Proceedings of the IASTED International Conference

Computational Bioscience (CompBio 2010)

November 1 - 3, 2010 Cambridge, Massachusetts, USA

ANATOMY GUIDED HYBRID DEFORMABLE MODEL FOR

RECONSTRUCTION OF BRAIN CORTEX FROM MR IMAGE
Somojit Saha Rohit Kamal Chatterjee Sarit Kumar Das Avijit Kar
Department of CSE Department of CSE Department of Mechanical Engineering Department of CSE
Jadavpur University BIT Mesra IIT Madras Jadavpur University
Kolkata, India Kolkata, India Chennai, India Kolkata, India
somojitsaha@yahoo.co.in rkchatterjee@bitmesra.ac.in skdas@iitm.ac.in avijit_kar@cse.jdvu.ac.in

ABSTRACT suffer from manual interpretations [1, 3, 4].

This paper presents an Anatomy Guided Hybrid Deformable Deformable model-based segmentation approaches
Model (AGHD) for fully automatic reconstruction of outer consider an object boundary as a whole and can make use of
cortical surface of brain from MR image. Apart from its fully a priori knowledge to constrain the segmentation problem
automatic nature, the algorithm requires tuning of the least and thereby it can overcome many of the limitations of
number of parameters and avoids all kinds of assumption and traditional image processing techniques [3]. Among two
approximation. This strength of the algorithm is a derivative types of deformable models, parametric deformable models
of rich tissue specific information of the images acquired represent curves and surfaces explicitly in its parametric
with signal attenuation from the gray matters. form and its popularity in medical image analysis is credited
Using signal nulling effect the acquisition protocol to the work of “snake” by Kass et al. [5, 6]. Other variant of
has enhanced the intensity difference between white and gray deformable models is geometric deformable model which is
matters in the reconstructed MR image. This leads to the based on the theory of curve evolution and geometric flow,
generation of a histogram where pixel values of different represents curves and surfaces implicitly as a level set of an
anatomical structures are distributed around separate evolving scalar function and used as a powerful technique for
dominant modes. An algorithm for automated multilevel computing interface motion [1, 4, 6]. A number of groups
thresholding for partitioning specific modes into initial brainhave attempted to implement both types of deformable
contour has been highlighted. models in cortical reconstruction as it is a fundamental step
Finally AGHD model has hybridized the essence of for brain image registration [7, 8], image-guided
traditional “snake” model and the Generalized Gradient neurosurgery [9, 10], brain geometry analysis [11, 12] and
Vector Flow deformable contour with precisefunctional mapping [13, 14]. Davatzikos and Prince used a
neuroanatomical guidance for accurate reconstruction of ribbon for mapping the cortex [15]. Davatzikos and Bryan
CSF/gray matter interface. The algorithm has been tested on used a deformable surface model to obtain a shape
a large dataset with great success and validated by a robust representation of cortex and proposed an active contour
index with highly encouraging outcome. algorithm for determining the spine of such a ribbon [16].
McDonald et al. designed an iterative algorithm for
KEY WORDS simultaneous deformation of multiple surfaces to segment
Magnetic resonance imaging, Image segmentation, MR brain images using cost function minimization [17].
Deformable model, Cortical reconstruction Kapur et al. also used a snake approach, along with EM
segmentation and mathematical morphology [18]. Teo et al.
1. Introduction incorporated knowledge of cortical anatomy with deformable
models, in which white matter and CSF regions were first
Human cerebral cortex reconstruction from MR image is one segmented, then the connectivity of the white matter was
of the challenging problems in computational neuroanatomy verified. Finally, a connected representation of the gray
because of its highly convoluted complex structures and matter was created by growing out from the white matter
marked variability within and across individuals. Apart from boundary [19]. Xu et al. proposed the generalized gradient
these inherent biological attributes of the anatomic shapes of vector flow (GGVF) deformable surface in conjunction with
interest, in case of MRI data, post-hoc processing is limited tissue membership functions for reconstructing the central
by factors like image contrast, resolution, SNR, RF field cortical layer halfway between the gray/white and gray/CSF
inhomogeneity, sampling artifact etc. which render the boundaries [4, 20]. McInerney and Terzopoulos developed
boundaries of the structures indistinct and discontinuous [1, topologically adaptive snakes (T-snakes) and implemented to
2]. That is why to integrate the cortical boundaries into a segment gray/white interface in MR brain image slice [21].
coherent mathematical description using model-free, low Apart from these parametric deformable models, Zeng et al.
level image processing technique such as thresholding, edge implemented geometric deformable models by developing a
detection and linking, region growing, relaxation labeling coupled surfaces approach for automatically segmenting a
and mathematical morphology operations produce inefficient volumetric cortical layer from 3D image. They used a set of
results since they consider local information only and highly coupled differential equations, with each equation

DOI: 10.2316/P.2010.728-015 599

determining the evolution or propagation of a surface within
a level set framework [22, 23].
Despite so many attempts and approaches, fully
automated reconstruction of the outer cortical surface is still
a big challenge. Though all approaches address the challenge
with success to a great extent, question of automation, close
initialization of the deformable models and validation of the
end result are still under investigation. All approaches suffer
from varied amount of human interaction, approximation and
assumption, heuristically tuning of several parameters and
above all the questionable representation of the true surface.
In case of deformable ribbon of Davatzikos et al. close
initialization and human interaction are needed to force the
ribbon into the sulcal folds [1, 15, 16]. Algorithm of
McDonald et al. is not only computationally expensive but
also requires tuning a number of weighting factors in the cost
function [1, 17]. Although volume measurement may be
reliable with the approach of Kapur et al., the shape of the
outer surface is poorly representative of the true surface [1,
18]. GGVF deformable surface of Xu et al. retrieves central
cortical layer, not the outer cortical surface. It may be useful
for feature based image registration but it can not meet the
other purposes like functional mapping or morphometric
analysis. Besides this, the method depends on human
interaction to extract intracranial tissue from entire brain
image, to compute an initial estimate of the cortical surface
and depends on tuning of a number of parameters empirically
[4, 20]. T-snake segments the gray/white interface not the
gray/CSF interface of outer cortical surface. In approach of
Zeng et al. coupling between the surfaces incorporates the
notion of an approximately fixed thickness separating the
surfaces everywhere in the cortex. Since three layered
granular type heterotypical cortex of postcentral gyrus,
superior temporal gyrus and part of hippocampal gyrus or
agranular type heterotypical gyrus of precetral gyrus differs
in thickness from six layered homotypical neocortex of the
remaining area [24, 25, 26], this approximation does not
reflect the true scenario.
Considering all these shortcomings of the existing
deformable models for segmentation of the cortical outer
surface from MR image of brain we have developed a novel
Anatomy Guided Hybrid Deformable (AGHD) Model to
make it fully automated and accurate. For this purpose we
tried to identify the basis of the drawbacks of all existing
models. Actually all deformable models with or without
incorporation of a priori knowledge of brain anatomy highly
depends on parameters derived from image information
namely pixel values of the classified tissues, relative
differences of their pixel values, edge maps, gradient and so
on. Apart from the structural variability of the complex brain
structure, common source of error is lack of specific
information from a particular tissue class. Different tissue
classes may share the same information, most importantly the
pixel value. If specific non-overlapping information does not
resides in the image at acquisition level mere post-hoc
processing is not enough to distinguish different tissue
classes depending on the image information. In spite of
advancement of post hoc processing, image acquisition plays
the central role to meet the objectives. Once the objective is
600
defined, image acquisition should be done in a way that leads
to minimum computational overheads with maximum
accuracy. For this application we have focused on the highest
contrast between white and gray matters as well as CSF and
gray matters at acquisition level to explicitly delineate the
CSF/gray interface without any post-hoc enhancement. The
acquisition protocol has been detailed in [27]. The resulting
images of this protocol have brilliant gray-white contrast
which is evident by the unique feature of the histogram. In
the histogram, pixels of specific brain structures namely
CSF, gray and white matters, background and bones have so
distinct gray levels that they group into dominant modes with
prominent valleys in between. CSF is confined in the
lowermost range of the gray scale and the gray matter,
background, white matter and bones are distributed in that
order towards the lighter direction of the gray scale [27, 28,
29]. Using this histogram feature automated thresholding in
conjunction with prior knowledge of brain anatomy segments
the brain with thin layer of CSF encapsulated by the
meninges [28, 29]. In this way, we have designed a fully
automated algorithm for extraction of intracranial tissues
including brain, CSF and meninges. Salient features of the
algorithm have been mentioned in section 2 with illustration.
External boundary of the meninges serves as the initial
estimator of the deformable contour. Then we have
hybridized the parametric classical snake of Kass et al. with
GGVF proposed by Xu et al. to exploit the power of greater
capture range and greater progression into boundary
concavity [4, 20, 30, 31] along with precise anatomical
guidance. Detail of the algorithm is discussed in section 3.
Result of the proposed AGHD model for segmentation of
CSF/gray interface is not only encouraging but also holds
some unique merits. The most outstanding feature of the
algorithm is complete automation without any assumption or
approximation. Accuracy of the result has been evaluated
quantitatively by a robust index proposed by Zizdenbos et al.
[32] with highly positive outcome.
2. Determination of the Initial Estimator
As T1-weighting produces better contrast between white and
gray matters and inversion-recovery sequence generates
greater T1 contrast [33], we used inversion-recovery spin
echo sequence to generate T1-weighted coronal view of the
brain. A consistent unique feature has been identified in the
histogram of MR image of brain acquired with the proposed
protocol for better visulization of the target gray structures.
This leads to generation of a histogram where pixel values of
different anatomical structures are distributed around
separate dominant modes. Global multilevel thresholding of
the histogram by heuristic approach, based on visual
inspection on an image is illustrated in Fig. 1. From this
illustration it is evident that CSF, gray matter, background,
white matter and bones are distributed respectively in the
lighter direction of the gray scale. This unique feature of the
histogram has been combined with the prior knowledge of
inherent anatomy of brain to develop an algorithm for
automated multilevel thresholding of the histogram for
partitioning specific modes into brain contour.

Fig. 1: Histogram analysis: 1st row shows histogram and the
original image, following rows illustrates histogram with
faded out thresholded region on left and the segmented image
on right.
At first third mode of the histogram that is
background is thresholded out automatically using histogram
statistics [34]. Pixels containing the gray value lower than the
lower threshold of the background must be part of CSF and
the gray matters (Cortical and subcortical) as first and second
mode of the histogram representative of CSF and gray matter
respectively. So, mere global thresholding of the entire image
with lower threshold of the background produce boundary of
601
the intracranial tissue consist of CSF and cortical gray
matter. Since meninges shares the same gray value as gray
matter, the segmented brain contour is encapsulated with
meninges of varied thickness. Though this global
thresholding produce perfect result for posterior and middle
part of the brain, nasal cavity is misclassified as brain tissue
in frontal area. This non-brain region is strategically
eliminated after identification of the erroneous slices.
Position of the centroid of the bony skull in the brain
bounding box plays a deterministic role in this strategy [28,
29]. Thus, the brain contour surrounded by CSF and
encapsulated by meninges is extracted automatically. One
pixel thick outer boundary of the meninges serves as the con-
tour initialization for cortical reconstruction. End results of
this automated histogram based segmentation spanning the
entire brain from posterior to anterior are depicted in Fig 2.
3. AGHD Model for Cortical Segmentation
After computation of the initial contour three stepped
Anatomy Guided Hybrid Deformable (AGHD) model has
been implemented for finding out the CSF/gray interface in
each slice. In first step, classical energy-minimization snake
of kass et al. [5] is implemented to drive the contour towards
inner margin of the meninges. In next step, under guidance of
topological distribution of CSF, snake is reinitialized in the
vicinity of the cortex after elimination of thick meningeal
fold and finally it is deformed according to GGVF model of
Xu et al. [4, 31]. The detail steps are as follows:
3.1 Classical Energy-Minimizing deformable model
In this approach, initially the contour is one pixel thick outer
most margins of the meninges derived from automated
histogram based segmentation described in section 3. Idea of
implementing the classical energy-minimizing snake
[1,4,5,6] is to drive the initial contour towards inner margins
of the meninges that is meninges/CSF interface.
The deformable contour is represented in the image
plane ( x, y ) ∈ ℜ 2 as v( s ) = ( x( s ), y ( s )) Τ , where x and y
are the coordinate functions and s ∈ [0,1] in the parametric
domain. The shape of the contour is typically determined by
variational formulation expressed as
Ε = Ε int (v( s )) +Ε ext (v( s ))ds. (1)
s
The functional can be viewed as a representation of the
energy of the contour, and the final shape of the contour
corresponds to the minimum of this energy. The first term
prescribes a priori knowledge about the model such as its
elasticity and rigidity and can be expressed as
1 2 2
∂v ∂ 2v
Εint 1 ∂s 2 ∂s 2 ds.
= w ( s ) +w ( s ) (2)
0
w1(s) and w2(s) two parameters dictate the simulated
physical characteristics of the contour. The second term in
(1) is the external force which couples the snake to the
image. In accordance with the calculus of variations, the
contour v(s) that minimizes the energy E must satisfy the
Euler-Lagrange equation
 Fig 2: Initial contour: Initial contour for deformable model derived by automated histogram based segmentation
∂ ∂v ∂2 ∂ 2v
w1 − 2 w2 2 − ∇ Ε ext = 0
∂s ∂s ∂s ∂s
A numerical solution of (3) is found by discretizing the
equation and solving the discrete system iteratively.
3.1.1 Anatomical guidance to define external force field
The basic objective of classical deformable model is to bring
the initial contour in the area containing CSF. For this
reason, we have designed an external force field by using
topological guidance of distributed CSF. For this purpose,
topological distribution of CSF is segmented out. Since we
have the image with very distinct pixel values for different
tissue types, a binary image of CSF and the background has
been produced by automated optimal global thresholding [35,
36] of first two mode of the histogram. Lower threshold of
the third mode of the histogram is determined by the
background removal algorithm [34] as follows:
In reconstructed MR data, background noise is
normally distributed white noise and has a Rayleigh
distribution in the image histogram [37] and expressed as
2
f f
Pn o is e ( f ) = exp −
σ 2
2σ 2
where σ is the standard deviation of the channel noise.
Global maxima or height of any Rayleigh curve rmax is
related to σ as σ = 0.607 rmax
All maxima h( fi ) of the histogram function h( f ) and the
corresponding gray values f i are traced. Using this value of
h( fi ) in (5), initially a Rayleigh curve ri ( f ) is generated at
each maximum. Similarity of ri ( f ) in the range σ i to the cut
off value f c = 2σ i is measured quantitatively with the
original histogram in the same range using a similarity index
derived from the Kappa statistic as proposed by Zijdenbos et
al. [32]. If area under ri ( f ) in the mentioned range is
denoted as Ri and area under the original histogram in the
same rang is denoted as H i , then the similarity index
becomes
Ri Hi
Si = 2
Ri + H i
where S i ∈ {0...1}
(3)
The relevant ri ( f ) for the highest Si is considered as the
Rayleigh for the noise distribution rnoise ( f ) . Peak of the
curve rnoise ( f ) coincides with the noise peak of the original
histogram. Then, rnoise ( f ) is scaled by a constant K and a
least square fit is performed by minimizing the error
expression
2
fc
f f2
h( f ) − K exp − (7)
f =0 σ 2
2σ 2
to get the best fit Rayleigh curve r ( f ) . Lower threshold of
r ( f ) is considered as the lower threshold of the third mode
that is image background.
Part of the histogram below this threshold is clearly
bimodal with well-defined peaks and valley for our image
dataset. First mode is of CSF and the second mode is of gray
matter. Because of this distinct bimodal distribution,
automated optimal global thresholding of the histogram up to
lower threshold of the third mode by the algorithm proposed
by Otsu [35, 36] efficiently segment out the distributed CSF
(4)
in the given image. Fig. 3 shows an example of segmentation
of CSF with intermediate result of lower threshold of the
third mode at gray level 146 and optimal global thresholding
at gray level 66. Logic operation of this binary image of CSF
(5) with the original image produces uniform gray value of zero
for entire CSF region instead of a range of gray values.
Thereby it produces a sharp distinct boundary at
meninges/CSF interface and CSF/gray matter interface with
enhanced gradient. Now, external potential is designed
whose local minima coincide with step edges and expressed
Ε
2
as ext
( x, y ) = − ∇ I ( x, y ) (8)
As meninges has got spatially varied thickness,
depending on the sharpness of the edge gradient and the
chosen parameters for the internal force, it may fail to come
completely inside the meninges. Besides these, in some area,
cortical surface may be attached to the meninges leaving no
Si visible subarachnoid space for CSF. Fig. 4 shows the end
result of this step. To overcome these entire problems,
topological guidance along with a prior knowledge of
(6) neuroanatomy has been incorporated in the next step.
602

(a)
(b) (c)
Fig. 3: CSF distribution: Histogram with best fit Rayleigh
curve, faded out part for optimal global thresholding and the
threshold for CSF distribution (a) applied on the image (b)
and the segmented CSF distribution (c)
3.2 Incorporation of anatomical guidance and
reinitialization of the deformable contour
CSF map which has been designed to enhance the gradient
across the edge adjacent to CSF, also provides topological
guidance to eliminate the thick part of meninges as well as
helps to design external force field for GGVF in the
following steps.
(a) (b)
Fig. 4: Removal of meninges: Classical deformable model
started with initial contour over outer boundary of the
meninges (a) and the final contour, red line in (b) resides
over inner boundary of the meninges
3.2.1 Removal of thick part of meninges
This part is designed to overcome the failure of classical
energy-minimizing snake in attempt to eliminate the
meninges completely. Since we have clear distribution of
CSF, the final contour of the classical snake is scanned pixel
603
by pixel along its length with inspection of surrounding eight
neighbors of each pixel. If any neighbor of the scanned pixel
contains the pixel value zero, it must be part of CSF and the
corresponding CSF set is traced by extraction of connected
component of pixel value zero [38]. If two pixels on the
length of the contour are adjacent to the common CSF set,
connected component of non-zero pixel value in between
them is identified and turned into zero in order to eliminate
those sets. In the thick meningeal fold where the classical
snake could not touch the meninges/CSF interface,
eliminated in this way on letting the cerebrum and
cerebellum the only tissue with non-zero pixel value.
3.2.2 Reinitialization of deformable contour
Before implementation of GGVF deformable model of Xu et
al., presence of the initial estimator must be ensured in the
capture range of the gradient vector field. For this purpose,
boundary of the cluster of non-zero pixel set (containing
cerebrum and cerebellum) is approximated by a polygon with
a method for finding minimum perimeter polygons proposed
by Kim and Sklansky [38, 39]. In this procedure boundary of
the cluster is enclosed by a set of two pixels thick
concatenated cells and the boundary is considered as rubber
band contained within the walls of a cell. If the rubber band
is allowed to shrink, it produces a polygon of minimum
perimeter that fits the geometry established by the cell strip.
This minimum perimeter polygon serves as the initial
contour for the final step of GGVF model.
3.3 GGVF deformable contour
Finally GGVF deformable model has been implemented for
extraction of CSF/gray matter interface by exploiting its
properties of greater capture range and superior convergence
of boundary concavity in comparison to the classical snake
model.
The method is started by defining an edge map f(x)
derived from the gray-level image. In our application we
have used Canny edge detector [40] to define the edge map
with very high gradient at CSF/gray matter interface because
of distinct topological guidance of distributed CSF as
described in section 4.1.1. The gradient of the edge map ∇f
has vectors pointing towards the edges, which are normal to
the edges.
With help of this edge map, GGVF field v is defined as
the equilibrium solution of the following vector partial
differential equation
vt = g ( ∇f )∇ 2 v − h( ∇f )(v − ∇f ) (9)
In (9) the first term is considered as the smoothing term,
since it alone produces a smoothly varying vector field. The
second term is referred as data term since it encourages the
vector field to be close to ∇f computed from the edge map.
Spatially varying weighting functions g (⋅) and h(⋅) are
applied to the smoothing and data term. For better
progression of the deformable contour in long, thin
indentation along the cortical boundary where two edges are
in close proximity, weighting functions are selected such that
g (⋅) gets smaller as h(⋅) becomes larger [4, 30, 31]. For this exactly similar pattern is 1 and index of more than 0.7 is
application, the weighting functions are as follows:
2
∇f
g ( ∇f ) = exp− (10)
κ In this paper we have tried to meet the basic challenge of
h ( ∇f ) = 1 − g ( ∇f ) (11)
The specification of κ determines to some extent the degree
of tradeoff between field smoothness and gradient
conformity.
Once the GGVF is designed it acts as an external force
field and replaces the potential force − ∇Eext in (3). Then
the equation is solved numerically by discretization and
iteration, in identical fashion to the traditional deformable
contour.
4. Test Results and Validation Checking
The proposed algorithm has been tested with 24 datasets of
coronal views of proposed acquisition protocol with excellent
accuracy. Fig. 5 shows final contour of the AGHD model
overlaid on the original image with intermediate outcome of
reinitialized deformable model for GGVF (discussed in
section 4.2.2) along with edge map for external force and
initial estimator.
For validation study, some slices were selected such
that the entire range of the image volume from anterior to
posterior was covered. A domain expert traced the target
manually on each slice and it was compared with the
automatically drawn head and brain contour using the
similarity index described by Zijdenbos et al.[32].
In case of segmented brain images of the coronal view
manually segmented cortical outer contour was highly
comparable to the automated ones and it is reflected by high
similarity indices, approximately 0.98 or more. Index for
considered as very good agreement [41].
5. Conclusion
reconstruction of outer cortical surface from MR image of
brain by developing a novel AGHD model which is
hybridization of traditional deformable model with GGVF
deformable contour along with prior precise anatomical
knowledge. The proposed algorithm claims its excellence in
terms of fully automation, tuning of the least number of
parameters and avoidance of any assumption or
approximation. No step of the algorithm, from elimination of
bones and soft tissues or initialization of the deformable
contour to final localization of AGHD model requires an iota
of human interpretation. Essence of the algorithm resides in
the rich non-overlapping tissue specific information in the
acquired datasets. Though the algorithm is specific for the
proposed acquisition protocol, the protocol is highly
applicable for the basic purposes of cortical reconstruction
viz. image registration, image-guided neurosurgery and
functional mapping when co-registered with functional
imaging modalities like SPECT, PET or fMRI since it
produces markedly different gray values for different class of
tissue.
In our data sets, though the signal from the gray
matters has been attenuated, it could not be nullified because
of hardware limitations and unavoidable computational
constraints [27]. However, the gray-white contrast is
markedly improved evidenced by distinct intricate cerebral
and cerebellar gray structures. Analysis of the histogram
reestablishes this subjective evaluation. The unique feature of
the histogram is that modal distribution of the gray values
directly correlated with a specific anatomical structure. From
the histogram one can easily identify a mode actually
representing a given anatomical structure. That is why

Fig. 5: Cortical segmentation: Initial estimator for classical deformable model in first column, edge map for external force
with superimposed reinitialized minimum perimeter polygon (red line) for GGVF deformable model in second column and
final contour of GGVF model with segmentation of CSF/gray matter interface (red line) in last column.

604
determination of the initial contour is easily be made by
global thresholding only. Actually, using global thresholding,
segmentation of the brain is possible only because of high
contrast of the adjacent tissues as well as full control to
manipulate it to meet the objective. As in our dataset pixel
values of CSF and gray matter are consistently lower than the
background noise, lower threshold of the background is
confidently chosen for the upper threshold for segmentation
of the initial brain boundary. As the brain is bounded by the
bony skull and the background noise never overlaps the
distribution of the bones, the chance of misclassification of
the bony part as brain is virtually eliminated. The only
shortcoming of the global thresholding is that the nasal cavity
is erroneously included in the segmented contour. This has
also been successfully eliminated by incorporating spatial
information of the intricate anatomy [28, 29]. As structural
organization of the bony cranium is more consistent than any
soft tissues we intend to shift the centroid towards bony
region to set a criterion. Sphenoid and the bony orbit help to
identify the anterior slices with coronal section of the frontal
lobe of the brain attached to the nasal cavities. No imprecise
morphological operations have been used in this algorithm
since our primary objective is measurement of subcortical
gray matters as well as quantitative study of cognition.
The initial brain contour derived in this way is
surrounded by CSF and encapsulated by meninges with
spatially varying thickness. Meninges is eliminated by using
traditional “snake” model and thus the deformable contour is
brought to the inner boundary of meninges. To design the
exernal force field we again exploit the unique feature of the
histogram that is modal distribution of the specific tissue
class for segmentation of topological distribution of CSF in
the image. By assigning the zero value to the distributed CSF
map and logic operation with the original image we actually
tries to have zero valued CSF and nonzero valued intracranial
tissues like cerebrum and cerebellum. This step is the most
crucial for success of the algorithm. By this way gradient of
CSF/meninges interface and CSF/gray matter interface
becomes very high which helps to define the external force
field for traditional snake as well as GGVF field in next step.
Besides this, it provides precise anatomical guidance to
eliminate thick meningeal folds which can not be eliminated
by the traditional snakes. Though some zero valued CSF
remain distributed in the ventricles, being far away of the
cortical gray matters they do not affect the following steps of
the algorithm anyhow.
Final contour of the traditional snake is scanned
pixel by pixel for its adjacency to the zero-valued CSF. In
case of any thick isolated meningeal fold, final contour of the
traditional snake passes arbitrarily through mid portion and
may not be adjacent to CSF. But first and last pixel of the
fold must be adjacent to the same CSF set. On checking this
condition after tracing these extreme points, non-zero set in
between is identified and turned into zero value to eliminate
it. This situation is more vivid around superior sagittal sinus.
This checking is made strategically not to affect the part of
the cortical folds which are attached to the meninges leaving
no CSF in between them. Thus all discrete non-zero valued
meningeal folds turns into zero selectively and the brain
605
remains as a set of non-zero value. Polygon of minimum
perimeter of that set serves as initial contour for GGVF
deformable contour and this initialization ensures starting of
the deformable contour in the capture range of the force field.
In our GGVF model, we have chosen a pair of weighting
function for better convergence of boundary concavity which
is very essential for our application to reconstruct the highly
convoluted cortical surface. The GGVF field computed using
this pair of weighting functions will conform to the edge map
gradient at strong edges, but will vary smoothly away from
the boundary. Despite medialness of the GGVF deformable
model, in this application the final contour perfectly lies over
CSF/gay matter interface, since the contour faces a very high
gradient at this interface when approaches externally. This
has occurred due to anatomical guidance of CSF distribution.
Instead of having a range of gray values, distributed CSF
posses gray value zero and gray matter distribution in the
gray scale is far away from zero. This produces very high
gradient at their interface and helps to segment out that
interface. The result is highly encouraging and this 2D
deformable contour is to be implemented as 3D surface for
reconstruction of the entire cortical surface.
7. References
[1] Isaac N. Bankman, Handbook of Medical Imaging
Processing and analysis, Academic Press, 2000
[2] Marijn E. Brummer, Russell M. Mersereau, Robert
L. Eisner, & Richard R. J. Lewine, “Automatic
Detection of Brain Contours in MRI Datasets”, IEEE
Transactions on Medical Imaging, 12(2), pp.153-166,
1993
[3] Tim McInerney and Demetri Terzopoulos,
“Deformable models in medical image analysis: a
survey,” Medical Image Analysis, 1(2), pp. 91-108, 1996
[4] C. Xu, “Deformable Models with Application to
Human Cerebral Cortex Reconstruction from Magnetic
Resonance Images”, Ph.D. thesis, The Johns Hopkins
University, Baltimore, MD, January 1999
[5] M. Kass, A. Witkin and D. Terzopoulos, “Snakes:
Active contour models”, International Journal of
Computer Vision, 1(4), pp. 321-331, 1988.
[6] S. Osher and N. Paragios, “Geometric level set
methods in imaging, vision and graphics,” Springer-
Verlag, 2003
[7] G. E. Christensen, S. C. Joshi and M. I. Miller,
“Volumetric Transformation of Brain Anatomy”, IEEE
Transactions on Medical Imaging, 16(6), pp.864-877,
1997
[8] Chris A. Davatzikos, “Spatial transformation and
registration of brain images using elastically deformable
models”, Computer Vision and Image Understanding,
66, pp. 207-222, 1997
[9] W. E. L. Grimson, G. J. Ettinger, T. Kapur, M. E.
Leventon, W. M. W. III and R. Kikinis, “Utilizing
segmented mri data in image-guided surgery”, Int. J.
Pattern Recog. Artif. Intell., 11(8), pp. 1367-1397, 1996
[10] M. Vaillant, C. Davatzikos, R. H. Taylor and R. N.
Bryan, “A path-planning algorithm for image-guided
neurosurgery”, in Lecture Notes in Comp. Sci.:
CVRMed-MRCAS’97, 1205, pp.467-476, 1997
[11] S. Joshi, J. Wang, M. I. Miller, D. C. V. Essen and
U. Grenander, “On the differential geometry of the
cortical surface,” in Proc. Of the SPIE: Vision Geometry
IV, 2573, pp. 304 -311, Aug. 1995
[12] L. D. Griffin, “The intrinsic geometry of the
cerebral cortex”, Journal of Theoretical Biology, 166(3),
pp. 261-273, 1994
[13] A. M. Dale and M. I. Sereno, “Improved
localization of cortical activity combining EEG and
MEG with MRI cortical surface reconstruction: A linear
approach”, Journal of Cogn. Neuroscience, 5(2), 162-
176, 1993.
[14] H. A. Drury and D. C. V. Essen, “Functional
specializations in human cerebral cortex analyzed using
the visible man surface-based atlas”, Human Brain
Mapping, 5, pp.233-237, 1997
[15] Chris A. Davatzikos and Jerry L. Prince, “An Active
Contour Model for Mapping the Cortex”, IEEE
Transactions on Medical Imaging, 14(1), pp.65-80, 1995
[16] Chris A. Davatzikos and R. N. Bryan, “Using a
deformable surface model to obtain a shape
representation of cortex”, IEEE Transactions on Medical
Imaging, 15(6), pp.785-795, 1996
[17] D. MacDonald, D.Avis and A. C. Evans, “ Multiple
surface identification and matching in magnetic
resonance images,” in SPIE Proc. VBC’94, 2359, pp.
160-169, 1994
[18] T. Kapur, W. Grimson, W. Wells and R. Kikinis, “
Segmentation of brain tissue from magnetic resonance
images”, Medical Image Analysis, 1, pp.109-128, 1996
[19] P. C. Teo, G. Sapiro and B. A. Wandell, “Creating
connected representations of cortical gray matter for
functional MRI visualization”, IEEE Transactions on
Medical Imaging, 16(6), pp.852-863, 1997
[20] C. Xu, D. L. Pham, M. E. Rettmann, D. N. Yu and
J. L. Prince, “Reconstruction of the human cerebral
cortex from magnetic resonance images”, IEEE
Transactions on Medical Imaging, 18(6), pp.467-480,
1999
[21] Tim McInerney and Demetri Terzopoulos, “ T –
snakes: Topology adaptive snakes”, Medical Image
Analysis, 4, pp.73-91, 2000
[22] X. Zeng, L. H. Staib, R.T. Schultz and J. S. Duncan,
“Volumetric Layer segmentation using coupled surface
propagation” in Computer vision and Pattern
Recognition, pp. 708-715. IEEE Comp. Soc., Los
Alamitos, 1998
[23] X. Zeng, L. H. Staib, R.T. Schultz and J. S. Duncan,
“ Segmentation and measurement of the cortex from 3D
MR images using coupled surface propagation”, IEEE
Transactions on Medical Imaging, 18(10), 1999
[24] R. S. Snell, “ Clinical Neuroanatomy for Medical
Students,” Lippincott-Raven Publishers, Fourth edition,
1997
[25] M J T FitzGerald and Jean Folan-Curran, “ Clinical
Neuroanatomy and Relate Neuroscience,” W.B.
Saunders, Fourth edition, 2002
[26] R.H. S. Carpenter, “Neurophysiology”, Arnold,
Fourth edition, 2003
[27] Somojit Saha, Sarit K. Das, Avijit Kar, “High
Resolution MR Image of Brain With Signal Attenuation
of Gray Matter”; IST, IEEE Instrumentation and
606
Measurement Society, Thessaloniki, Greece, pp. 311-
315, 2010
[28] Somojit Saha, Sarit k Das, Avijit Kar;
“Segmentation of MR Image of Brain: A New Method
Using Unique Histogram Features and Prior Knowledge
of Brain Anatomy,” IPCV , WORLDCOMP, Las Vegas,
Nevada, USA,2010
[29] Somojit Saha, Sarit k Das, Avijit Kar; “A New
Segmentation Technique for Brain and Head from High
Resolution MR Image Using Unique Histogram
Features,” IPTA (IEEE & EURASIP), Paris, France,
2010
[30] C. Xu and J. L. Prince. “Snake, shape, and gradient
vector flow,” IEEE Transaction on Image Processing,
7(3), pp. 131-139, 1998
[31] C. Xu and J.L. Prince, “Generalized gradient vector
flow external forces for active contours,” Signal
Processing, An International Journal, 71(2), pp. 131-
139, 1998
[32] A. Zijdenbos, B. Dawant, R. Margolin, & A.
Palmer, “Mophometric analysis of white matter lesions
in MR images: Method and Validation”; IEEE
Transactions on Medical Imaging, 13(4), pp. 716-724,
1994
[33] Zhi-Pei Liang, & Paul C. Lauterbur, Principle of
Magnetic Resonance Imaging – A signal processing
perspective, IEEE Press, 2000
[34] Somojit Saha, Sarit k. Das, Subhajit Das, Avijit Kar;
“An improved algorithm for automated segmentation of
the head contour from MR image”, VIIP, IASTED
Cambridge, UK, pp. 25-32, 2009
[35] N. Otsu “ A threshold selection method from gray
level histogram”, IEEE transaction on systems, man, and
cybernetics, SMC 9, pp 62-66, 1979
[36] R. M. Haralick and L. G. Shapiro, “Computer and
Robot Vision, vol. 1, Addison-wesley Publishing
company, 1992
[37] M. Henkelman, “Measurement of signal intensities
in the presence of noise in MR images”, Medical
Physics, 12(2), pp. 232-233, 1985
[38] Rafael C. Gonzalez, & Richard E. Woods, Digital
Image Processing 2nd ed. (Prentice-Hall, 2002).
[39] C. E. Kim and J. Sklansky, “Digital and cellular
convexity,” Pattern recognition, 15(5), pp. 359-367,
1982
[40] J. F. Canny, “A computational approach to edge
detection,” IEEE Transaction on Pattern Analysis and
Machine Intelligence,” 8(6), pp. 679-698, 1986
[41] John J. Bartko, “Measurement and Reliability:
Statistical Thinking Considerations”; Schizophrenia
Bulletin, 17(3), pp. 483-489, 1991