f ). Figure
9.15(b) is the dierence between these two horizontal proles.
Figure 9.16(a) shows the magnitude of the current density jJj for the
vertical injection current I1 , computed from the nite element model of the
saline phantom with the true resistivity distribution using the 3D forward
276
(a)
(b)
(c)
(d)
(e)
(f )
Figure 9.14. Magnetic ux density images at the centre slice Sc for the vertical injection
current I1 : (a) Bx , (b) By , (c) Bz , (d) Bx after denoising, (e) By after denoising, and (f ) Bz
after denoising.
MREIT images
277
(a)
(b)
Figure 9.15. (a) Horizontal proles at the centre of two Bz images in gure 9.14(c) and (f ).
(b) Dierence between two proles in (a).
278
(a)
(b)
(c)
Figure 9.16. Images of the magnitude of the current density jJj for the vertical injection
current I1 from (a) the 3D forward solver, (b) measured magnetic ux densities without
denoising, and (c) with denoising.
MREIT images
279
(a)
(b)
Figure 9.17. (a) Horizontal proles at the centre of two jJj images in gure 9.16(b) and (c).
(b) Dierence between two proles in (a).
280
Figure 9.18(a) shows the true resistivity image of the phantom. Here, the
resistivity distribution within the solution and sausage are assumed to be
homogeneous in each region. Using the J-substitution algorithm, gure
9.18(b) and (c) show reconstructed resistivity images without denoising
and with denoising, respectively. Figure 9.19 shows horizontal proles
around the centre of three resistivity images in gure 9.18. For the resistivity
image in gure 9.18(b) without denoising, the reconstructed average
resistivity values were 60.8 and 115.4 :cm in the solution and sausage,
respectively, compared to the true values of 50.5 and 123.7 :cm. For the
image in gure 9.18(c) with denoising, the average values were 60.9 and
117.7 :cm in the solution and sausage, respectively. The relative L2 -error
of the resistivity image is dened as
"
k k2
100 [%]
k k2
where and are the true and reconstructed resistivity image, respectively.
The computed relative L2 -errors were 32.3 and 25.5% for the images in gure
9.18(b) and (c), respectively.
Lee et al (2003a) discussed that the errors in reconstructed current
density and resistivity images were primarily due to the low SNR of the
0.3 Tesla experimental MRI scanner. Since they rotated the phantom to
get images of Bx and By in addition to Bz , misalignments of pixels among
dierent slices should have also caused a signicant amount of errors.
Their results suggest that we should use only one component of B such as
Bz to eliminate the troublesome subject rotation procedure. Furthermore,
recessed electrodes are desirable to avoid severe artefacts near copper
electrodes.
9.6.2. Images using the harmonic Bz algorithm
This section summarizes experimental results using the harmonic Bz
algorithm by Oh et al (2003). They used the same 0.3 Tesla experimental
MRI scanner described in the previous section. They constructed a cubic
saline phantom of 50 mm 50 mm 50 mm shown in gure 9.20(a). On
the four sides of the phantom, recessed electrode assemblies of
20 mm 20 mm 10 mm were positioned symmetrically. The phantom
was lled with a solution of 2 S/m conductivity (12.5 g/l NaCl and 2 g/l
CuSO4 ). Two cylindrical objects (5 ml of 20% polyacrylamide, 4.9 ml of
6% sodium-styrenesulfonate with molecular weight of 70 000, 15 mg of
CuSO4 , 0.1 ml of 10% ammonium persulfate and 4 ml of tetramethylethylenediamine) were located inside the phantom, and these objects are denoted
as A1 and A2 in gure 9.20. The conductivity value of A1 and A2 was
0.56 S/m. Their diameter was 14 mm and heights were 20 and 50 mm, respectively. Figure 9.20(b) and (c) show the diagrams of the phantom.
MREIT images
281
(a)
(b)
(c)
Figure 9.18. (a) True resistivity image assuming the sausage is homogeneous, (b) reconstructed resistivity image without denoising, and (c) with denoising.
282
Figure 9.19. Horizontal proles around the centre of three resistivity images in gure
9.18.
MREIT images
283
(a)
(b)
(c)
Figure 9.20. (a) Cubic saline phantom with four recessed electrodes. Diagrams of the
phantom: (b) top view and (c) front view (the recessed electrode on the frontal surface is
hidden). The conductivity values of the solutions A1 and A2 were 2, 0.56 and 0.56 S/m,
respectively.
Figure 9.21.
284
(a)
(b)
(c)
Figure 9.22. (a) MR magnitude image of the phantom with four recessed electrodes at
the axial imaging slice of S 9 (25 z 28:1 mm). Since the imaging slice is above the
object A1, as shown in gure 9.20(c), we can see only the object A2. (b) 82 82 image
of B1z at S 12 . (c) 82 82 image of B1z at S 6 .
MREIT images
(a)
(b)
(c)
(d)
(e)
(f )
285
Figure 9.23. (a) Positions of ve slices. Reconstructed conductivity images of the saline
phantom at slices of (b) #1, (c) #2, (d) #3, (e) #4 and (f ) #5. The relative L2 -errors are
in the range of 13.8 to 21.5%.
286
Figure 9.24. Typical horizontal proles of the conductivity images in gure 9.23. Solid
and dotted lines are the true and reconstructed proles, respectively.
MREIT images
(a)
287
(b)
Figure 9.25. Typical reconstructed images of the magnitude of current density distributions. (a) Imaging slice S 9 including only the object A1, and (b) dierent slice S 7 including
both A1 and A2.
288
Figure 9.26.
(a)
(b)
Figure 9.27. (a) MR magnitude image of the tissue phantom in gure 9.26 at the middle
imaging slice, and (b) reconstructed conductivity image at the same slice.
9.7.
289
9.8.
290
The rst problem has been the major technical limitation of MREIT and also
MRCDI. Now, the harmonic Bz algorithm provides a solution even though
this algorithm has a weak point in terms of noise tolerance. As we make
progress in better understanding the information embedded in the induced
magnetic ux density, we expect other algorithms with an improved stability
against measurement noise to appear soon. The second problem of artefacts
near electrodes can be eectively handled by using recessed electrodes. We
may also look for new electrode materials generating a negligible amount
of artefacts.
The third and fourth problems are interrelated. If the SNR of an MRI
scanner is large enough, we could easily reduce the amount of injection
current down to 0.1 or 1 mA. There are many factors determining the
amount of random noise in measured magnetic ux density images. First
of all, we should use an MRI scanner with a high main magnetic eld
and excellent eld homogeneity. Then, we may gradually increase the
voxel size until we obtain the amount of random noise that could be
tolerated by image reconstruction algorithms. Ecient denoising techniques based on the underlying physical principles should be developed
to enhance the SNR without sacricing the edge information in reconstructed images.
Reducing the amount of injection currents down to 0.1 or 1 mA is the
most challenging task in MREIT. With such a small injection current, the
induced magnetic ux density could easily be lower than the noise level.
Once we have minimized the amount of random noise in measured magnetic
ux density images, we have to rely on the signal averaging technique to
improve the SNR further. However, this will increase the imaging time
and may limit the practical applicability of MREIT.
From the BiotSavart law, the induced magnetic ux density (signal) is
determined by the current density distribution due to an injection current.
Though the relation between them is given in the form of a 3D convolution,
we can roughly expect a bigger signal where the current density is large.
Therefore, we should further investigate the optimal electrode conguration
including size, shape and location to minimize the regions where the current
density becomes small. It is desirable to sequentially inject multiple currents
through dierent pairs of electrodes so that each injection current will
produce bigger signals in dierent regions. Then, we could get an averaging
eect by using all of them in an appropriate way. Injecting a pattern of
currents with multiple current sources may also be helpful to generate
more or less uniform current density distribution. When multiple electrodes
are used, it will be benecial to measure all independent boundary voltage
data to provide extra compatibility conditions.
In terms of image reconstruction algorithms, a hybrid form combining
the advantages of dierent algorithms may turn out to be optimal as long
as it requires only one component of B such as Bz . Since conventional MR
References
291
REFERENCES
Beravs K, White D, Sersa I and Demsar F 1997 Electric current density imaging of bone by
MRI Magn. Reson. Imag. 15 90915
Beravs K, Frangez R, Gerkis A N and Demsar F 1999a Radiofrequency current density
imaging of kainate-evoked depolarization Magn. Reson. Imag. 42 13640
Beravs K, Demsar A and Demsar F 1999b Magnetic resonance current density imaging of
chemical processes and reactions J. Magn. Reson. 137 2537
Birgul O, Ozbekl O, Eyuboglu B M and Ider Y Z 2001 Magnetic resonance conductivity
imaging using 0.15 tesla MRI scanner Proc. 23rd. Ann. Int. Conf. IEEE Eng. Med.
Biol. Soc.
Birgul O, Eyuboglu B M and Ider Y Z 2003 Current constrained voltage scaled reconstruction (CCVSR) algorithm for MR-EIT and its performance with dierent probing
current patterns Phys. Med. Biol. 48 65371
Bodurka J, Jesmanowicz A, Hyde J S, Xu H, Estkowski L and Li S J 1999 Current-induced
magnetic resonance phase imaging J. Magn. Reson. 137 26571
Bodurka J and Bandettini P A 2002 Toward direct mapping of neural activity:
MRI detection of ultraweak, transient magnetic eld changes Magn. Reson. Med.
47 10528
292
Boone K, Barber D and Brown B 1997 Imaging with electricity: report of the European
concerted action on impedance tomography J. Med. Eng. Tech. 21 20132
Burnett D S 1987 Finite Element Analysis (Reading, MA: Addison-Wesley)
Carter M A 1995 RF Current Density Imaging with a Clinical Magnetic Resonance Imager
MS Thesis, University of Toronto, Canada
Chan T, Marquina A and Mulet P 2000 High-order total variation-based image restoration SIAM J. Sci. Comput. 22 50316
Eyuboglu M, Reddy R and Leigh J S 1998 Imaging electrical current density using nuclear
magnetic resonance Elektrik 6 20114
Eyuboglu M, Birgul O and Ider Y Z 2001 A dual modality system for high resolutiontrue conductivity imaging Proc. XI Int. Conf. Elec. Bioimpedance (ICEBI) 40913
Folland G 1976 Introduction to Partial Dierential Equations (Princeton, NJ, USA: Princeton University Press)
Gamba H R and Delpy D T 1998 Measurement of electrical current density distribution
within the tissues of the head by magnetic resonance imaging Med. Biol. Eng.
Comp. 36 16570
Gamba H R, Bayford D and Holder D 1999 Measurement of electrical current density
distribution in a simple head phantom with magnetic resonance imaging Phys.
Med. Biol. 44 28191
Gerkis A N 1996 An Enhanced RF Current Density Imaging Technique for Imaging Biological Media MS Thesis, University of Toronto, Canada
Ghiglia D C and Pritt M D 1998 Two-Dimensional Phase Unwrapping: Theory, Algorithms
and Software (New York: Wiley Interscience)
Ider Y Z and Birgul O 1998 Use of the magnetic eld generated by the internal distribution
of injected currents for Electrical Impedance Tomography (MR-EIT) Elektrik 6 215
25
Ider Y Z, Onart S and Lionheart W R B 2003 Uniqueness and reconstruction in magnetic
resonance.electrical impedance tomography (MR.EIT) Physiol. Meas. 24 591604
Joy M L G, Scott G C and Henkelman R M 1989 In vivo detection of applied electric
currents by magnetic resonance imaging Magn. Reson. Imag. 7 8994
Joy M L G, Lebedev V P and Gati J S 1999 Imaging of current density and current pathways in rabbit brain during transcranial electrostimulation IEEE Trans. Biomed. Eng.
46 113949
Khang H S, Lee B I, Oh S H, Woo E J, Lee S Y, Cho M H, Kwon O, Yoon J R and Seo J K
2002 J-substitution algorithm in magnetic resonance electrical impedance tomography (MREIT): phantom experiments for static resistivity images IEEE Trans. Med.
Imag. 21 695702
Kim S W, Kwon O, Seo J K and Yoon J R 2002 On a nonlinear partial dierential equation arising in magnetic resonance electrical impedance tomography SIAM J. Math
Anal. 34 51126
Kim Y J, Kwon O, Seo J K and Woo E J 2003 Uniqueness and convergence of conductivity
image reconstruction in magnetic resonance electrical impedance tomography Inv.
Prob. 19 121325
Kwon O, Woo E J, Yoon J R and Seo J K 2002a Magnetic resonance electrical impedance
tomography (MREIT): simulation study of J-substitution algorithm IEEE Trans.
Biomed. Eng. 48 1607
Kwon O, Lee J Y and Yoon J R 2002b Equipotential line method for magnetic resonance
electrical impedance tomography (MREIT) Inv. Prob. 18 10891100
References
293
Lee B I, Oh S H, Woo E J, Lee S Y, Cho M H, Kwon O, Seo J K and Baek W S 2003a Static
resistivity image of a cubic saline phantom in magnetic resonance electrical impedance tomography (MREIT) Physiol. Meas. 24 57989
Lee B I, Oh S H, Woo E J, Lee S Y, Cho M H, Kwon O, Seo J K, Lee J Y and Baek W S
2003b Three-dimensional forward solver and its performance analysis in magnetic
resonance electrical impedance tomography (MREIT) using recessed electrodes
Phys. Med. Biol. 48 197186
Mikac U, Demsar F, Beravs K and Sersa I 2001 Magnetic resonance imaging of alternating
electric currents Magn. Reson. Imag. 19 84556
Oh S H, Lee B I, Woo E J, Lee S Y, Cho M H, Kwon O and Seo J K 2003 Conductivity and
current density image reconstruction using harmonic Bz algorithm in magnetic resonance electrical impedance tomography Phys. Med. Biol. 48 310116
Oh S H, Lee B I, Lee S Y, Woo E J, Cho M H, Kwon O and Seo J K 2004 Magnetic resonance electrical impedance tomography: phantom experiments using a 3.0 Tesla MRI
system Mag. Reson. Med. in press
Park C, Park E J, Woo E J, Kwon O and Seo J K 2004a Static conductivity imaging using
variational gradient Bz algorithm in magnetic resonance electrical impedance tomography Physiol. Meas. 25 27569
Park C, Kwon O, Woo E J and Seo J K 2004b Electrical conductivity imaging using gradient Bz decomposition algorithm in magnetic resonance electrical impedance tomography (MREIT) IEEE Trans. Med. Imag. 23 38894
Saulnier G J, Blue R S, Newell J C, Isaacson D and Edic P M 2001 Electrical impedance
tomography IEEE Sig. Proc. Mag. 18 3143
Scott G C, Joy M L G, Armstrong R L and Henkelman R M 1991 Measurement
of nonuniform current density by magnetic resonance IEEE Trans. Med. Imag. 10
36274
Scott G C, Joy M L G, Armstrong R L and Hankelman R M 1992 Sensitivity of magnetic
resonance current density imaging J. Magn. Reson. 97 235254
Scott G C 1993 NMR Imaging of Current Density and Magnetic Fields PhD Thesis, University of Toronto, Canada
Seo J K, Kwon O, Lee B I and Woo E J 2003a Reconstruction of current density distributions in axially symmetric cylindrical sections using one component of magnetic ux
density: computer simulation study Physiol. Meas. 24 56577
Seo J K, Yoon J R, Woo E J and Kwon O 2003b Reconstruction of conductivity and
current density images using only one component of magnetic eld measurements
IEEE Trans. Biomed. Eng. 50 11214
Sersa I, Beravs K, Dodd N J F, Zhao S, Miklavcic D and Demsar F 1997 Electric current
density imaging of mice tumors Magn. Reson. Med. 37 4049
Webster J G ed. 1990 Electrical Impedance Tomography (Bristol, UK: Adam Hilger)
Weinroth A P 1998 Variable Frequency Current Density Imaging MS Thesis, University of
Toronto, Canada
Woo E J, Lee S Y and Mun C W 1994 Impedance tomography using internal current
density distribution measured by nuclear magnetic resonance SPIE 2299 37785
Woo E J, Lee S Y, Seo J K, Kwon O, Oh S H and Lee B I 2004 Conductivity images of
biological tissue phantoms using a 3.0 Tesla MREIT system 26th Ann. Int. Conf.
IEEE EMBS in press
Yan R T H 1997 Fast Radio-Frequency Current Density Imaging with Spiral Acquisition MS
Thesis, University of Toronto, Canada
294
Yoon R 2000 Characterization of Cortical Spreading Depression in Rat using RadioFrequency Current Density Imaging MS Thesis, University of Toronto, Canada
Zhang N 1992 Electrical Impedance Tomography based on Current Density Imaging MS
Thesis, Dept. of Elec. Eng., Univ. of Toronto, Toronto, Canada
Chapter 10
Electrical tomography for industrial
applications
Trevor York
10.1.
INTRODUCTION
The mathematical concept of tomography was rst suggested early in the 19th
century. About 100 years later an Austrian mathematician, Radon, extended
the ideas to objects with arbitrary shapes [1]. During the rst half of the 20th
century several independent workers, notably Bocage, Ziedses des Plantes,
Grossman and Watson, suggested methods for imaging a plane using x-rays.
In 1979 Godfrey Hounseld and Allen Cormack were jointly awarded the
Nobel prize for their pioneering work on computed x-ray tomography, a
concept that was, perhaps, anticipated by Gabriel Frank in 1940 [2]. The
basic aim of modern tomography is to determine the distribution of materials
in some region of interest by obtaining a set of measurements using sensors
that are distributed around the periphery. For instance, in medical applications the contrasting materials may be normal and cancerous tissue, and
for industrial applications the materials could be oil or gas in a pipeline. Tomographic measurements are non-intrusive, perhaps penetrating the wall of the
vessel but not entering into the medium, and also, ideally, non-invasive such
that the sensors are located on the outside of the wall. Each measurement
is aected, to a greater or lesser degree, by the location of materials in the
region of interest. Typically a source of energy is imposed on the vessel from
one orientation and a number of measurements are taken by distributed
sensors to create a projection of data. The source is then moved to provide
another projection and so on around the vessel until a frame of data is accumulated. Usually the frame of data is translated, in software, into a crosssectional image representing the distribution of materials. Tomography has
enjoyed considerable success in medical applications, for instance identication of tumours, particularly using x-rays as a source of energy, to identify
contrasting material density from the attenuation of the transmitted signal.
296
More recently magnetic resonance and electrical excitation [3], among others,
have emerged as alternative modalities oering particular features that might
be usefully exploited. Tomography, therefore, is inherently complex, involving
energization of a target region, multiple sensor electronics, data acquisition
and data inversion.
Fuelled by developments in personal computing and sensor design,
research into applications of tomography to industrial processes began to
gain popularity in the early 1990s. Techniques have been inuenced by
successes in medicine; however, in many cases, the demands of industrial
applications are signicantly dierent. It is not uncommon to require
many cross-sectional images per second, at low cost, using mobile equipment that is easy to operate and introduces no risk to the user. For these
reasons nucleonic techniques are often inappropriate and alternatives have
emerged. For instance, the literature includes descriptions of instruments
that are based on acoustic propagation, optical, infra-red and microwave
sources of energy [4, 5]. A particularly successful approach for industrial
applications involves electrical tomography. Three, relatively low frequency,
measurement modalities are used to determine distributions of conductivity
(resistance), permittivity (capacitance) and permeability (inductance), and
these are the subject of the present survey. Impedance tomography oers
the ability to measure both the resistive and reactive components. It
should be noted that microwave tomography is excluded from the present
discussion, operating at signicantly higher excitation frequencies, of the
order of GHz, where eects due to molecular structure start to become
signicant. The characteristics of the electrical modalities are summarized
in table 10.1.
Prediction of the electric elds that arise, and consequently the boundary
values, due to electrical excitation of specic distributions of materials, is
referred to as the forward problem. This is usually realized using nite element
modelling tools. The opposite process, to determine the distribution of
materials from the boundary values, is called the inverse problem. For x-ray
tomography the path of the signal is known to follow a straight line and the
only eect on the detected signal strength is due to material along that path.
This is a so-called hard eld problem. In contrast, for soft eld modalities
such as electrical tomography, material throughout the subject aects the
signal strength and presents a much more demanding challenge. Consequently
it is not yet possible to match the spatial resolution of the images that are
produced by hard-eld systems, although this is also, in part, due to the
increased number of measurements that are often taken in hard-eld systems.
An important decision when selecting an appropriate modality is whether the
reduced resolution is an acceptable price to pay in order to enjoy the accompanying benets.
Electrical tomography has motivated applications for process design
and validation, on-line monitoring and control. This can, for instance, lead
Introduction
Table 10.1.
Method
297
ECT
Measurand
Material properties
Typical material
Capacitance
C
Oil
De-ionized water
Non-metallic powders
Polymers
Burning gases
Resistance
(impedance)
RZ
Water/saline
Biological tissue
Geological materials
Semiconductors
Self/mutual
inductance
L=M
Metals
Some minerals
Magnetic materials
Ionized water
Capacitive
plates
ERT
Electrode
array
EMT
Coil array
298
of signicant sections. The intention is not to attempt to present an exhaustive introduction to industrial tomography, which can be readily found in
earlier publications such as Process Tomography: Principles, Techniques
and Applications by Williams and Beck [17], but to present a technical
audit for 2004.
Section 10.2 presents a summary of hardware for data acquisition that
has been reported for electrical tomography. The hardware is primarily
sourced from academic institutions, but includes two established commercial
instruments plus emerging systems. Section 10.3 addresses data processing
issues of image reconstruction and interpretation. Section 10.4 considers
contrasting applications of tomography that have made signicant progress
towards industrial benet.
10.2.
DATA ACQUISITION
Figure 10.1.
Data acquisition
299
300
[23], while others combine the functionality. The most popular approach for
industrial applications is to apply a sinusoidal current source to a pair of
electrodes, at a frequency of some tens of kilohertz, and to measure the
resulting electric potentials between other pairs of electrodes. This arrangement reduces eects due to contact impedance, although this is less important in many industrial applications compared with the medical eld in
which the interface is human tissue. This adjacent strategy provides high
sensitivity near the vessel walls, but is poor in the centre of the region.
Alternatively, other strategies can be adopted, for instance to inject current
between opposite electrodes. An adaptive current strategy, in which signals
of varying amplitude are injected concurrently on all electrodes in order to
optimize the eld distribution, is popular in the medical tomography community [23]. Measurements are taken concurrently on all electrodes and the need
for multiplexing the electrodes is removed. The required instrumentation is
considerably more complex for this approach and the resulting benets
have not yet proved suciently attractive to generate widespread interest
for industrial applications. Much eort is directed at providing a high quality
current source with high output impedance. However, a practical solution,
that has some merit, monitors a modest current source [24]. For industrial
applications metal walls pose a signicant problem as current leaks away
through the wall. A strategy to accommodate this uses common ground
return for transmitted and detected signals. An ERT system that resulted
from work done at UMIST has been developed into a commercial instrument
by Industrial Tomography Systems Ltd. (http://www.itoms.com).
Three recent projects have explored the design of ERT instruments that
specically aim to yield low-cost solutions [79, 80, 83]. The rst two use a bidirectional current pulse to excite the region, and this is related to the original
technique that was used for electrical capacitance tomographyas described
in section 10.2.2. Dierential voltages are measured around the vessel on the
positive and negative cycles. These values are subtracted to yield d.c. levels
representing resistance. Electrochemical eects are minimized by the use of
bipolar excitation. At the University of Cape Town [79] a commercial
DAQ card is used to transfer results into the host PC. The original version
employed a single multiplexed measurement channel and was tested at low
excitation frequencies of a few kilohertz. A modied version takes advantage
of parallel input ampliers and is synchronized by an embedded microcontroller. The authors claim a measurement rate of 500 frames per second.
Image reconstruction is performed o-line using the NewtonRaphson algorithm. The system that has been developed at the University of Aberdeen [80]
is intended for considering uid distribution in porous rock. It employs eight
planes of 24 electrodes and can acquire a frame of data, comprising
192 192 measurements, in 19 s. It is suggested that the system might oer
capture rates of a few hundred frames/s for a 16-electrode sensor. At
Tampere University of Technology [83], a 16-electrode ERT system is
Data acquisition
301
(a)
(b)
(c)
Figure 10.2. Prototype conducting ring sensor. (a) Complete sensor, (b) inner view of the
conductive ceramic ring, (c) electrical contacts on the outside wall of the conducting ring [82].
described for monitoring the air bubbles in pulp ow. The system that can
inject either sinusoidal waves or square pulses with some advantage
suggested the latter in terms of a sampling period.
A novel approach to the implementation of ERT sensors is described by
Wang et al [82]. Conventional ERT sensors use discrete electrodes that are
mounted on the inside wall of the vessel, and this can give problems when
the medium is discontinuous. For instance, consider a conducting aqeous
medium that is either stratied or contains gas bubbles. If a large gas
bubble is adjacent to a pair of electrodes then there is, essentially, no conduction between them. Wang et al have proposed a novel sensor in which the
discrete electrodes are replaced by a conductive ring that is inserted into
the wall. Contact can be made at any point and discontinuities are accommodated by the ring such that current can still be applied. This arrangement has
been modelled using 3D FEM, and results suggest a more uniform eld
within the vessel but with reduced eld strength and consequently sensitivity.
A value of 5 : 1 is suggested for the ratio of the conductivity of the ring
compared with the material in the vessel. A prototype sensor comprising a
38 mm ring with 16 electrical contacts has been manufactured from conducting ceramic having conductivity of 0.5 ms cm1 , as shown in gure 10.2.
Initial results of images of stratied ow in water are shown in gure 10.3.
302
Figure 10.3.
10.2.2.
Introductions to ECT have been presented previously by, for instance, Yang
[25]. Distributions of electrical permittivity are determined from measurements of current around the boundary of a vessel. For capacitance measurements, electrodes must have a large surface area in order to provide sucient
signal. Electrodes are often located outside the vessel, such that the technique
is non-invasive as well as non-intrusive. In contrast to ERT, an a.c. voltage
signal is usually applied to a drive electrode and the resulting current on the
remaining electrodes is measured. Typical excitation frequencies, to provide
sucient sensitivity, are about 1 MHz. The main dierence between the
various systems that have been described is the use of either sinusoidal or
Data acquisition
303
304
Data acquisition
Figure 10.4.
305
306
10.2.5.
Data processing
307
switch room, which is a safe area some 50 m from the lter. The philosophy
behind the design of the I.S. system has been to utilize, wherever possible,
existing certied components. This is achieved by taking an existing system
certication for a typical Zener barrier in a strain gauge conguration
and expanding on this using a series of certied I.S. relay modules.
The intrinsically safe EIT system is built on an earlier system that incorporates a commercial LCR instrument with a custom switch matrix [28].
Although the acquisition rate is slow, taking about 40 s for a 16 electrode
frame of ERT data, it is adequate for many applications that have modest
dynamics. The instrument is capable of measuring both the resistive and
reactive parts of the impedance. Industrial Tomography Systems Ltd. have
recently succeeded in obtaining certication for an intrinsically safe option
for their ERT system.
10.2.6.
10.3.
DATA PROCESSING
Mode
Comments
Aberdeen [80]
Barcelona [29]
Bergen [30]
ERT
ERT
ECT
ERT
ERT
Delft [31]
De Montfort [32]
ECT
EIT
ERT
Keele [20]
ERT
Kuopio [34]
EIT
Lancaster [26]
EMT
ECT
ECT
EIT
ECT
Organization
308
EIT
Moscow [38]
EMT
UMIST/Syngenta
LCT [109]
Sheeld [22]
EIT
ERT
Tabriz [91]
Tampere [83]
Thrace [39]
UCL [99]
UMIST [25]
UMIST Mk 1b [21]
ECT
EIT
EIT
EIT
ECT
ERT
UMIST Mk 2a [40]
UMIST [28]
UMIST [85]
VCIPT [19]
ERT
EIT
EIT
Multimodal
ECT
Warsaw [84]
Data processing
Rensselaer [23]
309
310
Data processing
311
they lie within the known physical limits of calibration, this seems to get rid
of any spurious artefacts and speeds up convergence to the true image. Without such truncation the image accuracy is signicantly degraded.
It is well known that pixel-based image reconstruction is an ill-posed
problem due to the limited number of measurements that are available in
each frame of data. Driven by the desire for interpretation of images,
parametric approaches have been suggested for void fraction in oilgas
ows [47] and determination of the size of the air core in a hydrocyclone
[48]. The latter case will be considered to illustrate the approach. A hydrocyclone was equipped with eight planes of 16 electrodes each for ERT. Xray photographs suggest the stability of a centrally located air core in a
correctly operating hydrocyclone. This information can be used to direct
the parameterization of the process such that the conductivity (s) is
modelled as
r a b3r 2 c10r2 12r 3
where r is distance of the air core from the boundary, and a, b and c are
parameters to be determined. The expected voltages can be calculated
numerically, using the four parameters, and the results compared with the
measurements. Optimization routines are then used to nd the best values
for the parameters and, hence, determine the most likely distribution of
materials in the hydrocyclone. A parametric approach can be very attractive
for the ecient reconstruction of high quality images for processes that have
well understood behaviour. Clearly, care should be taken to ensure that the
starting assumptions about the process, in this case the location and stability
of the air core, are valid under all conditions.
Motivated by the possibility of learning good solutions and an anity
for improved speed via parallel computation, a number of neurally inspired
approaches have been considered for processing tomographic data. Most of
these, for instance [4951], are based on derivatives of conventional multilayer perceptrons and have been implemented in software and tested o-line.
Results are interesting, for limited data sets, but have not yet revealed significant benet over conventional techniques. In addition, the multi-layer
perceptron networks suer from extensive learning cycles, which often
yield rigid network congurations, in terms of connectivity, that are not
readily updated when conditions change. One approach [52], using a socalled weightless neural network which is eectively an exotic look-up
table, has been implemented in hardware and tested on-line using an ECT
system. Although this approach oers some potential improvement in
speed, the quality of the resulting images to date are no better than those
from simple linear back projection, and more eort is needed if signicant
advantage is to be realized.
A signicant development is the EIDORS (Electrical Impedance and
Diuse Optical Tomography Reconstruction Software) project [53]. This
312
10.4.
Previous reviews have summarized early applications of electrical tomography [14]. Table 10.3 directs the reader to some interesting recent reports.
The following sections summarize developments in a number of contrasting
application areas. Much of the material has been extracted, with approval,
from earlier papers by the researchers involved. Rather than presenting
exhaustive lists, it is intended that reference to recent publications will
direct the reader to related earlier work. Criteria for selection of the applications presented here include progress towards industrial benet, contrasting
modalities, sensing challenge and on-going eort.
10.4.1.
313
Modality
Status
ECT
ERT
ERT
ECT
ECT
ERT
EMT
ERT
ERT
ERT
ERT
ECT
ECT
ERT
ECT, ERT
ECT
ERT
ERT, ECT
ECT
ECT
ERT
ECT
ECT
EMT
ERT
EMT
ECT
ECT
ERT
Industrial tests
Industrial tests
Industrial tests
Industrial tests
Industrial tests
Industrial tests
Industrial tests
Field tests
Field tests
Field tests
Field tests
Laboratory tests
Laboratory tests
Laboratory tests
Laboratory tests
Laboratory tests
Laboratory tests
Laboratory tests
Laboratory tests
Laboratory tests
Laboratory tests
Laboratory tests
Laboratory tests
Theoretical
Laboratory tests
Industrials
Laboratory tests
Laboratory tests
Laboratory tests
reactor vessel and stirrer arrangements were designed to mimic those that might
typically be encountered in the pharmaceuticals industry. For instance, a
retreat curve impeller (RCI), similar to those tted in 50% of pilot plant
stirred tanks in GSK Chemical Development, has been studied. A schematic
of the impeller is shown in gure 10.6. All impellers were coated with PTFE
to prevent interference of the impeller with the electrical eld.
Mixing time is often used to assess quantitatively the blending performance of stirred tanks. It was decided to study t99 , which is the time required
to reach 99% of homogeneity. Using conductivity probes it is possible to
detect as many dierent values of the mixing time as there are probes in
the reactor. All those values are equally valid and represent the mixing
time at a particular location in the tank. A value of t99 over the whole
314
Figure 10.5.
tank can be obtained by combining all these local measurements. Its value
will vary with the increase in the number of probes until it reaches a plateau
where an increase in the number of probes has only a marginal eect.
Using the adjacent current strategy for the 64 electrode ERT sensor
described above, there are eectively 1264 non-intrusive electrical conductivity probes so that a much higher data density is obtained when recording the
distribution of a tracer compared with the traditional method of inserting
conductivity probes.
The tracer distribution images obtained from the mixing time experiments were compared with computational uid dynamics (CFD) results, as
Figure 10.6.
Figure 10.7.
315
shown in gure 10.7. The tracer is seen to cover a large proportion of the
surface before being ingested into the bulk. After it reaches the impeller a
well mixed zone emerges. The nal layer to be mixed lies between the well
mixed impeller zone and the surface. The results suggest that there is some
advantage to adding material close to the bae and working with a liquid
height equal to the impeller diameter. Although there is reasonable agreement a shift in time steps is observed between the images from ERT and
CFD. Two possible reasons are suggested to account for this discrepancy.
First, CFD evaluates mixing time over the whole bulk. Second, the CFD
software may be unable to model large eddy structures which are known
to have an impact on mixing time.
Observation of the oscillations of the electrical conductivity over 20
pixels after tracer addition allow t99 to be deduced. The stirrer speed was
varied over a range so that measurements took place in the turbulent ow
regime (Re > 1000 for the RCI). In general, the mixing time measurements
showed good reproducibility and followed the expected trend, i.e. mixing
time decreased when increasing stirrer speed. The data obtained were
compared with correlations available from the literature for liquid height
equal to tank diameter. Figure 10.8 shows good agreement with the correlation described by Nienow [96].
Conclusions from this work suggest that ERT shows promise for online control of process mixing performance, as well as eciency evaluation
and optimization of reactor geometries. Results show successful modelling
and analysis of pharmaceutical mixing processes. ERT is capable of oering
superior mixing time information for vessel characterization purposes
compared with existing techniques, and can also provide valuable data for
316
Figure 10.8.
CFD validations. The authors plan for the work to evolve to an increased
level of process complexity with the study of multiphase, solid/liquid
systems.
10.4.2.
Figure 10.9.
317
318
Figure 10.10.
319
320
Figure 10.11.
Figure 10.12.
321
322
Figure 10.13.
the normal operation of the lter. The design has evolved to the mark IV
version, 50 mm diameter, as shown in gure 10.14.
. Materials of construction: In common with the majority of processes
operated within the chemical industry, the materials of construction of
the subject process unit were carefully selected to prevent erosion and
corrosion. The demonstration lter is predominantly hastelloy-C276, an
alloy of nickel, with a mesh fabricated from polypropylene. These
materials, together with PTFE, PVDF and viton, for the O-ring elastomer,
were used exclusively in the electrode assembly.
. Cable routing: The pressure vessel had no provision for additional anges
through which the 24 electrode cables could exit. Surprisingly, for such a
large vessel, the best solution involved routing the 24 cables through two
1 cm diameter air balance ports.
Figure 10.14.
323
Operational constraints: As the demonstration unit was also a manufacturing asset, access to get into the lter to t the electrodes was severely
restricted to an existing time window during the planned annual maintenance period. The eect of this was to limit the electrode installation
time to a four-day period each year. The usable resource was further
constrained as safety procedures dictated that to ensure a breathable
atmosphere within the vessel only two people could enter the unit at any
one time.
10.4.3.1.
Results
Figure 10.15 shows representative results that compare the level measurements of the ltrate in the vessel with the mean signal from the tomography
system. The eect of the slurry, acetic acid and water washes can be seen and
the tomographic measurements clearly track the process. The tomography
measurements lag behind the level measurements and it is reasonable to
assume that this is due to the time for the liquid to pass through the cake.
A simple algorithm, that assumes that the conductivity in regions of the
cake is reected by local measurements, has been used to provide a crude
estimation of the conductivity distribution. The cross-section is divided
into six regions and a representative image is shown in gure 10.16(a),
where the darker colour corresponds to a wetter region of the cake. The
time evolution of the wetness during a batch is also recorded, as shown in
gure 10.16(b). This and other information is available on a dedicated
web-site that is available on the Syngenta intranet. The information is
updated every 15 min and can be readily accessed by the plant operators.
The EIDORS 3D software toolsuite is being used to explore possibilities
for 3D image reconstruction. The model incorporates the vessel furniture,
such as hold-down bars and central metal pillar, and results using simulated
data are shown in gure 10.17. In this simulation two inhomogeneities
are introduced, representing above average and below average conductivity.
The reconstructed inhomogeneites are clearly visible in gure 10.17.
Unfortunately, eects due to the Zener barrier diodes in the intrinsically
safe instrument lead to diculties in reconstructing images from real
measurements and this aspect is currently under consideration.
The instrument has been operating on a continuous basis for about three
years. Results are repeatable and the electrodes are transparent to the
process. The main challenge is to deliver 3D images and this is being impeded
by the proliferation of metal current sinks in the vessel. Work is on-going to
produce an accurate forward model under these circumstances which will, in
turn, allow good images to be reconstructed. Subsequently, if the cost of
instruments can be signicantly reduced, then it is likely that the use of
the technology in related applications will spread and generate tangible
benets.
324
Electrical tomography for industrial applications
Figure 10.15.
Figure 10.16.
325
326
Figure 10.17.
10.4.4.
Figure 10.18.
327
328
Figure 10.19. Images at various times from the gravity-drop ow test. White represents
solids, black is air.
Figure 10.20.
Normalized correlogram.
329
Figure 10.21. Concentration (left-hand scale) and velocity (right-hand scale) against time
in centre zone.
artefacts caused by sharp-edged windows. This shaping is known as apodization and various apodization functions are programmed into the Tomoow
R100 ECT. The results presented here use the common Hanning window,
which is a smooth bell shape.
Figure 10.21 shows the concentration and velocity against time for the
central zone of a 13-zone map for a typical test with data acquisition of
200 frames per second. The dashed line shows the concentration in the rst
plane, light grey shows the concentration in the second plane and the
black line shows velocity.
The velocity of the plug starts at about 2.80 m/s, rising to about 3.70 m/s.
This speed increase is consistent with the fact that the lowest beads fall about
0.4 m before arriving at the upper plane of the sensor, while the upper beads
fall about 0.7 m. The beads falling from the funnel after the rst plug show a
steady velocity of about 3.70 m/s and though barely discernible in gure
10.19 the signals correlate well between the two planes, as shown in gure
10.21.
Integrating the whole ow period between 2 and 8 s gives an estimate of
volume of 2335 cm3 , compared with the actual value of 2379 cm3 within
2%. The plug between 2.5 and 3.2 s can be separately integrated and
shows a volume of 591 cm3 . This plug volume corresponds to a cylinder of
330
4.95 cm diameter and 30.7 cm length, which is the cylinder of beads from
the top of the valve to the top of the beads within the part-lled funnel, as
shown in lighter grey in gure 10.18. It appears then that as the valve is
opened the entire volume of the cylinder of beads supported by the valve,
both in the cylindrical section and within the funnel, drops as one accelerating mass down through the centre of the sensor. The remaining beads
within the funnel then trickle out in the manner of an egg-timer at a much
lower rate. An understanding of this type of behaviour will assist in the
design of industrial hoppers or silos, where many types of solids may be
dicult to discharge.
This work demonstrates the feasibility of making a owmeter for blown
and gravity-fed solids. A few technical challenges remain, for instance
calibration and varying moisture content of materials, but these are likely
to be solved in the near future. The main obstacle to implementing a full
scale commercial integrated owmeter is availability of capital on the 35
years scale to fund the large engineering programme to launch the product.
This would involve engineering design, integration of electronics, manufacturing route, marketing, distribution and servicing. The technical risk is
small, but the commercial risk is dicult to evaluate as there is not a current
market because such owmeters do not exist.
10.4.5.
331
Figure 10.22.
332
Figure 10.23.
Figure 10.24.
Figure 10.25.
333
to the passage of the slug front. Similarly, the last four images show the
passage of the slugs tail through the measurement plane.
The use of a twin plane system allows the shape of the slugs to be reconstructed, as shown in gure 10.25. The pixels lying along a vertical line
passing through the centre are selected from each frame. These are combined
to give a longitudinal cross-section of the slug, as shown in gure 10.26. Diculties associated with such images include limited spatial resolution in the
cross-sectional images, averaging of the concentration of solids along the
length of the electrodes and smearing of boundaries between phases.
If a model relating the dielectric permittivity to the bulk density is
known, it is possible to extract an average solids distribution from the
cross-sectional image. Using a simple linear relationship, the average
solids distribution is plotted in gure 10.27 as a function of frame number
Figure 10.26.
334
Figure 10.27.
Figure 10.28.
vertical pipe.
335
10.4.6.
Natural gas from a well contains condensable materials, such as water and
hydrocarbons, which must be separated from the gas stream. Traditional
336
Figure 10.29.
wet gas.
Figure 10.30.
337
diameter and has eight electrodes. The sensor is able to operate from 20 to
60 8C and pressure up to 150 bar. Ideally the sensor should be in direct
contact with the gas stream, but because of electrical insulation requirements
a very thin insulating layer has to be applied to the electrodes. In the present
design, a 0.5 mm PEEK inner sheath is used, to maintain high sensitivity.
Sensor 1 is located immediately down-stream of the airfoil. Sensor 2 is
located immediately up-stream of the vortex nder.
The sensor is calibrated using two materials having dierent, known,
permittivities to determine the wall capacitance and standing capacitance.
In this way the permittivity of a third material can be estimated. Experiments were conducted using an air/water ow Twister. Humidity was
varied from 20 to 95% and the temperature from 35 to 50 8C to obtain
dierent concentrations of water droplets. The linear back-projection
algorithm was used for rapid on-line monitoring and the Landweber iterative algorithm was used for more accurate o-line image reconstruction.
Figure 10.31 shows representative images using the iterative algorithm.
Without the airfoil water droplets are distributed almost uniformly over
the cross section of sensor. When the airfoil is in place, water is accumulated
on the walls of both sensors. Hollow cores of the vortex are suggested by the
dark regions.
338
Figure 10.31.
10.5.
SUMMARY
Summary
339
340
ACKNOWLEDGEMENTS
Many thanks to the following for approving the inclusion of their work and
for facilitating appropriate materials: Tom Dyakowski, Bruce Grieve, Andy
Hunt, Tony Peyton, Francois Ricard, Mi Wang and Wu Qiang Yang.
REFERENCES
[1] S R Deans 1983 The Radon Transform and Some of its Applications, Krieger
Publishing
[2] S Webb 1990 From the Watching of Shadows, Adam Hilger
[3] Mathematics and Physics of Emerging Biomedical Imaging 1996 National Research
Council, National Academy Press
[4] Measurement Science and Technology 1996 Special Issue on Process Tomography
7(3) 308315
[5] World Congress on Industrial Process Tomography, Buxton, UK (1999); Hannover,
Germany (2001); Ban, Canada (2003)
[6] Proc. of 1st European Concerted Action on Process Tomography (ECAPT)
Workshop, Manchester, UK (1992)
[7] Proc. of 2nd European Concerted Action on Process Tomography Workshop,
Karlsruhe, Germany (1993)
References
341
[8] Proc. of 3rd European Concerted Action on Process Tomography Workshop, Oporto,
Portugal, 2426 March (1994)
[9] Proc. of 4th European Concerted Action on Process Tomography Workshop, Bergen,
Norway, 68 April (1995)
[10] D M Scott and R A Williams eds 1995 Frontiers in Industrial Process Tomography I,
AIChE
[11] Proc. of Frontiers in Industrial Process Tomography II, Delft, Holland, 912 April
(1997)
[12] Special issue of the Chemical Engineering Journal, 77(1/2) (2000)
[13] Special issue of Measurement and Control, 30(7) (1997)
[14] M S Beck, T Dyakowski and R A Williams 1998 Process tomographythe state of
the art Trans. Inst. Meas. and Control 20(4) 163177
[15] C G Xie, N Reinecke, M S Beck, D Mewes and R A Williams 1995 Electrical tomography techniques for process engineering applications Chem. Eng. J. 56 127133
[16] K Boone, D Barber and B H Brown 1997 Review: imaging with electricity: report of
the European concerted action on impedance tomography J. Med. Eng. Technol.
21(6) 201232
[17] R A Williams and M S Beck 1995 Process Tomography: Principles, Techniques and
Applications, Butterworth Heinemann
[18] F J Dickin, B S Hoyle, A Hunt, S M Huang, O Ilyas, C Lenn, R C Waterfall, R A
Williams, C G Xie and M S Beck 1992 Tomographic imaging of industrial process
equipment: techniques and applications IEE Proc-G 39(1) 7282
[19] B S Hoyle, X Jia, F J W Podd, H I Schlaberg, H S Tan, M Wang, R M West, R A
Williams and T A York 2001 Design and application of a multi-modal process
tomography system Meas. Sci. Tech. 12(8) 11571165
[20] P Record 1994 Single-plane multifrequency electrical impedance instrumentation
Physiol. Meas. 15 A29A35
[21] M Wang 1995 Impedance sensorsconducting systems, in Process Tomography:
Principles, Techniques and Applications ed Williams R A and Beck M S, Butterworth
Heinemann
[22] A J Wilson, P Milnes, A Waterworth, R H Smallwood and B H Brown 2001 Mk
3.5A modular, multi-frequency successor to the Mk 3a EIS/EIT Physiol. Meas.
22(1) 4954
[23] R D Cook, G J Saulnier, D G Gisser, J Goble, J C Newell and D Isaacson 1994
ACT3: A high speed, high-precision electrical impedance tomography IEEE
Trans. Biomed. Eng. 41 713722
[24] A Hartov, R A Mazzarese, F R Reiss, T E Kerner, K S Osterman, D B Williams and
K D Paulsen 2000 A multichannel continuously selectable multifrequency
electrical impedance spectroscopy measurement system IEEE Trans. Biomed. Eng.
47(1) 4958
[25] W Q Yang 1997 Hardware design of electrical capacitance tomography systems
Meas. Sci. Tech. 7(3) 225232
[26] A J Peyton, A R Borges, J de Oliveira, G M Lyon, Z Z Yu, M W Brown and J
Ferreira 1999 Development of electromagnetic tomography (EMT) for industrial
applications. Part 1: Sensor design and instrumentation, in 1st World Congress on
Industrial Process Tomography, Buxton, UK, 1417 April
[27] R E Beissner, J H Rose and N Nakagawa 1999 Pulsed eddy current method: an
overview Rev. of Progress in Quant. NDE 18 469474
342
References
343
344
[60] W Daily and A Ramirez 1999 The role of electrical resistance tomography in the US
nuclear waste site characterization program, in 1st World Congress on Industrial
Process Tomography, Buxton, UK, 1417 April, 25
[61] A Binley, W Daily and A Ramirez 1999 Detecting leaks from waste storage ponds
using electrical tomographic methods, in 1st World Congress on Industrial Process
Tomography, Buxton, UK, 1417 April, 613
[62] M Gasulla, J Jordana and R Pallas-Areny 1999 2D and 3D subsurface resistivity
imaging using a constrained least-squares algorithm, in 1st World Congress on
Industrial Process Tomography, Buxton, UK, 1417 April, 2027
[63] R C Waterfall, R He, P Wolanski and Z Gut 1999 Monitoring ame position and
stability in combustion cans using ECT, in 1st World Congress on Industrial Process
Tomography, Buxton, UK, 1417 April, 3538
[64] R B White 2001 Using electrical capacitance tomography to monitor gas voids in a
packed bed of solids, in Proc. 2nd World Congress on Industrial Process
Tomography, Hannover, Germany, 307314
[65] M A Bennett, S P Luke, X Jia, R M West and R A Williams 1999 Analysis and ow
regime identication of bubble column dynamics, in 1st World Congress on
Industrial Process Tomography, Buxton, UK, 1417 April, 5461
[66] K L Ostrowski, R A Williams, S P Luke and M A Bennett 2000 Application of
capacitance electrical tomography for on-line and o-line analysis of ow patterns
in a horizontal pipeline of a pneumatic conveyer Chem. Eng. J. 77(1/2) 4350
[67] R Mann, S Stanley, D Vlaev, E Wabo and K Primrose 2001 Augmented-reality
visualisation of uid mixing in stirred chemical reactors using electrical resistance
tomography J. Elec. Imaging 10(3) 620629
[68] J J Cilliers, M Wang and S J Neethling 1999 Measuring owing foam density distributions using ERT, in 1st World Congress on Industrial Process Tomography,
Buxton, UK, 1417 April, 108112
[69] S J Wang, D Geldart, M S Beck and T Dyakowski 2000 A behaviour of a catalyst
powder owing down in a dipleg Chem. Eng. J. 77(1/2) 5156
[70] R A Williams, S P Luke, K L Ostrowski and M A Bennett 2000 Measurement of
bulk particulates on belt conveyor using dielectric tomography Chem. Eng. J.
77(1/2) 5764
[71] M Wang, S Johnstone, W J N Pritchard and T A York 1999 Modelling and mapping
electrical resistance changes due to hearth erosion in a cold model of a blast
furnace, in 1st World Congress on Industrial Process Tomography, Buxton, UK,
1417 April, 161166
[72] A Plaskowski, T Piotrowski and M Fraczak 2002 Electrical process tomography
application to industrial safety problems, in 2nd International Symposium on Process
Tomography, Wroclaw, Poland, 6372 (ISBN 83-7083-643-8)
[73] K Tomkiewicz, A Plaskowski, M S Beck and M Byars 1999 Testing of the failure of
solid rocket propellant with tomography methods, in 1st World Congress on
Industrial Process Tomography, Buxton, UK, 1417 April, 249255
[74] M H Pham, Y Hua and N B Gray 1999 Eddy current tomography for metal
solidication imaging, in 1st World Congress on Industrial Process Tomography,
Buxton, UK, 1417 April, 451458
[75] R Thorn, G A Johansen and E A Hammer 1999 Three-phase ow measurement in
the oshore oil industryis there a place for process tomography, in 1st World
Congress on Industrial Process Tomography, Buxton, UK, 1417 April, 228235
References
345
[76] E Yuen, D Vlaev, R Mann, T Dyakowski, B Grieve and T A York 2000 Applying
electrical resistance tomography (ERT) to soliduid ltration processes, in World
Filtration Congress 8, The Brighton Centre, Brighton, England, 37 April
[77] B D Grieve, J Davidson, R Mann, W R B Lionheart, T A York 2003 Process
compliant electrical impedance tomography for wide-scale exploitation on
industrial vessels, in 3rd World Congress on Industrial Process Tomography, Ban,
Canada, 25 September
[78] R A Williams and T A York 1998 Microtomographic sensors for microfactories, in
International Conference on Process Innovation and Intensication, G-Mex Centre,
Manchester, 2122 October
[79] A J Wilkinson, E W Randall, D Durrett, T Naidoo and J J Cilliers 2003 The design
of a 500 frames/second ERT data capture system and an evaluation of its
performance, in 3rd World Congress on Industrial Process Tomography, Ban,
Canada, 25 September
[80] J J A Van Weereld, D A L Collie and M A Player 2001 A fast resistance measurement system for impedance tomography using a bipolar DC pulse method Meas.
Sci. Tech. 12 10021011
[81] M Byars and J D Pendleton 2003 A new high-speed control interface for an electrical
capacitance tomography system, in 3rd World Congress on Industrial Process Tomography, Ban, Canada, 25 September
[82] M Wang, W Yin and N Holliday 2002 A highly adaptive electrical impedance
sensing system for ow measurement Meas. Sci. Tech. 13 18841889
[83] S Zhou and J Halttunen 2003 Monitoring of air bubbles in pulp ow based on
electrical impedance tomography, in 3rd World Congress on Industrial Process
Tomography, Ban, Canada, 25 September
[84] P Brzeski, J Mirkowski, T Olszewski, A Plskowski, W Smolik and R Szabatin 2003
Capacitance tomograph for dynamic process imaging, in 3rd World Congress on
Industrial Process Tomography, Ban, Canada, 25 September
[85] T A York, Q Smit, J L Davidson and B D Grieve 2003 An intrinsically safe electrical
tomography system, in IEEE International Symposium on Industrial Electronics, Rio
de Janeiro, Brazil, 912 June (ISBN 0-7803-7912-8)
[86] H S Tapp and A J Peyton 2003 A state of the art review of electromagnetic
tomography, in 3rd World Congress on Industrial Process Tomography, Ban,
Canada, 25 September
[87] H Griths 2001 Magnetic induction tomography Meas. Sci. Tech. 12 11261131
[88] S Ramli and A J Peyton 1999 Feasibility study of planar-array electromagnetic
inductance tomography, in 1st World Congress on Industrial Process Tomography,
Buxton, UK, 1417 April, 5461, 502510
[89] G Miller, P Gaydecki, S Quek, B T Fernandes and M A M Zaid 2003 Detection
and imaging of surface corrosion on steel reinforcing bars using a phase-sensitive
inductive sensor intended for use with concrete NDT 36 1926
[90] M He, Z Liu, L J Xu and L A Xu 2001 Multi-excitation-mode electromagnetic
tomography (EMT) system, in Proc. 2nd World Congress on Industrial Process
Tomography, Hannover, Germany, 247255
[91] J Frounchi and A-R Bazzazi 2003 High resolution rotary electrical capacitance
tomography system, in 3rd World Congress on Industrial Process Tomography,
Ban, Canada, 25 September
[92] Special Issue of Meas. Sci. Tech. 12 2001
346
References
347
[109] B D Grieve, T A York and A Burnett-Thompson 2004 Low cost, non-invasive, real
time, 3D, electrical impedance imaging: a new instrument to meet the needs of
industry, research and education, in APACT 04, The Assembly Rooms, Bath,
April
[110] R Halter, A Hartov and K D Paulsen 2004 Design and implementation of a high
frequency electrical impedance tomography system Phys. Meas. 25(1) 379390
Chapter 11
EIT: The view from Shefeld
D C Barber
11.1.
BEGINNINGS
349
11.2.
Looking back on the very early days it was clear that there were many things
we did not know about. We did not know about ill-posed problems and
regularization. We did know about reciprocity, but initially did not appreciate
the fact that there were only a limited number of independent current patterns.
It is of course obvious that if you have N electrodes there are only N 1
independent current patterns, but it wasnt obvious to us (or at least to me)
then. So the rst system Brian built generated data using all current bipolar
patterns, from adjacent to 1808 apart. We did see the sense in back-projecting
along equipotentials, so these were constructed for all current patterns (in 2D
with a circular boundary and point electrodes) and everything was backprojected. With 16 electrodes there were 1920 measurements and all of them
were used [1, 2]. We continued to do this until Andrew Seagar contacted us
from New Zealand and pointed out that we only had 104 independent
measurements. The logic was impeccable and Andrew came to join us.
Andrews thesis [3] was a model of rigour and claried many things for us.
It also used distributed current patterns! It was also realistically pessimistic
about the likely image quality we could expect. This was an early introduction
to the idea of ill-posed problems. I still think it took some time before it really
settled in. I certainly remained optimistic about how much image quality might
be improved for a long time after Andrew left us (perhaps because he was not
there).
At the time we did not call the technique EIT but applied potential
tomography (APT) [4]. This was because our experience to date with electrophysiological measurements had been with internally generated signals
(EMG, ECG etc.), and in the case of APT the currents were applied from
outside. Once other groups had taken up EIT it became clear that this was
the favoured name for the technique and we converted to it, but it was
hard to drop the name APT locally.
350
11.2.1.
Figure 11.1.
351
between a pair of adjacent electrodes and then moving these electrodes closer
and closer together, increasing the current as this is done to maintain the
voltage levels on the surface of the object. In the limit current input and
output (source and sink) are at the same point, which is dicult to realize
practically, but mathematically this is acceptable, just like any other
dipole. The equipotentials for a dipole drive were easy to compute and
formed the basis of our back-projection algorithm.
11.2.2.
We knew that simply back-projecting the voltage dierence was not correct.
We wanted to back-project a resistance value, and although the voltage
dierence between a pair of electrodes was dependent on the resistance
between the equipotentials, it was also dependent on the area between the
equipotentials. We knew we would have to normalize the data in some
way. The obvious way was to calculate the voltage between the two
electrodes when the resistance was some standard uniform value, measure
the equivalent voltage dierence on the object being imaged and then take
the ratio of these two values. Provided the current did not change, the
ratio of these values was the same as the ratio of the two resistances, the
standard value and the unknown value (assuming that the resistance changed
uniformly between the equipotentials), and so this value could be safely backprojected. Thus was born dierential imaging. A more subtle argument
convinced us that we should be back-projecting the logarithm of the ratio.
This could then be approximated by the ratio of the dierence in voltage
values divided by the reference value (or possibly the average of the
values). Since we were only equipped to deal with small changes in conductance (because the algorithm was linear), the dierence between log of ratios
or normalized dierences was of no real signicance.
There was one other feature which we added to the back-projection.
Simple back-projection clearly did not work very well near the edges of the
image being reconstructed. This was most obvious when reconstructing
point objects. Circular objects became elongated in a direction normal to
the boundary. The reason was not dicult to see if the equipotentials passing
through the object were inspected. At the boundary, all equipotentials are
normal to the boundary. Close to the boundary, the majority run in a direction close to the normal. But if we are going to be able to reconstruct the
point object accurately we need back-projections going through the object
in all directions. This is what happens in the CT algorithm. To make it
happen in EIT we need to give a larger weighting to those data projecting
along equipotentials at large angles to the normal to the boundary, and
smaller weights (because there are more of them) to those data backprojecting along equipotentials more parallel to the normal to the boundary.
We need isotropic back-projection. After some struggles with trigonometry
352
the appropriate weights were calculated and applied [5]. It was gratifying that
subsequently a much more rigorous analysis came up with the same result [6],
and in fact a subsequent analysis by us based on conformal transformations
(again this was all in 2D) provided a much simpler route to the weights [7].
With our initial approach to calculating the weights, it was only possible
to calculate weights for the case when the drive electrodes were adjacent,
and it was this fact that, from a reconstruction standpoint, dictated the use
of the adjacent drive conguration. Later it became possible to calculate
the weights for other bipolar drive congurations [7], but by then we had
moved on to other approaches to reconstruction. Fan-beam CT also uses a
weighted back-projection for similar reasons.
We knew that, by itself, the back-projection algorithm could not give
uniform resolution across the image. Resolution was always worst at the
centre, but improved as the point object being imaged moved towards the
boundary. Further analysis (again using conformal transforms based on
the work in Andrew Seagars thesis) produced a measure of the resolution
as a function of the distance of the point object from the centre. Clearly, if
the resolution was to be improved further we needed to perform some
image processing. Two approaches were tried. We found a radial transform
which (approximately) transformed the image into one with uniform resolution (the boundary went o to innity) and applied standard position
independent image lters, using fast Fourier transform (FFT) methods, to
improve resolution [5, 8]. The other approach constructed a simple
position-dependent enhancing lter and applied it to the image. This lter
was combined with the matrix used for back-projection to create a set of
reconstruction weights, and these weights went out with the rst APT
systems we produced. The decision to use this approach, rather than the
FFT method, was made largely on the basis of simplicity and speed. The
computer systems we were using were not very powerful. I am not sure I
would do the same today.
All the above was based on a linear model of reconstruction. We knew
that the problem was not linear, we knew that objects were 3D rather than
2D and did not have circular boundaries, and we knew that the equipotentials did not run through the object as though its resistance was uniform.
However, there was one overriding consideration which dictated our
choice of reconstruction methods and that was that we wanted to reconstruct
images using data taken from human subjects.
11.3.
DIFFERENTIAL IMAGING
Dierential imaging
Figure 11.2.
353
dishes of saline) we could possibly calculate this, but we did not have access
to and experience of nite element techniques then. If we had had such
methods, within the limits of our reconstruction algorithm, we could have,
in principle, produced images of the absolute distribution of resistivity. As
we had a Radiotherapy section within the department, we did have access
to techniques for making plastic moulds of parts of the body. In Radiotherapy these moulds are for patient immobilization, but in our case we
were looking for a copy of the body surface that we could ll with saline
to measure the reference data. We made a model of Rod Smallwoods arm
and inserted a ring of electrodes inside (drawing pins, points outwards!).
We then made a set of measurements on his arm, took his arm out, blanked
o the ends of the mould, lled it with saline and made a second set of
measurements. An image was reconstructed and turned out quite well, showing all the basic structures [2]. Figure 11.2 shows an example of the sort of
images we were able to obtain. These actually represented the rst absolute
images of human subjects, although the forearm was not an area of major
clinical interest! The bones could easily be seen (high resistivity is represented
by black) and possibly a layer of surface fat. We convinced ourselves we
could see other structures [2].
Although we considered this approach as a possible way of getting
images, it was obvious that it was not really practical. Attempts to directly
compute reference data were not very successful, but in the course of looking
at data from the head we did discover that images could be produced if we
concentrated on changes in resistivity. More importantly, we could also do
the same using data from the chest. So although static imaging was proving
dicult, it was possible to produce dynamic images from data which changed
over time and from then on, for many years, we focused on such imaging. We
354
eventually changed the name to dierential imaging, but the principles were
the same.
Dierential imaging was more than a convenience. The measured
voltages on the surface of an object are determined by the shape of the
object, the placing of the electrodes on the surface of the object and the
internal resistivity distribution. The rst two of these are usually dominant
and for successful reconstruction of resistivity distributions must be
accounted for in some way. As an example, the voltage dierence between
electrodes can be measured to 0.1% accuracy, and this sort of accuracy is
required if useful images are to be obtained. If electrodes are spaced
100 mm apart around a thorax, then a variation in positioning of 0.1 mm
will produce errors of 0.1%. So random electrode placement errors of
1 mm will produce measurement errors 10 times that due to noise [9]. We
felt that it was going to be dicult to determine electrode positions with
this accuracy. However, with dierential imaging this sort of error would
cancel out. This is discussed further in the Appendix.
The reconstruction algorithm also assumed that the electrode pairs were
equally spaced around a circular boundary. Now, in 2D, it can be shown that
all non-circular boundaries can be mapped to a circular boundary using a
conformal transformation. So any boundary with any electrode spacing
can be mapped on to the circle. The electrodes would no longer be placed
uniformly along this equivalent circular boundary, but provided we knew
where they were we could interpolate our data to that produced by electrodes
of uniform spacing. We developed an algorithm which would determine the
boundary shape (and electrode positions) from the measurements (to within
5% accuracy) [10] and an algorithm which would map the non-circular
boundary on to a circle [11], so we had the tools to convert all problems to
the ideal 2D case. Coupled with the use of dierential imaging to deal with
variations in electrode spacing, this went some way towards dealing with
the uncertainties in real data. Oddly enough we never followed this up. It
is dicult to recall the reasoning process which led us to put these results
to one side, but in part it was due to the realization that solving a 2D problem
was not the correct way to tackle 3D problems and partly because we thought
that we should be using a more principled approach to reconstruction,
namely the sensitivity matrix. When we had solved these problems it might
be appropriate to return to the ne details of shape correction. We knew
that, even if the above problems were solved, the assumption of uniform
resistivity for building the sensitivity matrix, or determining equipotentials
for back-projection, was going to run into diculties for situations (such
as the head) where there were signicant deviations from uniform resistivity,
so there were always going to be reconstruction artefacts. Putting in some
a priori information might help, but using this to determine the correct equipotentials and back-projecting along these equipotentials did not seem to
produce spatially correct images [12, 13], so a proper sensitivity matrix was
Collecting data
355
required. We were also being told, correctly, that our approach was only an
approximation, in the case of back-projection with little theoretical support,
and that better algorithms were available, based on sound principles, which
oered the prospect of accurate images of good resolution and that better
current patterns were available. Nevertheless, the dierential algorithm
was the only one to provide images of any quality from in vivo data. In
particular, it allowed us to collect data from 3D objects (humans) but
reconstruct images using a 2D algorithm. The images were not accurate
but looked sensible, and this was very encouraging.
Although there is only one physical property being measured, we can
talk about either resistivity or its reciprocal conductivity. When we moved
to the use of the sensitivity matrix rather than back-projection, the mathematics suggested that we should talk about conductibility, and from this
point on we produced images of changes in conductivity rather than changes
in resistivity.
We had started o by taking the ratio of the data before and after a
change of conductivity, and then the logarithm of the ratio (to get logarithms
of conductivity changes) and then the normalized dierence of the data. In
the limit of small changes in conductivity the last two data transforms
were equivalent. However, whereas our earlier analysis had supported the
view that we were imaging log changes in conductivity, the later sensitivity
matrix approach did not obviously support this view. This was not an
important issue in practice, but nevertheless continued to niggle away in
the background. Huw Griths continued to use ratios of logarithms [14]
and I now believe he was correct to do so. In fact the dierences between
these two approaches can be resolved quite easily. A reworking of the
Sheeld algorithm, including extension to complex data, is given in the
Appendix.
11.4.
COLLECTING DATA
From the beginning of our work we had put signicant eort into the
development of data collection equipment. The developmental approach
we took was heavily inuenced by the desire to collect data from patients,
which meant careful attention to the issues of safety and the problems
associated with electrode impedance, and the need to collect data quickly.
Although there have subsequently been several attempts to develop methods
of determining electrode impedance in vivo, we took the view that this was
not practically possible and that therefore all measurements would be
four-electrode, with current being driven between a pair of electrodes with
measurement of voltage between another pair. The need to collect data at
high speed, because we were looking at dynamic imaging, meant that the
data collection system had to be kept simple and robust.
356
11.4.1.
11.4.2.
The mark 2
The mark 1 machine was completely serial. A current pattern was applied
and the voltages between adjacent electrodes measured one after the other.
One clear improvement we could make was to collect the data in parallel.
We could only apply the current patterns one at a time, but there was no
reason why we could not collect the voltage data from each current pattern
Collecting data
Figure 11.3.
357
in parallel. This was an important step forward. We could collect data much
faster. More importantly, we could spend more time collecting each data
value with improved signal to noise. The machine which did this for us
was the mark 2. Having decided to go parallel, we also decided to go digital.
Demodulation and processing of the signals was made completely digital,
which further improved the signal-to-noise ratio. Given that we could collect
a complete set of high quality data 25 times a second, we decided we needed
to reconstruct and display data at this rate, in other words to go for a realtime system. This could only be done with a simple matrix-based reconstruction algorithm, which of course we had. The reconstruction time on the mark
1 system, using by todays standards a very modest PC, was about 1 s, so
although we could collect data at much faster frame rates the data had to
be processed o-line. In the mark 2 system (gure 11.4) we decided to use
Figure 11.4.
358
a recently developed processor called the transputer, which was fast enough
to implement the reconstruction within the time for data collection. The
novel feature of the transputer was that it was specically designed to be
linked together with other transputers to form an array of processors,
across which computations could be distributed. There was even a parallel
language, OCCAM, developed for it. We linked together four transputers:
one to acquire data, one to reconstruct the images, one to display the
images and one to manage the others. This was a cutting edge approach at
the time and worked remarkably well. Images of an insulating rod moving
in a tank of saline were a common demonstration. Perhaps one of the
most impressive and evocative sequences was the visualization of a stream
of water or concentrated saline poured into a tank of isotonic saline. Used
in dierential mode, the system allowed us to visualize in real time the
changes in conductivity in the heart as it moved through the cardiac cycle.
We still used a.c. current, but this time at 20 kHz. We used the mark 2 to
try to simultaneously identify ventilation defects in the lung by gating data
analysis to breathing, and perfusion defects by gating data analysis to the
cardiac cycle [16]. The aim was to try to detect pulmonary embolism. The
principle was sound, but technically it was very dicult.
11.4.3.
Limitations
For all the reasons given previously, the images were not very reliable. If we
took the electrodes o and replaced them on the patient we would not reliably
get the same images. If the patient moved signicantly between collecting the
reference data set and the second data set there would be artefacts in the
images. Unlike other imaging systems, images of nominally the same part of
the anatomy on two dierent subjects often looked very dierent from each
other. We could produce images of the lungs during respiration, and of the
heart, and obtain gastric emptying curves, but only the latter experiments
seemed to have any practical applications [17]. No one else was faring any
better. The problem was not one of reconstruction algorithms as such. By
this time we had moved on to reconstruction using sensitivity matrices. We
felt that the ad hoc nature of the back-projection algorithm precluded the
possibility of being able to signicantly improve the resolution using this technique. In addition, it was not obvious how this approach could be extended to
3D, which we were beginning to think about. We also wanted to try to improve
resolution by adding more electrodes104 measurements give an eective
pixel size of just under 10% of the image diameter and on a good day we
could obtain a resolution (in a phantom) with our 16-electrode system consistent with this result. With 64 electrodes we could expect to obtain an eective
pixel size of 3% of the image diameter, and if our object was a thorax we would
be talking about a resolution of the order of 1 cm, comparable with a gamma
camera. The problems around our assumptions of circularity were still there,
Multifrequency images
359
but resolution seemed a more pressing problem, and once we had dealt with
resolution we could return to the other issues.
11.5.
MULTIFREQUENCY IMAGES
The mark 3
We decided to build a third data acquisition system, the mark 3. This had 16
electrodes and could collect data at eight frequencies. Previous experience with
the marks 1 and 2 had identied the diculty of making measurements using
the same electrodes through which, at a dierent part of the data collection
cycle, current was owing. The electrodes had to be switched between a current
source and a high impedance voltage measurement system. We knew that this
would cause problems at higher frequencies because of capacitive eects in the
electronics. We therefore decided to separate the 16 electrodes into two
interleaved sets of eight electrodes each, one set for current drive (in adjacent
pairs) and the other set for voltage measurement, an approach used for
other reasons by other groups elsewhere. This simplied the electronics
360
Figure 11.5. Multi-frequency images. Each of these is a dierential (inspiration, expiration) image at the named frequency.
Multifrequency images
361
that the diculties of constructing static images from in vivo data might not be
as dicult as I had always supposed, at least for well-conditioned systems.
11.5.2.
Marks 3a and 3b
This machine became the mark 3a because we were still interested in the
possibility of high resolution and so, having used the 3a as a test bed for
the electronics, we built a 64-electrode system, the mark 3b. This gave us,
in principle, 961 independent measurements which should have delivered
three times the resolution of the marks 1 and 2. This, of course, turned out
to be wishful thinking, but this system did enable us to explore 3D imaging
(more on this below), although it had not been explicitly designed for this and
was not completely optimal for the purpose.
Finally, we developed the mark 3.5 (gure 11.6). In this minimalist
system we reduced the number of electrodes to eight (largely because we
were planning to work with neonates), returned to the idea of using the
same electrodes for current drive and voltage measurement, and expanded
the number of frequencies to 30 in order to determine the ColeCole
362
Figure 11.7. Data collected of absolute lung resistivity from 155 normal infants over the
rst three years of life.
parameters more accurately. We stayed with the triaxial cables from the
mark 3 because they improved accuracy at high frequencies. The number
of independent measurements is 20, which removed all worries about conditioning, and we have used this system to obtain some interesting results on
Figure 11.8.
363
neonatal lung development. In particular we have been able to use these data
to determine the absolute conductivity of lung tissue. This was done using a
model of the thorax. By treating the lung conductivity as a free parameter it is
possible to determine the absolute conductivity of the lung as a function of
frequency. This allowed us to follow the way the impedance spectra of the
lungs changes with age (gure 11.7), and hence quantify the relation between
lung composition and impedance spectra. This approach brings us back to
the original idea which stimulated our interest in EIT, the determination of
body composition (the fat to lean ratio). The system could collect data
from adults as well as neonates (gure 11.8).
I suspect the eight-electrode multi-frequency conguration is probably
close to the optimum for practical 2D EIT.
11.6.
All our work so far had been concerned with 2D imaging, or treating dierential image data as though it was from 2D objects. We knew that this was
not strictly justied. The mark 3b had sucient electrodes to allow us to
collect data over the surface of an object. We concentrated on a 3D conguration consisting of four layers of 16 electrodes, again with an interleaved
pattern on each layer (gure 11.9). This conguration worked well, even
Figure 11.9. Three-dimensional data collection. The images are dierential ventilation
images at eight levels through the chest.
364
though, because the mark 3b had been designed for 2D use, we were not able
to take full advantage of the benets of driving and collecting between layers.
Peter Metherall developed a 3D version of the reconstruction algorithm, and
demonstrated 3D dierential images and images at dierent frequencies.
This work resulted in a paper in Nature [20]. We collected data from the
chest and were able to reconstruct reasonable 3D images of respiration
and cardiac activity, but did not go on to explore other truly 3D geometries,
for example those that might be associated with the breast. Connecting many
electrodes to a patient was not a fast or reliable thing to do, and only a
limited amount of 3D in vivo data was collected. In addition, the dierential
algorithm can run into a problem in 3D which is not found in 2D. The data
used for reconstruction are based on the ratio of two data sets. For 2D data,
at least theoretically, all data have a non-zero value. However, in the case of
3D data it is possible for some data to be truly zero. This can arise, for
example, when the drive electrode pair and the receive electrode pair are
orthogonal to each other. Taking ratios of such zero or near-zero data can
produce large reconstruction errors. With absolute imaging this should not
be a problem, but with dierential imaging it could be quite serious. In practice we identied drive/receive combinations which suered from this
problem, and did not use the data from these when we reconstructed the
images.
11.7.
CLINICAL STUDIES
We have performed various clinical studies using EIT. Perhaps the most
successful were gastric emptying studies, since it did seem possible that
EIT could be used clinically for measuring the rate of gastric emptying without the need for ionizing radiation, especially for paediatric subjects [21]. We
also investigated the use of EIT for lung disease [22], including PE. However,
the technique has not proved robust or reliable enough to be useful for
routine clinical investigation. The multi-frequency work and the measurement of absolute lung conductivity oers some insights into the development
of the neonatal lung [2325]. Absolute conductivity can be used to determine
lung density and air volume. The major use of this appears to be in measuring
lung water and in controlling levels of lung positive pressure when ventilators
are in use. This work has pointed the way to tissue characterization via multifrequency measurements, and Brian Brown has shown how such measurements may be used to dierentiate between normal and diseased cervical
tissue [26]. This may represent the best opportunity so far for impedance
measurements to make a clinical impact, although imaging has not been
used in this work to date. Other groups are also investigating clinical applications and the epilepsy work of the UCL group is particularly interesting,
but formidable technical challenges still remain.
365
I would like to take stock of what I see as the present state of EIT. Medical
EIT as an imaging procedure still represents a signicant technical challenge.
Progress seems slow. The success of EIT depends on the quality of image
reconstruction and it seems to me that no really signicantly new improvements in reconstruction have been published since the mid 1990s. I think it
is possible to draw some general conclusions about the state of EIT at present
and oer them here.
11.8.1.
366
11.8.2.
11.8.3.
Humans are 3D
367
So where does this leave medical EIT? We believe it leaves us with the need
to clearly formulate what problems can be solved. We have to nd clinical
problems which we can solve with robust methods, and then apply those
methods properly. To achieve robust reconstructions in the presence of
random noise and positioning uncertainties, we have to work within the
constraints of a well-posed problem, and the simplest way to do this is
to reduce the number of electrodes. To illustrate this point consider the
case of a 16-electrode 2D system. With an adjacent drive conguration
the condition number is >105 , which is far too large for reliable reconstruction. If we restrict ourselves to condition numbers of 100 then we
can only use 50 of the singular values, i.e. reconstruct images with 50
independent pixels. A 12-electrode adjacent drive system has 54 degrees
of freedom, which would bring it close to the margin. An eight-electrode
system has only 20 degrees of freedom but a condition number of 30.
Data collection and reconstruction with such a system, whilst of poor
resolution, should be relatively robust, and this has been conrmed with
the mark 3.5.
11.8.5.
Some suggestions
368
(d) Test out EIT with anatomically realistic models. There are plenty of
image data around to build such models and many have been built.
There are sucient data on the electrical properties of tissue to allow
physically realistic models to be built and good 3D FE software to
solve them. Demonstration of images derived from such models
would have far greater impact than yet another set of images derived
from a 2D circular mesh!
11.9.
In the form that it has taken so far, it seems unlikely that EIT will be a major
routine clinical tool. Having said that, there are at least two commercial EIT
systems: Transcan (Siemens) for breast imaging and our own eight-electrode
system (Maltron). The most likely applications, in my view, are the breast
and lungs, and if signicant progress could be made in these areas then
EIT might have a future. EIT has been a rich source of funding and research
projects, it has certainly improved greatly our understanding of what determines the impedance of tissue and has furthered many an academic career.
These are valuable aims in themselves, but EIT shows no evidence of
achieving its other goal, which is to provide support for routine health
care. Credibility is wearing thin and it is time to realize some of the promises
made over the past 20 years, or close the shop.
A1
@Apr ; pd ;
@
A2
and by forming
gpr ; pd
@Apr ; pd ; =@
g pr ; pd
A pr ; pd ;
A3
369
A4
where B is a function which is only dependent on the shape of the object and
the position of the electrodes, and h is a function which, although dependent
on the position of the dipoles and the conductivity distribution, is (hopefully)
less dependent on shape than A. The dipole position parameters in h are
dotted to reect the fact that they are the true positions mapped in some
way to t h. Then as before
g pr ; pd
@h p_ r ; p_ d ; =@
gpr ; pd
h p_ r ; p_ d ;
A5
370
A6
A7
Sij ci Fij :
@ logci
@gj
@ci @ logci gj
A8
Equation (A7) relates changes in the logarithm of the boundary values as the
conductivity changes from some reference value to changes in the logarithm
of conductivity values. In previous work we approximated logg by
g=gref , and this approach also ignored the contribution of c in the construction of F. In all our work we had constructed F for uniform reference distribution, so in practice the F we used was the same as the F above, apart from a
scaling factor. Equation (A7) represents a generalization of the Sheeld
algorithm to nonlinear reference distributions.
Complex data
In general, S will be complex. Dehghani has shown that S, for the case of
uniform but complex conductivity, can be written as
S S
A9
References
371
where S is the sensitivity for the real uniform case and is a complex
constant. If we multiply S by the uniform c c, where c is real and
is also a complex constant, then
g Sc g:
1=gj Sij ci
A10
REFERENCES
[1] Barber D C, Brown B H and Freestone I L 1983 Imaging spatial distributions of resistivity using applied potential tomography (APT), in Proceedings of 8th Conference
Information Processing in Medical Imaging ed F Deconinck (Dordrecht: Martinus
Nijho ) 446462
[2] Barber D C, Brown B H and Freestone I F 1983 Experimental results of electrical
impedance tomography, in Proceedings of the 6th International Conference on Electical
Bio-impedance, Zadar, Yugoslavia, Medical Jadertina XV: Supplementary Issue 15
[3] Seagar A D 1983 Probing with low frequency electric currents, PhD thesis, University
of Canterbury, Christchurch, NZ
[4] Barber D C and Brown B H 1984 Applied potential tomography J. Phys. E: Sci.
Instrum. 17 723733
[5] Barber D C and Brown B H 1986 Recent developments in applied potential tomography, in Proceedings of 9th Conference on Information Processing in Medical Imaging
ed S Bacharach (Dordrecht: Martinus Nijho) 106121
[6] Santosa F and Vogelius M 1988 A back-projection algorithm for electrical impedance
imaging. Technical note BN-1081, Department of Mathematics, University of
Maryland, College Park, MD 20742, USA
[7] Barber D C Image Reconstruction in Applied Potential TomographyElectrical
Impedance Tomography INSERM, Unite 305, Toulouse, France.
[8] Barber D C and Seagar A D 1987 Fast reconstruction of resistance images Clin. Phys.
Physiol. Meas. 8 Suppl. 2A 4754
[9] Barber D C and Brown B H 1988 Errors in reconstruction using linear reconstruction
techniques Clin. Phys. Physiol. Meas. 9 Suppl A 101104
372
[10] Kiber M A and Barber D C 1991 Estimation of boundary shape from the voltage
gradient measurements, in Proc. Electrical Impedance Tomography, Copenhagen,
University of Sheeld, 5259
[11] Barber D C and Brown B H 1991 Shape correction in APT image reconstruction, in
Proc. Electrical Impedance Tomography, Copenhagen, University of Sheeld 4451
[12] Avis N J, Barber D C, Brown B H and Kiber M A 1992 Back-projection distortions in
applied potential tomography images due to non-uniform reference conductivity
distributions Clin. Phys. Physiol. Meas. 13 Suppl A 113117
[13] Avis N J, Barber D C, Brown B H and Kiber M A 1991 Distortions in applied potential tomographic images due to non-uniform reference distributions Proc. IEEE
EMBS 13 2021
[14] Griths H, Leung H T and Williams R 1992 Imaging the complex impedance of the
thorax Clin. Phys. Physiol. Meas. 13 Suppl. A 7781
[15] Brown B H, Lindley E, Knowles R and Wilson A J 1990 A body-worn APT system
for space use, in Proc. Electrical Impedance Tomography, Copenhagen, University of
Sheeld 162167
[16] Brown B H, Sinton A M, Barber D C, Leathard A D and McArdle F J 1992 Simultaneous display of lung ventilation and perfusion on a real-time EIT system, in Proc.
14th Ann. Conf. IEEE EMBS, Paris 17101711
[17] Avill R, Mangnall Y F, Bird N C, Brown B H, Barber D C, Seagar A D, Johnson A G
and Read N W 1987 Applied potential tomography: A new non-invasive technique
for measuring gastric emptying Gastroenterology 92 10191026
[18] Brown B H, Barber D C, Wang W, Lu L, Leathard A D, Smallwood R H, Hampshire
A R, Mackay R and Hatzigalanis K 1994 Multi-frequency imaging and modelling of
respiratory related impedance changes Physiol. Meas. 15 Suppl. 2A 111
[19] Noble T J, Morice A H, Channer K S, Milnes P, Harris N and Brown B H 1999 Monitoring patients with left ventricular failure by electrical impedance tomography Eur.
J. Heart Failure 1 379384
[20] Metherall P, Barber D C, Smallwood R H and Brown B H 1996 Three-dimensional
electrical impedance tomography Nature 380(6574) 509512
[21] Lamont G L, Wright J W, Evans D F and Kapila L 1988 An evaluation of applied
potential tomography in the diagnosis of infantile hypertrophic pyloric stenosis
Clin. Phys. and Physiol. Meas. 9 Suppl. A 6569
[22] Campbell J H, Harris N D, Zhang F, Brown B H and Morice A H 1994 Clinical applications of electrical impedance tomography in the monitoring of changes in intrathoracic uid volumes Physiol. Meas. 15 Suppl. 2A 217222
[23] Hampshire A R, Smallwood R H, Brown B H and Primhak R A 1995 Multifrequency
and parametric EIT images of neonatal lungs Physiol. Meas. 16 Suppl. 3A 175189
[24] Brown B H, Primhak R A, Smallwood R H, Milnes P, Narracott A J and Jackson M J
2002 Neonatal lungscan absolute lung resistivity be determined non-invasively?
Med. Biol. Eng. 40 388394
[25] Brown B H, Primhak R A, Smallwood R H, Milnes P, Narracott A J and Jackson M J
2002 Neonatal lungsmaturational changes in lung resistivity spectra Med. Biol.
Eng. 40 506511
[26] Brown B H, Tidy J, Boston K, Blackett A D, Smallwood R H and Sharp F 2000 The
relationship between tissue structure and imposed electrical current ow in cervical
neoplasia The Lancet 355 892895
Chapter 12
EIT for medical applications at Oxford
Brookes 19852003
C McLeod
374
A signicant dierence between our approach and that of RPI is in our use of
independent current-application and voltage-measurement electrodes.
As the theoretical and mathematical modelling work progressed, curiosity
demanded some real experimental work. Mike and Bill had successfully
simulated conductivity distributions, applied current patterns to them and
calculated the resulting voltage patterns; the voltage patterns and current
patterns and added noise could be given to the reconstruction algorithm
which reproduced a recognizable version of the conductivity pattern. Dale
Murphy, another bio-engineer who had been working with Lionel in
Paediatrics, and Chris McLeod, another bio-engineer who had moved from
Paediatrics to Engineering at Oxford Polytechnic, adapted some of the
circuitry which had been used in the single-channel impedance work and
added programmable current sources to produce OXPACT-1, the Oxford
Polytechnic adaptive current Tomograph, in 1987. The performance was
very poor and no images were ever obtained. A great deal was learnt about
the precision needed in the hardware, particularly if the current sources were
to perform correctly when connected together on a conductive object. John
Lidgey, an Engineering lecturer specializing in analogue circuit design,
contributed many ideas for improving the sources [6].
For perspective, an alternative method had been developed by the
Sheeld group, amongst others, involving the use of only a single current
source; the current output could be measured continuously and it did not
have other sources to react with. The current source was applied in turn to
each adjacent pair of electrodes and voltage measured on the remaining
electrodes. From these, equipotential regions were calculated and a weighted
back-projection algorithm applied to produce a conductivity image. The
method works best when applied in a dierence modefrom some reference
physiological state, the dierences in conductivity during a cycle of heart or
breathing activity are imaged.
Any multiple-source system had to have identical sources, or sources
which could be programmed precisely, which would maintain the
programmed current during large impedance changes. Impedance changes
within the body are small, but the electrode contact impedance varies rapidly
due to movement. In the mid-1980s the extra complexity of the instrumentation for multiple-source systems and the success of the adjacent-drive systems
pioneered by Barber and Brown in Sheeld prompted many groups to avoid
the multiple-source method.
The computational task in reconstructing images from the measurements from 32 electrodes for a complete set of current patterns was very
time-consuming for the available computers. A second applied mathematics
post-graduate, Kevin Paulson, joined the group to work on, amongst other
things, reducing the computation time. These were the days of 16 MHz
clock speed PCs and 1 Mbyte memory size. Data les were transferred
from the acquisition system PC to the reconstruction PC on a 514 inch
375
oppy disk. Kevin experimented with Inmos Transputers and the Intel i860
vector co-processor, and achieved some improvement, but not much more
than could be achieved by waiting for the next generation of faster PC chipsets. The extra complexity introduced by having programs written in Occam
for the Transputers and C on the host PC, and the cost of using non-PC
boards and the diculty in maintaining such software, taught us many
lessons. It became clear that faster computers were never going to make
EIT practical and that more sophisticated inversion algorithms were
required. The time required to calculate an EIT image is limited by the
complexity of solving the matrix equation of the form Ax b. Given the
matrix A, with N rows and columns, and data vector b, ON 3 calculations
are required to nd the EIT image vector x. For an EIT system with M electrodes the matrix A has M 2 rows and columns, and so calculating the EIT
image requires OM 6 calculations. If the number of electrodes in an EIT
system are doubled, the time required to calculate the image increases by a
factor of 64. Kevin introduced the concept of optimal measurement patterns
that parallels optimal current patterns. When both sets of optimal patterns
are used, only M of the M 2 possible measurements are non-zero. The
POMPUS algorithm calculates the EIT image using only these non-zero
measurements and so scales as OM 6 . For a 32-electrode system the
POMPUS algorithm is over 32 000 times faster than the standard algorithm.
This development has made possible 3D and high resolution EIT systems
[7, 8].
By 1989 the EIT Group, as we named ourselves, consisted of Mike, Bill
and Kevin, who were primarily working on reconstructionthough no
distinction was drawn between system software and algorithm workand
Chris working on hardware and the low-level hardware drivers with help
from John Lidgey on the current sources. Various undergraduates helped
build some parts, but it was clear that a larger eort was required for building
a more suitable system. The rst electronics postgraduate, Ching (QS) Zhu,
joined the group for the development of the OXPACT-2 system.
Amongst the design changes introduced was the use of voltage sources
for delivering current. This was achieved by measuring the transfer admittance matrix and then calculating the voltage settings required to generate
the required current pattern. The transfer admittances are measured by
applying voltages to the electrodes and measuring the resulting currents.
Errors in the measurements and calculations are iteratively reduced by
using Landwebers algorithm to rene the voltage pattern until the desired
currents are set. Making high-accuracy current sources at high frequencies
(in our case, the design specication for the system was to operate at 10,
40, 160 and 640 kHz) was extremely dicult, so the voltage source idea
was attractive. It also prompted the realization that it does not matter
whether voltages are applied and currents measured or currents applied
and voltages measured, as long as a reasonable basis could be applied. The
376
377
Figure 12.1.
378
Figure 12.2.
2D 30 cm diameter tank.
379
.
.
380
Posterior
Anterior
Figure 12.3(a).
Figure 12.3(b).
381
Tomograph to use excitation at up to 640 kHz (the design allows 10, 40, 160
and 640 kHz). The system included much more digital circuitry, taking
samples at up to 10 million/s. This allowed greater exibility in using the
acquisition section, and greater accuracy through using digital signal generation, ltering and signal demodulation. The number of electrodes remained
the same: 32 for current sources and 32 for voltage measurements. A multiplexer selected the electrode for voltage measurement and measurements
were made sequentially during each applied current pattern. In this respect
the system diered from the contemporary RPI ACT3 system [9], which
has a dedicated processor for each electrode and which measures voltage
on the electrodes through which current is being delivered.
A 3D systemOXBACT-4was built with very limited funding for
tank studies. It was designed for static imaging and to test 3D reconstruction
algorithms. It could therefore be slow and be based on commercially available PC analogue input and digital output cards. The current sources (192)
and the voltage measurement multiplexer (816 channels) were custom-built
to match the eight-layer, 24 current electrodes/layer design. The current
electrodes occupied 30% of the cylindrical surface area and each current
electrode had four voltage electrodes associated with it, one in the centre
of the electrode and one mid-way to the adjacent current electrodes. The
arithmetically-adept need to know that the other 48 voltage electrodes
formed another layer beyond the last current layer. The electrode arrays
and connections were made accurately on exible printed circuit boards
and the tank cast around them in bre-glass. The tank is 30 cm diameter
and 120 cm high, with the electrode region occupying the middle third, as
seen in gure 12.4.
Ching Zhu left to join a medical electronics company in North America
and Dr Yu Shi joined us from the Toulouse group. Yu Shi wrote a wonderful
user interface on the host PC, and mastered the intricacies of the DSP which
drove the acquisition system. A pair of bre optics joined the two parts,
providing a fast, electrically isolated link. That left body shape and the
2D3D issue outstanding. As it happened, all the volunteers for the trial
studies with the new system had very similar chest shape, and a one-sizets-all FE mesh was created whose boundary was well described by only
four Fourier components. FE meshing programmes were appearing in Shareware schemes by this stage, so we nally produced images which had some
chance of convincing non-believers that there was truth in the resultssee
gure 12.5. Of course there was, and still is, no way to verify the truth of
the conductivity values, as there is little data on warm, blood-lled, living
tissue.
Kevin, Mike and Chris initiated a small parallel project on impedance
spectroscopy (EIS), to try to get conductivity data from living human
tissue from a small probe placed on exposed tissue [11, 12]. That work
progresses when funding allows. What it did show was that the quality of
382
Figure 12.4.
3D tank.
Figure 12.5.
mesh.
383
to be valuable. The EIT imaging was therefore only carried out at 40 kHz,
which allowed reasonable currents to be applied and good measurements
to be made.
The new fast system allowed sets of images to be made and hence timeseries analysis could be applied to these (gure 12.6). Nacer Kerrouche
replaced Yu Shi who had left for Australia, and his main work became the
time-series analysis which Bill had started. We applied Principal Component
and Fourier Analysis to the image sets and found that Fourier generated
much clearer and more helpful data. In retrospect, it is quite obvious that
it should, as there are no signicant or non-cyclical movements of tissue
around the chest. The obvious rhythms appeared at respiratory and cardiac
frequencies and there is often a small component at a much longer period
(c. 25 s), which we have speculated may be caused by the autonomic
system. More data is needed to investigate this feature.
Signicant stang changes forced changes in the emphasis of the
Groups work. Bill moved to a very good post at UMIST and added his
own brand of EIT to the existing expertise there. Kevin moved to the RutherfordAppleton laboratory; not far away, but concentrating on other scientic problems. Although we maintain links with both of them, their drive in the
project is greatly missed. This was partly oset by the arrival of Andrea
Borsic from Turin, who came to work on developments of the reconstruction
technique such as anisotropic smoothing and the Total Variation method. In
addition, he was also responsible for a short paper at a medical imaging
summer school on the localization of the sense of humour using a modied
evoked response method.
After many studies on those still-willing volunteers (gure 12.7) in the
laboratory, we felt ready to impose on patients and got ethical approval
for studies on a group of patients in Intensive Care, who had severe
cardio-respiratory problems. The patients were on articial ventilators and
had problems with uid accumulating in their lungs. This ought to be the
cue for some interesting abnormal lung images, but unfortunately the data
from these patients was too poor to reconstruct at all. The outcome will be
perhaps the biggest step change of the whole development, incorporating
advances in:
.
.
.
.
.
384
Figure 12.6. Fourier analysis of an image set: magnitude and phase at the respiratory and
cardiac frequencies. From [13].
Figure 12.7.
385
The importance of the software environment was brought home when Yu Shi
left; his (excellent) coding of the Texas Instruments TMS320C40 digital
signal processor as the data acquisition controller in the C language and
assembler is dicult for his successors to maintain. This is mainly due to
the intricacy of the assembly language for this processor. It is more generally
true for the environment in which university research takes placea succession of bright young researchers come and then go. Over the years, the starting point for new work becomes more sophisticated and the learning curve
correspondingly longer. For the new system we are attempting to separate
the system developer functions, written in a low-level language, from the
EIT researcher functions, written in MATLAB. Fortunately, the boundary
can be dened very simply as the Tomograph applies a current pattern (a
vector) and measures all the voltages (another vector). The EIT researcher
can dene the current vectors and send them, and wait for the returning
measurements. Functionally, whether it is a calibration function or an
imaging session is immaterial.
OXBACT-5 is the name of the new system. In it there are technological
developments whose plans were presented at the Colorado meeting, whose
implementation has taken longer than expected and some of whose results
should be ready for the Gdansk meeting (June 2004). The last year has
been spent on hardware development, so no truly EIT results have been
coming out in that period. The eort will be more justiable if these systems
are used by other groups; we hope that such inter-group co-operation will
help the whole EIT eld to establish the benets of the method, and see it
contribute to patient monitoring and diagnosis in the way we imagined
when we were all motivated to work on its development.
386
The long view of the project is that we believe that technically the
optimal methods are the right ones to pursue; it is more dicult to obtain
absolute conductivity values, but these data should be more valuable than
dierence data for dening the state of tissues. The spatial resolution of
any EIT method with a nite electrode set is limited by the number of
independent data, so more electrodes will give more resolution. In practice,
the limit on number is set by what is possible in an acceptable clinical
technique. In this respect the non-contacting magnetic or inductive methods
have an advantage, but at the expense of providing less precise data.
Electrode technology is improving independently with the development of
micro-needle arrays and non-contacting physiological signal sensors.
The recent interest shown by Microsoft [14] in using the resistance and
conductivity of the body for data entry and signalling, respectively, will
stimulate an orders-of-magnitude increase in EIT, though probably under
another name.
Today the inaccuracy in knowing the 3D spatial co-ordinates of the
electrodes on the surface of a human body remains the biggest error. The
electronics continue to improve and get cheaper, following Moores law
for computing. The software techniqueswhile they remain publicallow
new developers to build on the growing knowledge bases of incorporating
a priori data, and of solving large and complex ill-posed inverse problems.
The following have contributed to the project in chronological order of
start-date:
Lionel Tarassenko
Bill Lionheart
QS (Ching) Zhu
Matthew Rose
Jean-Louis Lottiaux
Andrea Borsic
Mike Pidcock
Kevin Paulson
Tieying Duan
Evelyn Morrison
Nacer Kerrouche
Alex Yue
Dale Murphy
Chris McLeod
Chris Denyer
Annabelle Le Hyaric
Svetlana Jouravleva
Dimitar Kavalov
Peter Furner
John Lidgey
Yu Shi
Mark Bode
REFERENCES
[1] Tarassenko L, Murphy D, Pidcock M and Rolfe P 1985 The development of imaging
techniques for use in the newborn at risk of intraventricular haemorrhage, in Proceedings of the International Conference on Electric and Magnetic Fields in Medicine and
Biology, London
[2] Breckon W and Pidcock M 1988 Some mathematical aspects of electrical impedance
tomography, in Mathematics and Computer Science in Medical Imaging ed M A
Viergever and Todd-Poporek, 204215, Springer
[3] Breckon W and Pidcock M 1988 Ill-posedness and non-linearity in electrical
impedance tomography, in Information Processing in Medical Imaging ed C N de
Graaf and M A Viergever, 235244, Plenum
[4] Isaacson D 1986 Distinguishabilities of conductivities by electric current computed
tomography IEEE-TMI MI-5(6) 9195
References
387
Chapter 13
The Rensselaer experience
J Newell
13.1.
EARLY DEVELOPMENTS
Early developments
389
Figure 13.1. This is ACT 0. It is a coil of copper wire wound around a wooden stick. At
intervals along the coil, wires are connected, which can be connected using clip leads to
electrodes around the inside edge of a circular saline tank. The intervals are irregular,
proportional to a sinusoid. The ends of the coil are connected to the output of a Radio
Shack audio amplier, driven by a signal generator. The result is a set of voltages in a
spatial sinusoid around a circular tank. Data are obtained from a hand-held multimeter,
and recorded by pencil and paper. (The student who spent a summer collecting and analysing this data has since earned a PhD.)
Isaacson wanted some realistic estimates of the noise levels that could be
achieved in a multi-channel instrument. Jon Newell had a laboratory where
electronic experiments in water baths could be done. We did the rst experiments to demonstrate the feasibility of detecting targets in water baths when
the targets were not near the electrodes (see gure 13.1). A sinusoidal pattern
of a.c. voltages was applied to 32 copper electrodes installed at the periphery
of a plastic pie transport dish. There were detectable changes when conducting
targets were placed in the bath, even near the centre of the bath.
This was enough encouragement to interest David Gisser in designing a
computer-controlled set of current sources, and a multiplexed voltmeter [3, 4].
This rst instrument, called an Adaptive Current Tomograph (ACT 1), was
built on a single perforated circuit board with wire-wrap technology (gure
13.2). Its multiplexed voltmeter converted the 12 kHz working signal to a DC
level that was passed to the computer through a commercial I/O board with
an A/D converter. Currents were specied digitally through the same board,
under the control of a language called ASYST. The result was a slow, imprecise
system with 32 current sources. Images were reconstructed from these data
using a non-iterative algorithm, which takes the rst step toward minimizing
the least-squares error between the measured voltages and the voltages
predicted from a uniform conductivity estimate. In the single-step algorithm
used, that rst step is just a constant conductivity. Dave designed this
390
Figure 13.2. This is ACT 1. Arrayed from left to right are 16 dual D/A converters at the
middle of the board, and current sources above and below each, to give a total of 32 current
sources. There are four multiplexers adjacent to the electrode connectors at the bottom
edge. The real and quadrature voltmeters are at the left end. The 50-pin cable to the
data acquisition card in the computer would connect at the upper left. Construction is
wire wrap.
algorithm, and it was written by Steve Simske as a Masters thesis [12]. It has
been the mainstay of our imaging eorts since 1988.
One of the rst results of this instrument was the discovery that in a real
saline phantom tank with real electrodes, the reconstruction algorithm overestimated the conductivity of the saline by as much as 15%. This was because
of the metal electrodes at the periphery, which were not modelled, but which
lowered the voltages by providing alternate current paths. In response, Dave
and his student, Kuo-Sheng Cheng, developed the complete electrode
model [5], which accounted for the conductivity of the electrodes, the gap
between them, and the interface impedance between the electrolyte and the
metallic conductor. This model agreed with the experimental results to
within the accuracy of the data.
The original ACT 1 instrument was designed with a synchronous detectorsensitive only to the real part of the target conductivity. Almost as an
afterthought, we added a quadrature voltmeter, and made a few images of
the reactive component of conductivity. We were pleased to see that aluminium
targets could be distinguished from bright copper targets by the permittivity of
the aluminium oxide layer on the former, although both had similar high
conductivity.
When it became clear that both conductivity and permittivity contained
valuable information, we developed a display and analysis system [38] that
accounted for the interaction between them, rather than simply reconstructing and displaying the results from the real and quadrature voltmeters [14].
In those early years of EIT, guring out what to do was almost as much
of a challenge as actually doing it. Everyones choices were strongly inuenced by their starting assumptions, and it has been interesting to see how
our systems have evolved along with those of the other groups in the eld.
Reconstruction algorithms
13.2.
391
The rst images made by ACT 1 were reconstructed by the NOSER algorithm, mentioned above. This algorithm has a number of properties that
allow it to take advantage of the data obtained by the ACT hardware. In
order to be able to invert the matrix relating voltage to current, the matrix
must be regularized, which has the eect of smoothing the image. This
adds stability and suppresses noise, but at the cost of blurring sharp boundaries in the image. Selection of the appropriate degree of regularization
required an empirical study of typical geometric and electronic noise sources,
and the reconstruction of several images with dierent regularization levels,
to reach a workable compromise. This algorithm in its general form was also
fairly slow, and required a few minutes to reconstruct each image on the SUN
workstation available at that time. The original slow algorithm for circular,
2D geometry has since been extended to incorporate non-circular shapes in
two and three dimensions, and to work in real time.
In 1997, NOSER was expanded to include out-of-round geometries for
the 2D case. Hemant Jain made manual measurements of a subjects chest
and made a reconstruction mesh by hand that ts that geometry. He also
made phantom tanks in elliptical shapes, and reconstructed their images
with various targets in elliptical meshes [38] (gure 13.3).
Another geometrical adaptation was made by Cathy Caldwell, who
wrote a reconstruction algorithm for the case of an array of 16 electrodes,
arranged in a circle within the volume to be imaged and 16 others at the
periphery [A35]. This geometry can be achieved by introducing a catheter
with electrodes into the esophagus to improve the image quality near the
heart. Other applications, for example in urology, may also treat the
unknown volume as an annulus with interior and exterior electrodes.
Table 13.1.
This table summarizes the dierent reconstruction algorithms this group has
developed.
Author
Year
Speed
Dim.
Geometry
Iterations
Approach
Ref.
Simske
Cheney
Goble
Caldwell
Edic
Jain
Edic
Mueller
Blue
Mueller
Choi
1990
1991
1992
1993
1995
1997
1998
1999
2000
2003
2004
slow
slow
slow
slow
fast
slow
fast
slow
fast
slow
slow
2
2
3
2
2
2
2
3
3
2
3
round
round
round
annulus
round
any
round
planar
round
round
two planes
1
NA
1
1
1
1
many
1
1
none
1
indirect,
direct
indirect,
indirect,
indirect,
indirect,
indirect,
indirect,
indirect,
direct
indirect,
12
16
24
A35
32
38
40
42, 43
44
49
A57
linearize
linearize
linearize
linearize
linearize
linearize
linearize
linearize
linearize
392
Figure 13.3. This gure shows the eects of using a reconstruction mesh that closely
approximates the actual shape of the body being studied. On the left is a non-circular
simulated phantom with two inhomogeneities. When the resulting voltage data are used
to reconstruct an image on a circular mesh, the middle gures are obtained. Important
artefacts are observed. On the right are the results of reconstructing the image on a
mesh that approximates the original. The artefacts are not present.
Steve Simske, who wrote the code for the original reconstruction
algorithm, called it NOSER, an acronym for Newtons One-Step Error
Reconstructor. In 1998, Peter Edic wrote and incorporated a forwardsolver algorithm that enabled NOSER to become a multi-step algorithm.
We were pleased and somewhat surprised to learn that allowing more
iterations did not markedly improve the resulting images [40] (gure 13.4).
Margaret Cheney introduced a novel reconstruction algorithm that
makes use of a layer stripping approach to solve the nonlinear inverse
problem directly, rather than forming a small-perturbation linearization as
NOSER and other algorithms do [16, 23]. This algorithm worked well with
simulated data, but was too sensitive to error to be practical with
experimental data.
Reconstruction algorithms
393
Figure 13.4. This is a static image of a saline lled tank with agar phantoms of the heart
and lungs. The actual resistivity values are shown in the bottom-left drawing, and the static
resistivity image is on the right.
Figure 13.5. On the left is a phantom like that in gure 13.4. On the right is a conductivity
image of that phantom, reconstructed using the d-bar algorithm. The conductivity range is
from 185 to 662 mS/m.
394
More recently, Jennifer Mueller and David Isaacson have used scattering theory to develop a direct inversion algorithm called the d-bar method.
This algorithm uses deep ideas from inverse scattering and boundary value
theory, proposed by A. Nachman [45, 47]. An example of its application
to a test phantom is given in gure 13.5. The absolute conductivities reported
Figure 13.6. At the upper left is an empty tank phantom, in which a cubical metal
inhomogeneity (not shown) was suspended at precisely known locations. At the upper
right, the 3D volume in which the conductivity is reconstructed is shown. Below are
images of reconstructed conductivity in slices through each of eight layers below the top
electrode plane. Results are shown for four dierent target depths below the top electrode
layer: (a) 3 mm, (b) 6 mm, (c) 9 mm, and (d) 12 mm. Conductivity scales are dierent
among cases (a)(d).
Hardware
395
by the d-bar algorithm are generally closer to the truth than the NOSER
results.
Our immediate plans are to study breast cancer in a conguration
similar to an x-ray mammogram. Rectangular arrays of electrodes will be
placed on opposite sides of the breastthis requires a reconstruction
algorithm for this geometry. Tzu-Jen Kao and Myoung Hwan Choi have
developed such an algorithm, presently using just 32 electrodes. A test
tank or phantom suitable for this geometry is shown in gure 13.6, along
with one example of the result from the reconstruction algorithm working
from real conductivity data obtained with ACT 3.
13.3.
HARDWARE
We expanded the hardware capability of our system in 1988 with the introduction of ACT 2 (gure 13.7), a 64-electrode system built with considerable
help from the Corporate Research and Development Center of GE [4]. This
system was built on eight double-sided circuit boards with eight channels
each. It could obtain the data for a 32-electrode image in a few seconds, a
signicant improvement over ACT 1 (see table 13.2). Its other characteristics
were similar.
Shortly thereafter, we began the design of ACT 3, a signicantly faster
and more accurate instrument (gure 13.8). It is a property of impedance
imaging systems that if any region of the eld changes during the acquisition
Figure 13.7. This is ACT 2. It contains eight boards with eight current sources on each.
The real and quadrature voltmeters are upper right, above the power supply. The ribbon
cable connects to the data acquisition card in the supporting computer. Construction is
two-sided printed circuit boards.
396
Table 13.2. A summary of the technical characteristics of the hardware systems we have
developed.
System
Architecture
Construction
Frequency
(kHz)
Imax ma,
peak
Frame rate
(frames/s)
In service
ACT I
ACT II
ACT III
32 ch/bd 1 bd
8 ch/bd 8 bd
1 ch/bd 32 bd
Wire wrap
2-sided boards
2-sided boards
12
15
28.8
5.0
5.0
0.85
1/30
1/5
20
ACT IV
8 ch/bd 8 bd
12-layer boards
0.51000
various
20
1987
1988
Slow, 1991
Fast, 1993
2004
of the data, all parts of the image are degraded. This was the motive for
designing ACT 3 to acquire data in a much shorter time, for use in imaging
the chest. We wanted the aperture time for an image to be a small fraction of
a cardiac cycle [30]. This was achieved by the rst version of the instrument,
but it was not able to reconstruct or display these data rapidly. Another
major change was to reconstruct and display data in real time [32]. ACT 3
also incorporated a high speed A/D converter, operating in an oversampling/undersampling mode to achieve high accuracy and high speed
Figure 13.8. This is ACT 3. Each of 32 electrodes is connected to a circuit board. There
are 32 such boards in two rows of 16. Only the front edge of each board is visible. The
instrument is controlled from the keyboard and rear monitorthe monitor displays the
images in real time. Construction is two-sided printed circuit boards.
Hardware
397
with high rejection of noise outside its narrow frequency bandwidth. The
current sources were also designed to have very high output impedance.
This is necessary because the load impedances for dierent electrodes can
be very dierent, and if the output impedance of a current source is not
high, some of the current it produces does not go to the load as desired. A
high output impedance is obtained in ACT 3 by adjusting a negative capacitance circuit and an output resistance circuit using digitally controlled potentiometers. Each channel can be connected to a calibrating circuit which
measures its output impedance. The digital potentiometers are then adjusted
iteratively to attain an output impedance above 10 M
, with an output
capacitance below 0.5 pF.
In 1998, we began the design of an instrument for breast cancer detection, based on a commercial data acquisition board. The manufacturer of
this board made some assertions about its capabilities that turned out not
to be true, and we wasted a lot of eort on a system that ultimately failed.
We then began the design of ACT 4, a faster, multi-frequency, 64-electrode system designed for breast imaging [50] (gure 13.9). This machine is
being built at the time of writing, having been simulated in software and
partly prototyped. Its technical characteristics are summarized with those
of its predecessors. Its major technical characteristic is its exibility; by
using programmable digital signal processors and eld programmable gate
Figure 13.9. This is ACT 4 at the time of writing. Modular design and construction uses
eight and 12 layer circuit boards in surface mount technology.
398
arrays, this instrument can be tailored to many data acquisition and image
display schemes. This single instrument contains a current source and a
separate voltage source for each electrode. The current sources are adjustable
using digital potentiometers, so as to have very high output impedance at
each of 610 operating frequencies over a wide spectrum. The intent is to
compare the quality of the data achievable with these sources, with the
signal-to-noise ratio achievable with voltage sources adjusted in software
to provide desired current levels. A complicated automatic calibration
scheme is used for the voltage sources, and their high precision may allow
a comparable overall system signal-to-noise ratio for the current and voltage
sources.
13.4.
APPLIED CURRENTS
Optimal currents
399
noise) from the skin. For these reasons, the eects of the skin are eliminated
or greatly reduced.
There is a rationale for the approach we have adopted. Spatial noise
introduced by, for example, errors in electrode placement or dierences in
electrode impedance, occurs at high spatial frequency. In systems which
apply currents, these artefacts are minimized by applying patterns with low
spatial frequency. They are exaggerated by patterns with high spatial
frequency. Patterns applying current between pairs of electrodes contain
high energy at high spatial frequency, and less energy at low frequencies.
There is, therefore, a noise-reducing eect of applying low-frequency current
patterns.
13.5.
OPTIMAL CURRENTS
Figure 13.10. This is a static image of a simulated thorax with realistic geometry and electrical properties. The top image is the phantom simulated. The middle image shows an
FEM reconstruction of the resistivities, using the canonical trigonometric current patterns.
The bottom image shows the increased contrast obtained when the current patterns have
been optimized by eight iterations of the optimizing algorithm.
400
patterns is just below it. We then applied the iterative optimal current
algorithm for eight iterations, with the result shown at the bottom of
gure 13.10. Clearly, the contrast and dynamic range of the reconstructed
resistivities are closer to the simulated values when optimal currents are used.
13.6.
When optimal currents are used in vivo, the number of iterations should be
limited to 24 because of the variations in the actual data due to cardiac
and ventilatory events. Figure 13.11 shows the rst four iterations of the
current optimizing algorithm, producing static images of a non-circular
chest. The contrast of the high-resistivity skin at the periphery and the central
lungs improves with each iteration.
Figure 13.11. These images are reconstructed from data obtained from a subject whose
chest had the shape shown. The reconstruction algorithm used a mesh adapted to this
shape. The four images show the result of using current patterns that approach the optimal
patterns. Note the range of conductivities displayed with each image, indicated by the
numbers above the grey scale. The original image with the canonical trigonometric
patterns has a range from 242 to 608 mS/m. After three iterations of the current-optimizing
algorithm, the image reconstructed from the data obtained with the new currents has a
range from 121 to 1477 mS/m.
13.7.
3D
In vivo applications
401
Figure 13.12. Four rows of eight electrodes each were applied to a subjects chest. A 3D
reconstruction algorithm was used to form a static image of the relatively conductive heart
(light grey) and less conductive lungs (dark grey). Two views of the reconstructed image are
shown, from above and in front of the subject (bottom left), and from above but behind the
subjects right side (bottom right).
13.8.
IN VIVO APPLICATIONS
We have conducted and published several studies in living subjects [20, 35,
43, 44, 48]. In a 1996 investigation of acute pulmonary edema in dogs, we
demonstrated the ability of the ACT 3 system to monitor the development
of acute pulmonary edema, induced by intravenous infusion of oleic acid
[35]. Changes in impedance images were correlated with post-mortem assessment of lung water. We also studied several acutely ill patients in a surgical
intensive care unit in 1993. These were early studies using ACT 3, which
conrmed our ability to use it in the ICU with minimal interference to clinical
routines. We detected a case of tension pneumothorax in one patient, which
was conrmed a few hours later by x-ray. These studies were of an exploratory nature, and they taught us a lot about how to use the system, but did not
yield publishable results. Three years later we studied a few more patients in a
coronary care unit, and related impedance changes to x-ray appearance of
pulmonary edema. A general correlation was found, and valuable experience
402
Figure 13.13.
Paying for it
403
13.9.
PAYING FOR IT
This work has been funded by the US taxpayers, and a few private sources.
Dave Isaacson got the rst National Science Foundation grant in 1987 for
two years support, and we got a follow-up grant from the National Institutes
of Health in 1988 for three years.
What happened next involved my guardian angel. In the mid-1980s,
when this work was getting started, I had been involved for many years
with a large-scale projectfunded by the National Institutes of Healthto
study trauma. When that project was competitively renewed in 1988, we
included an EIT proposal which was very favourably reviewed, but the
404
13.10.
PEOPLE
This project started with David Isaacsons work in 1985. When his rst paper
was nearly completed, he asked me to do some simple measurements of noise
levels, to illustrate what might be achievable in the real world. We recruited
an undergraduate student, Denise Angwin, who spent a summer getting data
from a saline-lled tank with copper electrodes driven by a Radio Shack
audio amplier (see gure 13.1). This gave some useful results, but took
too long. David Gisser, a senior Professor in Electrical Engineering was
well known to me from a couple of decades of collaboration in the trauma
research project. He joined us in 1986, and designed ACT 1, a system with
32 computer-controlled current sources and a multiplexed voltmeter. Results
from this system were encouraging, we decided we needed a faster system,
and began the design of ACT 2. By early 1988 that machine was in service,
and producing encouraging results. As we started the design of ACT 3,
around 1989, Gary Saulnierof the Electrical, Computer and Systems
Meetings
405
13.11.
MEETINGS
Most of the work done in the early years of impedance imaging was in
Europe, with major support from the European Community through a
406
Concerted Action in Impedance Tomography. Thanks to a liberal interpretation of the term Europe by Brian Brown who ran that programme,
we have participated in all the major meetings of that group. David
Holder also obtained EPSRC support for meetings in London, which we
have also joined with pleasure. Since 1995, there has been closer collaboration between those working in impedance imaging and the International
Conference on Electrical Bio-Impedance. These ties were greatly strengthened by Dr Eberhard Gersing, who organized a joint meeting of CAIT
and ICEBI in Heidelberg in 1995. Subsequent meetings in Barcelona and
Oslo have further developed these collaborations.
In working closely together, dierences in personality and style can
present challenges to everyone. In 1987, Dave and I travelled together to
Lyon for the CAIT Conference. But we almost didnt get there. We planned
to take the train to New York one morning, and y to Paris that evening. We
were to meet at the station. The train arrived on time, and several dozen other
passengers and I boarded. No sign of David. I hurried through the whole
train, searching. No David. As the scheduled departure time approached,
the platform was deserted, and the train doors were closed except for one,
just behind the engine. There I stood with the conductor, waiting. With a
minute to go, David appeared, and, seeing no activity on the platform, set
his suitcase down against a wall and sat on it to wait. I shouted, he walked
over, boarded the train, the conductor waved to the engineer, and we were
o to Lyon.
13.12.
CONCLUDING REMARKS
We have been working with this technology for around 18 years, as of June
2004, and perhaps it is appropriate to look back and look ahead with a longer
term view. In retrospect, I think we have been well served by the use of
multiple current sources, and the use of all available voltage measurements.
Our progress has been slowed by the technical challenges of the analogue
circuits required, but the basic EIT problem is dicult and ill-posed, and
requires the highest quality data that can be obtained if one is to draw
rm conclusions about its use.
At the time of writing, I can see some areas where I wish we had made
dierent decisions about the latest system, ACT 4. It is designed to have
many desirable features in a single instrument. It has both current and
voltage sources, available over a wide frequency range on 64 electrodes in
a small package operating at high speed. The development of this system
has been slowed, and made more expensive by our decision to use very
small circuit boards with high component density. A lot could be learned
without using as many as 64 electrodes. If tissue spectroscopy of the breast
is useful, we could improve spatial resolution by expanding a smaller
Complete Bibliography
407
408
Complete Bibliography
409
410
SELECTED ABSTRACTS
[A34] Newell J C, Edic P M, Saulnier G J, Isaacson D, Cheney M and Gisser D G 1993
Real-time adaptive current tomography, in Proc. European Community Concerted
Action on Impedance Tomography, Barcelona, Spain, 2225 September, pp 2930
[A35] Caldwell C, Cheney M and Isaacson D 1993 Impedance imaging using interior and
exterior measurements, in Physiological Imaging, Spectroscopy and Early-Detection
Diagnostic Methods ed R L Barbour and M J Carvlin. SPIE Proceedings series vol
1887
[A38] Newell J C, Isaacson D, Saulnier G J, Cheney M, Gisser D G, Edic P M, Ren X and
Larson-Wiseman J L 1994 Electrical impedance imaging of thoracic admittivity in
normal man, in Proc. World Congress on Medical Physics and Biomedical Engineering, Rio de Janiero, Brazil, August, p 604, OS22-2.1
[A48] Newell J C, Blue R S, Isaacson D, Saulnier G J and Ross A S 2001 Phasic threedimensional impedance imaging of cardiac activity, in Proc 3rd EPSRC Conf.,
London, April
[A57] Choi M H, Kao T-J, Isaacson D, Saulnier G J and Newell J C 2004 A simplied
model of a mammography geometry for breast cancer imaging with electrical
impedance tomography, in Proc. IEEE-EMBS Conf. 26, in press, #592, 2.4.2
Appendix A
Brief introduction to bioimpedance
David Holder
A.1.
The resistance and the capacitance of tissue are the two basic properties in
bioimpedance.
Resistance is a measure of the extent to which an element opposes the
ow of electrons or, in aqueous solution as in living tissue, the ow of ions
among its cells. The three fundamental properties governing the ow of electricity are voltage, current and resistance. The voltage may be thought of as
the pressure exerted on a stream of charged particles to move down a wire or
migrate through an ionized salt solution. This is analogous to the pressure in
water owing along a pipe. The current is the amount of charge owing per
unit time, and is analogous to water ow in a pipe. Resistance is the ease or
diculty with which the charged particles can ow, and is analogous to the
width of a pipe through which water owsthe resistance is higher if the pipe
is narrower (gure A.1).
They are related by Ohms law:
V (voltage, Volts) I (current, Amps) R (resistance, Ohms
:
The above applies to steadily owing, or d.c. current (direct current).
Current may also ow backwards and forwardsa.c. (alternating current).
412
Figure A.1.
Resistance has the same eect on a.c. current as d.c. current. Capacitance (C)
is an expression of the extent to which an electronic component, circuit or
system, stores and releases energy as the current and voltage uctuate with
each a.c. cycle. The capacitance physically corresponds to the ability of
plates in a capacitor to store charge. With each cycle, charges accumulate
and then discharge. Direct current cannot pass through a capacitor. A.c.
can pass because of the rapidly reversing ux of charge. The capacitance is
an unvarying property of a capacitive or more complex circuit. However,
the eect in terms of the ease of current passage depends on the frequency
of the applied currentcharges pass backwards and forwards more rapidly
if the applied frequency is higher.
For the purposes of bioimpedance, a useful concept for current travelling through a capacitance is reactance (X). The reactance is analogous to
resistancea higher reactance has a higher eective resistance to alternating
current. Like resistance, its value is in Ohms, but it depends on the applied
frequency, which should be specied (gure A.2).
The relationship is
Reactance (Ohms) 1=2 Frequency Hz Capacitance Farads:
When a current is passing through a purely resistive circuit, the voltage
recorded across the resistor will coincide exactly with the timing, or phase,
of the applied alternating current, as one would expect. In the water ow
analogy, an increase in pressure across a narrowing will be instantly followed
by an increase in ow. When current ows across a capacitor, the voltage
recorded across it lags behind the applied current. This is because the back
and forth ow of current depends on repeated charging and discharging of
the plates of the capacitor. This takes a little time to develop. To pursue the
water analogy, a capacitor would be equivalent to a taut membrane stretched
Figure A.2.
413
414
Figure A.3. The voltage that results from an applied current is in phase for a resistor (A)
and 908 out of phase for a capacitor.
Figure A.4.
415
416
Figure A.5. (a) The cell modelled as basic electronic circuit. Ri and Re are the resistances
of the intracellular- and extracellular-space, and Cm is the membrane capacitance. (b)
ColeCole plot of this circuit.
417
Figure A.6. (a) The movement of current through cells at both low and high frequencies.
(b) Idealized ColeCole plot for tissue.
418
occurs is known as the centre frequency (Fc), and is a useful measure of the
properties of an impedance. In real tissue, the ColeCole plot is not exactly
semicircular, because the detailed situation is clearly much more complex;
the plot is usually approximately semicircular, but the centre of the circle
lies below the x-axis. Inspection of the ColeCole plot yields the high- and
low-frequency resistances, as the intercept with the x-axis, and the centre
frequency is the point at which the phase angle is greatest. The angle of depression of the centre of the semicircle is another means of characterizing the tissue
(gure A.6(b)).
Over the frequency ranges used for EIT and MIT, about 100 Hz to
100 MHz, the resistance and reactance of tissue gradually decreases. This is
due to the simple eect of increased frequency passing more easily across capacitance, but also because cellular and biochemical mechanisms begin to operate, which increases the ease of passage of the electrical current. A
remarkable feature of live tissue is an extraordinarily high capacitance, which
is up to 1000 times greater than inorganic materials, such as plastics used in
capacitors. This is because capacitance is provided by the numerous and closely
opposed cell membranes of cells, each of which behaves as a tiny capacitor.
Over this frequency range, there are certain frequency bands where the phase
angle increases, because mechanisms come into play which provide more capacitance. They may be seen as regions of an increased decrease of resistance in a
plot of resistance against frequency, and are termed dispersions. At the low
end of the frequency spectrum, the outer cell membrane of most cells is able
to charge and discharge fully. This region is known as the alpha dispersion
and is usually centred at about 100 Hz.
As the frequency increases, from 10 kHz10 MHz, the membrane only
partially charges and the current charges the small intracellular space structures, which behave largely as capacitances. At these higher frequencies the
current can ow through the lipid cell membranes, introducing a capacitive
component. This makes the higher frequencies sensitive to intracellular
changes due to structural relaxation. This eect is largest around 100 kHz,
and is termed the beta dispersion. At the highest frequencies, dipolar
reorientation of proteins and organelles can occur, and aect the impedance
measurements of extra- and intracellular environments. This is the gamma
dispersion, and is due to the relaxation of water molecules and is centred
at 10 GHz. Most changes between normal and pathological tissues occur
in the alpha and beta dispersion spectra.
A.3. OTHER RELATED MEASURES OF IMPEDANCE
A.3.1.
Resistance and reactance, as described above, are xed measures of individual components or samples. It is useful to be able to describe the general
Figure A.7.
measured.
419
The eect of changing the length or cross-sectional area of the tissue sample
The resistance and reactance fully describe the impedance of tissue, but there
are several other related measures which are, sometimes confusingly, used in
the EIT and Bioimpedance literature. These arise because dierent, reciprocal, terms may be used to describe the ease, as opposed to the diculty, of
passage of current. Secondly, with respect to capacitance, one can choose
to use the eective resistance at a given frequencythe reactance, or the
intrinsic property of the material, the capacitance, which is independent of
frequency. Each of the dierent measures may be suxed with -ivity to
yield its general property. Finally, the measure given may refer to the
complex impedance rather than the in- or out-of-phase component.
Not all permutations, fortunately, are widely used. These are the most
common: the conductivity is the inverse of resistivity and is given in
Siemens/m or, more usually, S/m. The admittance is the inverse of
impedance, and so is a combined measure of in- and out-of-phase ease of
420
Figure A.8. (a) The two-electrode measurement as a block diagram, and (b) modelled as a
simple electrical circuit. The two overlapping rings represent a constant current electrical
source.
421
Figure A.9. The four-electrode measurement as (a) a block diagram, and (b) modelled as
a simple electrical circuit.
A.5.
The Sheeld mark 1, and the several similar systems which have been used to
make clinical and human EIT measurements, only record the in-phase,
resistive, component of the impedance. This is because unwanted capacitance
in the leads and electronics introduce errors. Fortunately, these are all out-ofphase and so can be largely discounted by throwing away the out-of-phase
data. For the same reason, images are generated of dierences over time,
as subtraction like this minimizes errors. As a result, the great majority of
clinical EIT images are a unitless ratio between the reference and test
image data at a single frequency. More recently, systems have been
constructed and tested which can measure at multiple frequencies, and
provide absolute impedance data. As these are validated, and come into
wider clinical use, then we may expect to see more absolute bioimpedance
parameters, such as resistivity, admittivity, centre frequency, or ratio of
extra- to intracellular resistivity, in EIT image data.
422
FURTHER READING
Duck F A 1990 Physical properties of tissue: a comprehensive reference book. London:
Academic Press
Gabriel C, Gabriel S and Corthout E 1996a The dielectric properties of biological tissues: I.
Literature survey Physics in Medicine and Biology 41 22312249
Gabriel S, Lau R and Gabriel C 1996b The dielectric properties of biological tissues: II.
Measurements in the frequency range 10 Hz to 20 GHz Physics in Medicine and Biology 41 22512269
Geddes L B L E 1967 The specic resistance of biological materiala compendium of data
for the biomedical engineer and physiologist Med. Biol. Eng. 5 271293
Grimnes S and Martinsen G 2000 Bioimpedance and bioelectricity basics. London:
Academic Press
Appendix B
Introduction to biomedical electrical
impedance tomography
David Holder
One of the attractions but also diculties of biomedical EIT is that it is interdisciplinary. Topics which are second nature to one discipline may be incomprehensible to those with other backgrounds. Not all readers will be able to
follow all the chapters in this book, but I hope that the majority will be
comprehensible to most, especially those with a medical physics or bioengineering background. Nevertheless, the reconstruction algorithm or
instrumentation chapters may be dicult to follow for clinical readers, and
some of the clinical terminology and concepts in application chapters may
be unfamiliar to readers with Maths or Physics backgrounds. This chapter
is intended as a brief and non-technical introduction to biomedical electrical
impedance tomography. It is didactic and explanatory, so that the more
detailed chapters in the book which follow may be easier to follow for the
general reader. It is intended to be comprehensible to readers with clinical
or life sciences backgrounds, but with the equivalent of high school physics.
A non-technical introduction to the basics of bioimpedance is presented in
Appendix A, and may be helpful for any reader wishing to refresh their
understanding of the basics of electricity and its ow through biological
tissues. As it is intended to be explanatory, key references and suggestions
for further reading are included, but the reader is recommended to the
detailed chapters in the main body of the book for detailed citations.
B.1.
HISTORICAL PERSPECTIVE
The rst published impedance images appear to have been those of Henderson
and Webster in 1976 and 1978 (Henderson and Webster 1978). Using a rectangular array of 100 electrodes on one side of the chest earthed with a single large
electrode on the other side, they were able to produce a transmission image of
424
the tissues. Low conductivity areas in the image were claimed to correspond to
the lungs. Shortly after, an impedance tomography system for imaging brain
tumours was proposed by Benabid et al (1978). They reported a prototype
impedance scanner which had two parallel arrays of electrodes immersed in
a saline lled tank, and which was able to detect an impedance change inserted
between the electrode arrays.
The rst clinical impedance tomography system, then called applied
potential tomography (APT), was developed by Brian Brown and David
Barber and colleagues in the Department of Medical Physics in Sheeld.
They produced a celebrated commercially available prototype, the Sheeld
Mark 1 system (Brown and Seagar 1987), which has been widely used for
performing clinical studies, and is still in use in many centres today. This
system made multiple impedance measurements of an object by a ring of
16 electrodes placed around the surface of the object.
The rst published tomographic images were from this group in 1982 and
1983. They showed images of the arm in which areas of increased resistance
roughly corresponded to the bones and fat. As EIT was developed, images
of gastric emptying, the cardiac cycle and the lung ventilation cycle in the
thorax were obtained and published. The Sheeld EIT system had the advantage that 10 images/s could be obtained, the system was portable, and the
system was relatively inexpensive compared to ultrasound, CT and MRI
scanners. However, since the EIT images obtained were of low resolution
compared to other clinical techniques such as cardiac ultrasound and x-ray
contrast studies of the gut, EIT did not gain widespread clinical acceptance
(see Holder 1993, Boone et al 1997, Brown, 2003, for reviews).
Around the same time, a group in Oxford proposed that EIT could be
used to image the neonatal brain (Murphy et al 1987). They developed a
clinical EIT system and obtained preliminary EIT images in two neonates.
Their system used 16 electrodes placed in a ring around the head, but in
contrast to the Sheeld system, the current was applied to the head by
pairs of electrodes which opposed each other in the ring in a polar drive
conguration. This maximized the amount of current which entered the
brain and therefore maximized the sensitivity of the EIT system to impedance
changes in the brain.
Since the rst ush of interest in the mid to late 1980s, about a dozen
groups have developed their own EIT systems and reconstruction software,
and publications on development and clinical applications have been produced
by perhaps another twenty or so. Initial interest in a wide range of applications
at rst has now settled into the main areas of imaging lung ventilation, cardiac
function, gastric emptying, brain function and pathology, and screening for
breast cancer. Convincing pilot and proof of principle studies have been
performed in these areas. In 1999, FDA approval was given to a method of
impedance scanning to detect breast cancer, and the system has been marketed
commercially (http://imaginis.com/t-scan/eectiveness.asp), but it is not yet
EIT instrumentation
425
clear how widely it is being used. In other areas, EIT has not yet broken into
routine clinical use.
B.2.
EIT INSTRUMENTATION
EIT systems are generally about the size of a video recorder, but some may be
larger. They usually comprise a box of electronics and a PC. Connection to
the subject is usually made by coaxial cables a metre or two long, and ECG
type electrodes are placed in a ring or rings on the body part of interest. All
will sit on a movable trolley, so that recording can be made in a clinic or outpatient department. A typical system is shown in gure B.1.
B.2.1.
A single impedance measurement forms the basis of the data set which is used
to reconstruct an image. Most systems use a four-electrode method, in which
constant current is applied to two electrodes, and the resulting voltage is
recorded at two others. This minimizes the errors due to electrode
impedance. The transfer impedance of the subject with this recording geometry is calculated using Ohms law (gure B.2). The current applied is
approximately one tenth of the threshold for causing sensation on the
skin. It is insensible and has no known ill eects. Most single frequency
systems apply a current at about 50 kHz. At this frequency, the properties
of tissue are similar to those at d.c., in that the great majority of current
travels in the extracellular space, but electrode impedance is much lower
than at d.c., so there are less instrumentation errors. At 50 kHz, a single
measurement usually takes less than 1 msec.
Figure B.1.
426
Figure B.2.
EIT instrumentation
427
Figure B.3. Essential components of an EIT system. The system shown is for a single impedance measuring circuit with connection to electrodes
using a multiplexer. More complex systems may have multiple circuits attached directly to electrode pairs. The demodulator converts the a.c.
recorded signal into a steady d.c. voltage for both resistance and reactance, although the reactance signal is discarded in many systems as the
stray capacitance renders it inaccurate. The subject and electrode impedances (R(e)) are represented as resistances.
428
Figure B.4. Sources of error in impedance measurements. There are two main sources of
error. (1) A voltage divider exists, formed by the series impedance of the skin and input
impedance of the recording instrumentation amplier. Under ideal circumstances, the
skin impedance is negligible compared to the input impedance of the amplier, so that
the voltage is very accurately recorded (upper example). In this example, skin impedance
is 100 kOhms and input impedance is 100 MOhms, so the loss of signal is negligible. In
practice, the stray capacitance in the leads, coupled to high skin impedances, may cause
a signicant attenuation of the voltage recordede.g. to 90%, if the input impedance
reduces to 1 MOhm (lower example). In this diagram, only one side of a dierential amplier is shown, for clarity. This attenuating eect may be dierent for the two sides of the
amplier. This leads to a loss of common mode rejection ability, as well as absolute
errors in the amplitude recorded. (2) The ideal current source is perfectly balanced, so
that all current injected leaves by the sink of the circuit. The eect of stray capacitance
and skin impedance may act to unbalance the current source. Some current then nds
its way to ground, either by the ground, or by the high input impedance of the recording
circuit. This causes a large common mode error. The common mode rejection ratio may be
poor because of the eects in (1), so that the recorded voltage is inaccurate.
common mode errors on the recording side due to impaired common mode
rejection as a result of stray capacitance (see Boone and Holder 1996 for a
review) (gure B.4).
B.2.2.
Data collection
EIT instrumentation
429
Figure B.5. Data acquisition with the Sheeld Mark 1 system. A constant current is
injected into the region between two adjacent electrodes, and the potential dierences
between all other pairs of adjacent electrodes are measured. The current drive is then
moved to the next pair of adjacent electrodes, and the measurements repeated and so on
for all possible current drive pairs. It is not possible to measure potential dierences accurately at the pair of electrodes injecting current, so there are 208 (13 16) measurements in
a data set.
430
Figure B.6. Miniature Sheeld Mark 1 APT system designed for the Juno space mission
(courtesy of Prof. B. Brown).
Figure B.7. UCLH Mark 1 EIT system, intended for ambulatory recording in subjects
being monitored on a ward for epileptic seizures. A small headbox is on a lead 10 m
long, so that the subject may walk around near their bed during recording.
EIT instrumentation
431
Electrodes
The great majority of clinical measurements have been made with ECG type
adhesive electrodes attached to the chest or abdomen (gure B.1). Although
the four-electrode recording system should in theory be immune to electrodeskin impedance, in practice it is usually necessary to rst reduce the skin
impedance by abrasion. Similar EEG cup electrodes have been used for
head recording.
In the mid 1980s convenient exible electrode arrays were designed and
reported for chest imaging, but did not become commercially available, so
now most groups use ECG or EEG electrodes (McAdams et al 1994).
Some specialized designs have been developed for the special case of imaging
the breastprecise positioning may be achieved by radially movable motorized rods arranged in a circle (gure B.8).
B.2.4.
432
Figure B.8.
USA.)
reciprocity, or to assume that the subject is primarily resistive at low frequencies, and adjust the phase detection accordingly (Fitzgerald et al 2002).
As many EIT systems are prototypes, it is helpful to calibrate them on
known test objects. Some employ agar test objects, impregnated with a
saline solution, in a larger tank which contains saline of a dierent concentration. These can be accurate if images are made quickly, but the saline
will diuse into the bathing solution, so that the boundaries can become
uncertain (Cook et al 1994). Others have employed a porous test object
such as a sponge, immersed in the bathing solution in a tank, so that the
impedance contrast is produced by the presence of the insulator in the test
object (Holder et al 1996a). Many tanks have been cylindrical; more realistic
ones have simulated anatomy, such as the head (Tidswell et al 2003), or used
biological materials to produce multifrequency test objects. Typically, the
spatial resolution of test objects in tanks is about 15% of the image diameter
(gure B.9).
B.2.5.
Most EIT work has used EIT as a dynamic imaging method, in which images
of the impedance change compared to a baseline condition are obtained. An
EIT instrumentation
433
Figure B.9. Example of image quality with a modern multifrequency EIT system from
Dartmouth, USA. (Courtesy of Prof. A. Hartov.)
Figure B.10. Example of EIT of gastric emptying, collected with the Sheeld Mark 1 EIT
system, and 16 electrodes placed around the abdomen.
434
Figure B.11. Example of cardiac imaging, collected with the Sheeld Mark 1 EIT system,
and a ring of 16 electrodes placed around the chest.
between the model of the body part used in the reconstruction software and
the actual object imaged. To reduce these, impedance changes are reconstructed with reference to a baseline condition; if the electrode placement
errors in the baseline images and the impedance change images are the
same, then these errors largely cancel if only impedance change is
imaged. Although the dynamic imaging approach minimizes reconstruction
errors, it limits the application of EIT to experiments in which an
impedance change occurs over a short experimental time course; otherwise,
electrode impedance drift may introduce artefacts in the data which cannot
be predicted from the baseline condition. As dynamic imaging cannot be
used to image objects present at the start of imaging and therefore in the
baseline images, dynamic EIT cannot be used to obtain images of tumours
or cysts. This contrasts with images obtained with CT, which can obtain
static images of contrasting tissues such as tumours. Dynamic imaging
has been used for almost all clinical studies to date in all areas of the
body.
In principle, it should be possible to produce images of the absolute
impedance. Unfortunately, image production is sensitive to errors in instrumentation and between the model used in reconstruction and the object
imaged. Pilot data has been obtained in tanks (Cook et al 1994) and some
preliminary images in human subjects (Cherepenin et al 2002, Soni et al
2004).
Dynamic EIT images typically use one measurement frequency, usually
between 10 and 50 kHz, to make impedance measurements. An alternative
approach is to compare the dierence between impedance images measured
at dierent measurement frequencies, termed EITS (EIT spectroscopy).
This technique exploits the dierent impedance characteristics of tissues
435
B.3.
B.3.1.
Back-projection has been very successful for the simple case of 16 electrodes
in a plane, but suered from the need for two assumptionsthat the problem
was 2D, and that the initial resistivity was uniform. Most systems now
employ a more powerful method, based on a sensitivity matrix (gure
B.13). This is based on a matrix, or table, which relates the resistivity of
each voxel in the subject, and hence, images, to the recorded voltage
measurements.
436
Figure B.12.
437
(a)
(b)
Figure B.13. Explanation of sensitivity matrix. (a) The sensitivity matrix. This is shown
guratively for a subject with four voxels and four electrode combinations. Each
column represents the resistivity of one voxel in the subject. Each row represents the
voltage measured for one electrode combination. The current from one current source
ows throughout the subject, but the voltage electrodes are most sensitive to a particular
volume, shown in grey. The resulting voltage is a sum of the resistivity in each of the voxels
weighted by the factor S for each voxel, which indicates how much eect that voxel has on
the total voltage. (b) The forward case. In a computer program, all the sensitivity factors
are calculated in advance. Given all the resistivities for each voxel, the voltages from each
electrode combination are easy to calculate. (c) The inverse. For EIT imaging, the reverse is
the casethe voltages are known; the goal is to calculate all the voxel resistivities. This can
be achieved by inverting the matrix. This is straightforward for the simple case of four
unknowns shown here, but is not in a real imaging problem, where the voltages are
noisy, and there may be many more unknown voxels than voltages measured.
438
(c)
Figure B.13.
(Continued)
anatomical meshes need to contain many more cells than a few hundred,
especially if in 3D, so the matrix may contain tens of thousands of
columnsone for each voxeland a few hundred rows. If the resistivities
of each voxel are given, then the expected voltages for each electrode combination may be easily calculated. This is termed the forward solution and is
simply a simulation of the situation in reality (gure B.13(b)). Its use is to
generate a sensitivity matrix. This is produced by, in a computer simulation,
varying resistivity in each voxel, and recording the eect on dierent voltage
recordings. This enables calculation of the sensitivity of a particular voltage
recording to resistance change in a voxelthe s factor in gure B.13.
To produce an image, it is necessary to reverse the forward solution. On
collecting an image data set, the voltages for each electrode combination are
known, and, by generating the sensitivity matrix, so is the factor relating
each resistance to these. The unknown is the resistivity in each voxel. This is
achieved by mathematically inverting the matrixwhich yields all the
resistivities (gure B.13(c)). In principle, this can give a completely accurate
answer, but this is only the case if the data is innitely accurate, and that
there are the same number of unknownsi.e. voxels requiring resistance
estimates, as electrode combinations. In general, none of these is true. In
particular, in many of the voxels, very little current passes through, so the
sensitivity factor for that cell in the table is near to zero. Just as dividing by
zero is impossible, dividing by such very small numbers causes instabilities
in the image. This is termed an ill-posed matrix inversion. There is a well
established branch of mathematics which deals with these inverse problems,
and matrix inversion is made possible by regularizing the matrix. In principle,
this is performed by undertaking a noise analysis of the datanoisy channels
with little signal-to-noise are suppressed, so that the image production by
Clinical applications
439
B.4.
CLINICAL APPLICATIONS
B.4.1.
B.4.1.1.
The great majority of clinical studies have been performed with the Sheeld
Mark 1 system, so that most published studies of accuracy have mainly been
440
(a)
(b)
Figure B.14. (a) Calibration studies with the Sheeld Mark 1 system in a saline lled
tank. The tank was lled with saline, which was varied to give dierent contrasts with
the test object of a cucumber. The cucumber may be seen in the correct location for all
contrasts, but with more accuracy and greater change near the edge (Holder et al 1996).
(b) Images taken with 3D linear algorithm in a latex head-shaped tank, with or without
the skull in place. The algorithm employed a geometrically accurate nite element mesh
of the skull and tank (Bagshaw et al 2003).
with this prototype system. In saline lled tanks, the Sheeld Mark 1, with
its 16 electrodes and back-projection algorithm, produces somewhat blurred
but reproducible images (gure B.14(a)). In general, the spatial accuracy is
about 15% of the image diameter, being 12% at the edge and 20% in the
centre (see Holder 1993 for a review). More recent studies with more
Clinical applications
441
advanced systems, including those in 3D in the thorax and head, are roughly
similar (Metherall et al 1996, Bagshaw et al 2003b) (gures B.9, B.14(b)). In
general, in human images where the underlying physiological change is well
described, such as gastric emptying (Mangall et al 1987), lung ventilation
(Barbas et al 2003), lung blood ow (Smit et al 2003), or cardiac output
(Vonk et al 1996), images have a similar resolution with mild blurring, but
the anatomical structures can be identied with reasonable condence. In
the more challenging areas such as imaging breast cancer (Soni et al 2004),
or evoked activity or epileptic seizures in the brain (Tidswell et al 2001,
Bagshaw et al 2003a), some individual images appear to correspond to the
known anatomy, but these are not suciently consistent across subjects to
be used condently in a clinical environment.
B.4.1.2.
Variability
442
The clinical application which has received the greatest interest has been
imaging of lung ventilation and cardiac output (Frerichs 2000). Large
impedance changes occur during ventilation, as air enters and leaves the
Clinical applications
443
lungs. Although the images have a relatively low resolution, several pilot
studies have conrmed that reasonably accurate data concerning ventilation
can be continuously obtained at the bedside (Harris et al 1988, Kunst et al
1998). EIT therefore has the potential to image ventilation. Although the
feasibility of imaging this with the Sheeld Mark 1 system was established
in the 1980s, the method has not yet been taken up into clinical use. This is
presumably because good imaging methods already exist for assessing lung
function and pathology, and the portability of EIT was not considered
sucient to outweigh relatively poor spatial resolution. However, recently,
there has been fresh interest in this application, led by Amato and colleagues
(Kunst et al 1998, Barbas et al 2003, Hinz et al 2003, Victorino et al 2004). In
operating theatres or Intensive Care Units, there is a growing body of thought
that, in ventilated patients, the outcome is improved if ventilation is adjusted
so that no regions of lung stay collapsed; EIT is suciently small and rapid to
enable continuous monitoring at the bedside to achieve this.
Pilot studies have also shown that EIT has reasonable accuracy in
imaging in emphysema (Eyuboglu et al 1995), pulmonary oedoema (Noble
et al 1999), lung perfusion with gating of recording to the ECG (Smit et al
2003), and perfusion during pulmonary hypertension (Smit et al 2002).
However, although of physiological interest, these applications have not
yet been taken up as being suciently accurate for clinical use.
All the above studies have employed the Sheeld Mark 1 or similar 2D
systems with a single ring of electrodes; it appears that this gives sucient
resolution to enable optimization of ventilator settings when compared to
concurrent CT scanning (Victorino et al 2004). Studies have also been
performed in the thorax with more advanced methods. A method for 3D
imaging of lung ventilation created great interest on publication in 1996
(Metherall et al 1996), but this requires the use of four rings of 16 electrodes
each and has not been taken up for further clinical studies, presumably
because of practical diculties in applying this number of electrodes in
critically ill subjects. The above studies have used EIT at a single frequency
and relied on its anatomical imaging capability for the proposed clinical use.
An alternative philosophy, developed in the Sheeld group, has been to go
to lower spatial resolution and extract EITS parameters of the lung function
in conditions such as respiratory distress or pulmonary oedoema, on the
principle that such conditions diusely aect the lung and the method will
be more reliable. The characteristics of adult (Brown et al 1995) and neonatal
(Brown et al 2002) lungs have been obtained in normal subjects, but this has
yet to be taken up in further studies in pathological conditions.
B.4.2.3.
Breast tumours
444
Brain function
There are already excellent methods for imaging brain anatomy and
functionx-ray CT, MRI and functional MRI. EIT has the potential,
however, to oer a low-cost portable system for imaging brain abnormalities like epileptic activity or stroke, where it is not practicable to undertake
serial or rapid imaging in a large scanner. For example, it could enable takeup of a new treatment for stroke. New thrombolytic (clot-busting) drugs
have been shown to improve outcome in acute stroke, but must be administered within three hours of the onset. Neuroimaging must be performed
rst, in order to determine the cause of the stroke; about 15% of strokes
are due to a haemorrhage, and thrombolysis must not be given in these
patients, as it may make the haemorrhage extend. In practice, it is not possible to obtain and report a CT scan in the recommended 30 min. EITS could
be available in casualty departments and used to provide images which
would enable distinction of haemorrhagic from ischaemic stroke, and so
enable the rapid use of thrombolytic drugs. It also has the potential for
imaging the small impedance changes associated with opening of ion
channels during activity in the brain, which is not presently possible by
any other method and would be substantial advance. Unfortunately, EIT
of the brain has to overcome the diculty of injecting current through the
resistive skull.
Systems optimized for brain imaging have been developed at University
College London. Imaging through the skull with reasonably good resolution
has been shown to be possible, mainly by using widely spaced electrodes
for current injection (Bagshaw et al 2003a). A series of pilot studies in anaesthetized animals with electrodes placed directly on the brain, and the
Sheeld Mark 1 system, conrmed that suitable changes could be imaged
in stroke (Holder 1992b), epilepsy (Rao et al 1997) and evoked activity
(Holder et al 1996b). In humans, with recording in 3D using scalp electrodes,
reproducible impedance changes have been recorded during physiologically
evoked activity (Tidswell et al 1999) and epilepsy (Fabrizi et al 2004), but
Current developments
445
the reconstructed images were noisy and did not reveal consistent changes.
At the time of writing, trials are in progress to assess the utility of EIT in
acute stroke and epilepsy with improved multifrequency hardware and
reconstruction algorithms.
B.5.
CURRENT DEVELOPMENTS
This review has covered applications with conventional EIT. There are two
new methods, with considerable potential, which are still in technical
development, and have not yet been used for clinical studies. Magnetic
induction tomography (MIT) is similar in principle to EIT, but injects
and records magnetic elds from coils. It has the advantages that the position of the coils is accurately known and there is no skin-electrode
impedance, but the systems are bulkier and heavier than EIT. In general,
higher frequencies have to be injected in order to gain a sucient signalto-noise ratio. Until now, spatial resolution has been the same or worse
than EIT. The method could oer advantages in imaging brain pathology,
as magnetic elds pass through the skull, and may in the thorax or abdomen
if the method can be developed to demonstrate improved sensitivity over
EIT. MR-EIT (magnetic resonance-EIT) requires the use of an MRI scanner. Current is injected into the subject and generates a small magnetic
eld that alters the MRI signal. The pattern of resistivity in three dimensions
may be extracted from the resulting changes in the MRI images. This therefore loses the advantage of portability in EIT, but has the great advantage
of high spatial resolution of MRI. It could be used to generate accurate
resistivity maps for use in models for reconstruction algorithms in EIT,
especially for brain function, where prior knowledge of anisotropy is
important.
Biomedical EIT is, at the time of writing, in a phase of consolidation,
where optimized EIT systems are still being assessed in new clinical situations. Almost all clinical studies have been undertaken with variants of
the 2D Sheeld Mark 1 system. Several groups are near completion of
more powerful systems with improved instrumentation and reconstruction
algorithms, with realistic anatomical models and non-linear methods. The
most promising applications appear to be in breast cancer screening,
optimization of ventilator settings in ventilated patients, brain pathology
in acute stroke and epilepsy, and gastric emptying. Although there is a
commercial application in breast cancer screening with an impedance scanning device, EIT has yet to full its promise in delivering a robust and
widely accepted clinical application. Well funded clinical trials are in
progress in the above applications, and there seems to be a reasonable
chance that one or more, especially if using improved technology, may
prove to be the breakthrough.
446
REFERENCES
Avill R, Mangnall Y F, Bird N C, Brown B H, Barber D C, Seagar A D, Johnson A G and
Read N W 1987 Applied potential tomography. A new noninvasive technique for
measuring gastric emptying Gastroenterology 92 10191026
Bagshaw A P, Liston A D, Bayford R H, Tizzard A, Gibson A P, Tidswell A T, Sparkes
M K, Dehghani H, Binnie C D and Holder D S 2003 Electrical impedance tomography of human brain function using reconstruction algorithms based on the nite
element method Neuroimage 20 752764
Baisch F J 1993 Body uid distribution in man in space and eect of lower body negative
pressure treatment Clin. Investig. 71 690699
Barbas C S, de Matos G F, Okamoto V, Borges J B, Amato M B and de Carvalho C R
2003 Lung recruitment maneuvers in acute respiratory distress syndrome Respir.
Care Clin. N. Am. 9 401418, vii
Benabid A L, Balme L, Persat J C, Belleville M, Chirossel J P, Buyle-Bodin M, de
Rougemont J and Poupot C 1978 Electrical impedance brain scanner: principles
and preliminary results of simulation TIT. J. Life Sci. 8 5968
Boone K, Barber D and Brown B 1997 Imaging with electricity: report of the European
Concerted Action on Impedance Tomography J. Med. Eng Technol. 21 201232
Boone K G and Holder D S 1996 Current approaches to analogue instrumentation design
in electrical impedance tomography Physiol Meas. 17 229247
Brown B and Seagar A 1987 The Sheeld data collection system Clinical Physics and
Physiological Measurements 8 9197
Brown B H 2003 Electrical impedance tomography (EIT): a review J. Med. Eng Technol.
27 97108
Brown B H, Leathard A D, Lu L, Wang W and Hampshire A 1995 Measured and expected
Cole parameters from electrical impedance tomographic spectroscopy images of the
human thorax Physiol Meas. 16 A57A67
Brown B H, Primhak R A, Smallwood R H, Milnes P, Narracott A J and Jackson M J 2002
Neonatal lungs: maturational changes in lung resistivity spectra Med. Biol. Eng
Comput. 40 506511
Cherepenin V A, Karpov A Y, Korjenevsky A V, Kornienko V N, Kultiasov Y S, Ochapkin
M B, Trochanova O V and Meister J D 2002 Three-dimensional EIT imaging of breast
tissues: system design and clinical testing IEEE Trans. Med. Imaging 21 662667
Cook R D, Saulnier G J, Gisser D G, Goble J C, Newell J C and Isaacson D 1994 ACT3: a
high-speed, high-precision electrical impedance tomograph IEEE Trans. Biomed. Eng
41 713722
Eyuboglu B M, Oner A F, Baysal U, Biber C, Keyf A I, Yilmaz U and Erdogan Y 1995
Application of electrical impedance tomography in diagnosis of emphysemaa
clinical study Physiol Meas. 16 A191A211
Fabrizi L, Sparkes M, Holder D S, Yerworth R, Binnie C D and Bayford R 2004 Electrical
Impedance Tomography (EIT) During Epileptic Seizures: Preliminary Clinical
Studies. XII International Conference on Bioimpedance and Electrical Impedance
Tomography, Gdansk, Poland
Fitzgerald A J, Holder D S, Eadie L, Hare C and Bayford R H 2002 A comparison of
techniques to optimize measurement of voltage changes in electrical impedance
tomography by minimizing phase shift errors IEEE Trans. Med. Imaging 21
668675
References
447
448
References
449