Computed Radiation Imaging: Physics and Mathematics of Forward and Inverse Problems

Table of Contents

Cover image

Front Matter

Copyright

Preface

1. Radiation Imaging

1.1. Why Radiation?

1.2. Imaging Modalities

1.3. Direct and Reconstructed Imaging

1.4. The Forward and Inverse Problems

1.5. Forward and Inverse Mapping

Introduction

2. Radiation Transport

2.1. Introduction

2.2. Variables

2.3. Cross Sections

2.4. Boltzmann Transport Equation

2.5. Source-Free Steady-State Problem

2.6. Steady-State Problem in Void

2.7. Point-Kernel Method

2.8. Charged Particles

3. Measurement Models

3.1. Formulation

3.2. Scaling

3.3. Measured Response

3.4. Sensitivity

3.5. Variability

3.6. Components

3.7. Image Grid

3.8. Idealization

3.9. Computer Coding

4. Transmission

4.1. Basic Model

4.2. Physical Relevance

4.3. Discretization

4.4. Nature of Radiation Source

4.5. Secondary Radiation

4.6. Scattering

4.7. Sensitivity

4.8. Variability

5. Emission

5.1. Embedded Radiation

5.2. Induced Emission

5.3. Discretization

5.4. Sensitivity

5.5. Sources

5.6. Interfering Effects

6. Scattering

6.1. Introduction

6.2. Single-Scattering Model

6.3. Multiple Scattering

6.4. Compton Scattering

6.5. Neutron Elastic Scattering

6.6. Discretization

6.7. Sensitivity

Introduction

7. Features

7.1. Discretization

7.2. Well-Posed Problem

7.3. Existence

7.4. Uniqueness

7.5. Continuity

7.6. Ill-Posed Problem

7.7. Ill-Conditioning

7.A. Basics of Functional Analysis

8. Formulation

8.1. Matrix

8.2. Functional

8.3. Analytic

8.4. Probabilistic

8.A. Probabilistic Basis of Maximum-Likelihood and Least-Squares Methods

9. Preprocessing of Measurements

9.1. Number of Measurements

9.2. Frequency Analysis

9.3. Spatial Filtering of Noise

9.4. Consistency and Smoothing

10. Matrix-Based Methods

10.1. Error Propagation

10.2. Singular Value Decomposition

10.3. Least Squares

10.4. Regularization Methods

10.5. Regularization-Parameter Determination

10.6. Iterative Methods

10.7. Nonlinear Problems

10.8. Software

11. Functional Optimization

11.1. Formulation

11.2. Effect of Number of Measurements

11.3. Sensitivity to Measurement Uncertainty

11.4. Minimization

11.5. Search Methods

11.6. Genetic Evolution

11.7. Simulated Annealing

11.8. Neural Networks

12. Analytic Methods

12.1. Radon Transform

12.2. Two-Dimensional Fourier Transforms

12.3. Backprojection

12.4. Fan-Beam Transmission Tomography

12.5. Cone-Beam Transmission Tomography

12.6. Emission Imaging

12.7. Scatter Imaging

12.8. Computer Codes

12.9. Wavelet Transforms

13. Probabilistic Methods

13.1. Bayesian - Minimum Information

13.2. Poisson Distribution

13.3. Normal Distribution

13.4. Maximum a posteriori (map)

13.5. The Monte Carlo Method

14. Incomplete Problems

14.1. Incompleteness

14.2. General Solution Methods

14.3. Estimation Maximization

14.4. Markov Random Fields

15. Testing

15.1. Ideal Problem

15.2. Noisy-Ideal Problem

15.3. Independently-Simulated-Data Problem

15.4. Laboratory Problem

15.5. Image Quality

15.6. Test Objects

16. Post-Processing

16.1. Image Convolution

16.2. Image Degradation

16.3. Frequency Filtering

16.4. Matrix Based

16.5. Statistical Methods

16.6. Optimization

16.7. Blind Deconvolution

Bibliography

Front Matter

Computed Radiation Imaging

Computed Radiation Imaging

Physics and Mathematics of Forward and Inverse Problems

Esam M.A. Hussein

Department of Mechanical Engineering

University of New Brunswick

Fredericton, N.B. Canada

AMSTERDAM • BOSTON • HEIDELBERG • LONDON • NEW YORK • OXFORD • PARIS • SAN DIEGO • SAN FRANCISCO • SINGAPORE • SYDNEY • TOKYO

Copyright

Elsevier

32 Jamestown Road London NW1 7BY

225 Wyman Street, Waltham, MA 02451, USA

First edition 2011

Copyright © 2011 Elsevier Inc. All rights reserved.

No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, recording, or any information storage and retrieval system, without permission in writing from the publisher. Details on how to seek permission, further information about the Publishers permissions policies and our arrangement with organizations such as the Copyright Clearance Center and the Copyright Licensing Agency, can be found at our website: www.elsevier.com/permissions.

This book and the individual contributions contained in it are protected under copyright by the Publisher (other than as may be noted herein).

Notices

Knowledge and best practice in this field are constantly changing. As new research and experience broaden our understanding, changes in research methods, professional practices, or medical treatment may become necessary.

Practitioners and researchers must always rely on their own experience and knowledge in evaluating and using any information, methods, compounds, or experiments described herein. In using such information or methods they should be mindful of their own safety and the safety of others, including parties for whom they have a professional responsibility.

To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume any liability for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions, or ideas contained in the material herein.

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

Library of Congress Cataloging-in-Publication Data

A catalog record for this book is available from the Library of Congress

ISBN 978-0-123-87777-2

For information on all Elsevier publications visit our website at elsevierdirect.com

This book has been manufactured using Print On Demand technology. Each copy is produced to order and is limited to black ink. The online version of this book will show color figures where appropriate.

Preface

Esam M.A. Hussein

Fredericton, October 2010

Imaging with atomic/nuclear radiation takes two forms: (1) a direct imprint as in radiography, or (2) a computed image as in computed tomography (CT) scans. Both forms of imaging are an integral part of modern medical-diagnostic practice, and are utilized in research and industrial imaging. Photons, in the form of x-rays or gamma-rays, are typically employed in such imaging systems, but neutrons are also utilized when photons cannot provide the needed indications. Radiography gives a direct imprint of the radiation that succeeds in penetrating the interrogated object. Therefore, radiographic imaging does not require much post-exposure numerical processing to extract the image features. In the case of computed radiography (CR), ¹ some computations are done to enhance the image quality and enable image manipulation, but such computations are not essential for producing the raw image. In digital radiography (DR), ² x-rays are captured directly, providing a digital image (without intermediate storage) that can also be numerically manipulated for optimum viewing.

¹Computed radiography (CR) employs a phosphorus plate that captures x-ray photons in gaps and stores their energy in its crystal structure. The stored energy is later extracted, in the form of light, by a low-energy laser, which is collected and converted into electrical signals, stored and processed for viewing.

²Digital radiography (DR) involves the use of

CCD

(charged-coupled device) cameras or photodiodes to capture light from a scintillation screen when exposed to x-rays, or a direct photoconductor material that produces electric signals.

Any numerical handling of a readily available image is an image enhancement and/or manipulation process. Computed imaging refers here strictly to images produced only after numerical processing of measurements, because the detected radiation, even when digitized, cannot give the required image features. For instance, a tomograph³ can only be obtained by numerical reconstruction using many radiation projections of radiation-transmitted through an object, while a radiograph gives a single projection without numerical reconstruction. Therefore, one speaks of computed tomography (CT) or computer-assisted tomography (CAT) scans, which implies that the image involves both computations and the acquisition of many projections through a scan. Computed imaging is also involved in positron emission tomography (PET), single photon emission tomography (SPECT), as well as scattering-based imaging methods.

³The word tomo is Greek for a slice or a section.

This book provides a framework for the computed imaging problem by presenting it as a forward (direct) problem, and as an inverse problem. The forward problem tackles the physical aspects of imaging via mathematical models that relate the to-be-imaged physical parameters to the collected measurements. The inverse problem reconstructs numerically the imaged parameters from the measurements. This categorization is not new and is used in the solution of the inverse problem of image reconstruction. However, in addressing the inverse problem, the forward problem is often taken as a fait accompli. This is in spite of the fact that the forward problem inherently hides many of the physical simplifications of imaging methods, and contributes to some of the artifacts in the reconstructed images that are not attributable to measurement uncertainties. A good investigation of the forward problem of a particular imaging technique is necessary, not only to provide a foundation for understanding the physics of the technique, but also to help in elucidating the often overlooked limitations of imaging methods. Part I of this book covers the formulation of the forward problem for various radiation imaging modalities. Part II presents a general formulation of the inverse problem.

This book is directed toward graduate students and researchers. Therefore, it takes a critical view of all material presented, so that readers become fully aware of the inherent physical and numerical capabilities and limitations of the methods presented for both the forward and inverse problems. This necessitates an in-depth discussion of the founding physical and mathematical principles upon which the examined methods are based. Parts of this book can be adopted in senior undergraduate courses, by selecting a few chapters of each part of the book, so that the essential fundamentals are covered.

The book is written in a manner that makes it understandable by readers with a background in physics, mathematics, computer science, or engineering. This is done by avoiding, as much as possible, the use of specialized terminology without clear introduction and definition. Mathematicians interested in the numerical aspects of the inverse problem will be exposed to the physical foundations of the forward problem. On the other hand, physicists and engineers should be able to follow the mathematical logic of the presented methods without being bewildered by the jargon of abstract mathematics. By making the material readable by a wide variety of scientists, it is hoped to encourage those outside the imaging research community to involve themselves in medical or industrial imaging, while providing those within the imaging field with a thorough look at the physical and mathematical challenges of the problem.

The author expresses his gratitude to John Bowles of Inversa Systems Ltd. and Esmaeil Enjilela of the University of New Brunswick for reading draft chapters of this book, and providing valuable comments.

1. Radiation Imaging

This Chapter addresses the questions: (1) Why use radiation imaging? Why not any other form of imaging? (2) How is such imaging accomplished? (3) What is reconstructive (computed) imaging? How does it differ from direct (imprint) radiographic imaging? (4) What is a forward problem? What is an inverse problem? (5) Why image reconstruction is an inverse problem? (6) What is inverse mapping? How the inversion process is affected by uncertainties? When and why will the problem be ill-posed?

Keywords: Radiography, Computed Tomography (CT), Tomosynthesis, Emission Tomography, Forward and Inverse Problems/Mappings, Ill-posed problem.

1.1. Why Radiation?

The word radiation is utilized to cover electromagnetic ( em) waves, as well as nuclear particulate radiation (alpha, beta, positrons, and neutrons). The most rudimentary form of imaging is that provided by natural vision, which relies on the deflection of visible light to the eye. Thermal imaging is a form of visualization based on the detection of infrared radiation that emanates from the surface of a hot body. Radar (radio detection and ranging) can also be viewed as a crude imaging process of seeing far-away objects. Radar utilizes em waves in the microwave (above 1 GHz) range and observes their reflection off surfaces. At the extreme end of the radiofrequency band (a few 100 GHz), where the wavelength is in the order of millimeters, electromagnetic radiation can penetrate some organic materials, such as clothing. These so-called millimeter waves are reflected, upon encountering material interfaces, due to the change in the dielectric constant. The intensity of the reflected waves can be used in surface imaging, as done in some body screening security systems.

Imaging with the above methods requires the occurrence of a discontinuity in specific bulk physical properties: light's index of refraction for optical viewing, thermal conductivity and emissivity for thermal imaging, and dielectric constant or electric conductivity in the case of radar. This limits their visualization capabilities to certain applications.

One imaging method that relies on microscopic (rather than bulk) properties is magnetic resonance imaging ( mri). In this method, a powerful magnetic field is applied to align nuclei with non-zero spin) in a particular direction. Some of these nuclei are then temporarily disoriented by radiofrequency ( rf) pulses. As the disturbed nuclei realign, they emit rf signals that are used to determine their location and characterize the surrounding medium. Obviously, if non-zero spin nuclei are present, no imaging information is obtainable. The presence of a magnetic material also interferes with the operation of mri. As such, mri is limited in use to certain materials. This restricts the visualization capability of mri to certain applications. Therefore, mri can also be viewed as an em imaging method that relies on microscopic (rather than bulk) properties.

.

At the high frequency end of the em spectrum, the wavelength of radiation becomes so low that it ceases to be sensitive to bulk properties and begins to be affected by individual atoms. For soft (low energy) x-rays, the wavelength range (in the nanometer range) becomes comparable to the spacing between atoms in a crystallized structure, and the phenomenon of x-ray diffraction begins to occur, providing a useful tool for examining crystal formations. As the frequency increases farther (into the hard x-ray and gamma-ray range), the wavelength becomes so small that the em radiation resembles a quantum of radiation and possess corpuscular capabilities. Each photon quantum becomes like a bullet that moves in the vast vacuum space between atoms, until it encounters an atom with which it can interact. This corpuscular nature of x- and gamma-rays gives them their penetrating properties, and makes them attractive probes for seeing deep within matter. Since this high-frequency radiation interacts with individual atoms, it is not affected by bulk optical, electrical, or thermal properties and can be used for imaging all types of materials.

MHz) are employed for this purpose. Sonic waves are reflected when they encounter an interface between two different materials, due to the change in acoustic impedance (which depends on the speed of sound in the material and the material's density). By sending pulses of ultrasonic waves via a small probe and recording the time of travel and intensity of reflected signals, the position of the reflected surface and the change in acoustic impedance at the surface can be determined. This information is used to construct an image of interfaces as a probe scans an object. Ultrasonic imaging is, therefore, a method for visualizing interfaces, but, when too many interfaces are present, its indications become quite difficult to decipher into an image. It also requires contact between the probe and the surface of the inspected object to provide the acoustic coupling needed for efficient transfer of the ultrasonic waves into and out of the medium. Imaging ocean floors and submerged objects with sonic waves (known as sonar imaging) functions also by monitoring the reflection of sound waves at interfaces.

Nuclear-particle radiation is capable of penetrating matter regardless of its bulk properties. However, charged particles, such as alpha and beta particles, are limited in their penetrability, though highly accelerated charged particles can reach some considerable depth in certain materials. On the other extreme, there are the neutrinos and cosmic rays that are so penetrating that they are not affected much by matter and are difficult to detect. Neutrons, being neutral particles with some mass, are good candidates for imaging, since they are affected by matter and are readily detectable. Therefore, neutron imaging is used in some applications.

1.2. Imaging Modalities

The modality used in imaging with non-ionizing radiation (light and microwaves) and sound (ultrasonic and sonar) is mainly that of scattering (reflection or refraction) of waves bouncing off surfaces and interfaces. Magnetic resonance imaging relies on the emission modality of imaging, as it records emitted secondary rf waves. However, the most common modality of imaging with x-rays, gamma-rays, and neutrons, relies on the transmission modality, by measuring the radiation that succeeds in penetrating the imaged object, passing through it, without interacting with the intervening medium. The dominance of the transmission modality in imaging with penetrating radiation is due to the fact that this radiation interacts with matter at the microscopic (the atomic/nucleus) level. Since most of the matter is made of empty space, with the nuclei and atomic electrons occupying a very small portion of the overall space, most of the radiation passes through matter without interaction. Therefore, the strength of the non-interacting transmitted radiation is quite high, while the interacting radiation that does not succeed in passing through carries information on the nature of the interrogated object. As such, transmitted radiation provides a negative impression of the material it passes through; the denser the material, the weaker the transmitted signal and vice versa. This is the basis of radiography, schematically in Figure 1.1(a), where the internal details of an object are projected by the penetrating radiation on a receiving plate (a film or a scintillating screen). The projected images can be directly interpreted by the viewer in terms of relative material density (dark/light) and their geometric mapping (projection) in the density distribution.

Radiography enables the projection of the internal details of an object on a screen. However, like any other type of projection, it compresses the internal details in the direction of projection, i.e. all material traversed by an incident radiation ray is projected as a single point on the receiving plate. In order to decipher the material information along each ray, many intersecting projections must be taken. Then, with the aid of numerical manipulations, the material attributes at each point (or more precisely, each elementary area) can be determined. This is the essence of transmission tomography, schematically depicted in Figure 1.1(b); more widely know as computed tomography ( ct) or computer-assisted tomography ( cat). The word tomos is Greek for a section or a slice. Therefore, tomographic imaging is usually performed slice-by-slice. The word computed in ct signifies the fact that an image is obtained after performing certain computations. This also makes it necessary to acquire measurements in a digital form, unlike in radiography where analog (e.g. film) recording of the measurements is sufficient. We also speak of ct scans, since the source and the detector assembly has to scan the object at various directions to provide the necessary multiple and overlapping projections; although in some cases, the source and detectors are kept fixed, and the object itself is rotated.

When distinct features are expected in the interrogated object, the effect of intervening material around these features can be reduced by acquiring digital radiography at different source orientations while fixing the receiver's location. A series of image slices, parallel to the imaging plane, can subsequently be numerically reconstructed from the acquired radiographs to depict the distribution of these features at different planes. This is due to the fact that features at different depths in the reflected object are projected differently at various angles of source orientation. This imaging technique is known as digital tomosynthesis (Grant, 1972), or simply tomosynthesis. It is quite attractive for use in mammography, where tumors may be masked by dense breast tissue. The technique is also useful in angiography, chest diagnosis, dental imaging, and orthopaedic imaging of joints (Dobbins III and Godfrey, 2003, Haaker et al., 1985). The concept of tomosynthesis is a numerical form of earlier analog or geometric tomography (also called laminography) systems, which involved a linear and opposing motion of the an x-ray source and a radiographic screen around a pivot to generate an image of the fulcrum plane parallel to the film. A point in the object on the fulcrum plane appears on the same location of the film as the source-film assembly moves, while points above and below this plane appear at different locations and as such their imprint on the film is blurred. In tomosynthesis, this is accomplished numerically by performing a shift-and-add operation of the logarithm of measurements (to linearize the response of transmitted radiation). The shifting process aligns the imprints of distant features on each tomographic plane. Features in the focal (fulcrum) plane then line up exactly on top of each other, so that the addition of the projections enhances their presence. On the other hand, features above or below the fulcrum plane are distributed in the added image to different locations, and their imprints are blurred.

Transmission imaging methods, as Figure 1.1(a) and (b) show, are global in nature; i.e. they image everything in between the source and receiver with the same resolution. Sometimes, however, the resulting image does not portray well particular regions in the object, due to their low density or similarity to the surroundings. If such feeble regions are accessible, a contrast material with higher absorption ability can be introduced into such regions to enhance their presence in the image; as done in the x-ray imaging of the gastrointestinal tract by the introduction of barium liquid. In other cases, particularly for examining physiological functions, a radiation source is introduced as a pharmaceutical, and the emitted radiation is monitored. This modality of emission imaging, schematically shown in Figure 1.1(c), can provide a radiogram by direct viewing of the recorded emission at a particular direction. Alternatively, emissions recorded in a digital form at different directions can be numerically manipulated to provide a tomograph. Single photon emission tomography ( spect) and double photon emission tomography, better known as positron emission tomography ( pet), are the medical forms of this emission imaging modality. While photon (gamma-ray) emitters, such as ⁹⁹mT, ¹²³I, and ¹³³Xe, are used in spect; positron emitters, such as, ¹⁵O, ¹³N, ¹¹C and ¹⁸O, are utilized in pet. Each emitted positron is combined with an electron in the medium, with the two annihilating each other to produce two photons that emanate in opposite directions; hence the double photon emission terminology. The radioisotopes used in emission imaging have short half-lives to enable high decay rate. In this imaging modality, the strength and spatial distribution of the embedded source is deciphered. As radiation from the source travels toward the detector, it is attenuated by the intervening material, which reduces the intensity of the recorded radiation and, in turn, shadows the image. Such attenuation should be accounted for using a priori anatomical information, obtained perhaps from a ct image. While emission tomography is compatible with medical imaging, it is not always easy to embed a radioisotope within an industrial medium. However, if the medium is readily radioactive, as in spent nuclear fuel, emission imaging can be utilized.

The scatter imaging modality is a good candidate for industrial imaging, particularly when the inspected object is too thick or too dense to allow transmission imaging and when emission imaging is not possible. Moreover, scatter imaging can also be utilized when access to opposite sides of the object is not available. Though scatter imaging is the modality of imaging surfaces and interfaces with light, microwaves and ultrasonic waves, scatter imaging with penetrating radiation is much more complex, because radiation experiences no abrupt change as it crosses from one medium to another. The radiation scattering signal can also be quite convoluted, since source radiation can scatter many times before leaving the object and reaching a detector. Nevertheless, imaging based on once-scattered radiation (i.e. single scattering) have emerged. Figure 1.1(d) shows a schematic of a scatter imaging modality. Notice in this arrangement that radiation scattered to a particular direction from a unidirectional source is shown to be detected. This can be done with the aid of a collimated detector and is necessary in order to know from where the detected radiation emerges. Many directions of scattering can be monitored, and various arrangements can be adopted. Scatter imaging can be viewed as a form of emission imaging, with a virtual source at the scattering point. Therefore, scatter imaging produces images of the intensity of scattering, hence material density. This is unlike transmission imaging, which provides an indirect measure of density, since it detects radiation attenuation. The attenuation coefficient is a function of the material density, but can also be affected by material composition, while the scattering coefficient can be made, by proper choice of radiation type and energy, to be a function of only material density. Therefore, a combination of scatter and transmission imaging can provide both density and composition maps. In fact, the motivation for developing scatter imaging systems arose in medical imaging to obtain composition information, which is important for understanding tumor histology, and density maps are needed for planning for radiotherapy, where radiation deposition is strongly density dependent.

1.3. Direct and Reconstructed Imaging

Two types of images can be extracted from any of the imaging modalities shown in Figure 1.1: direct and computed. Direct images are those depicted by the measurements, as in the case of transmission radiography or the radiograms obtained from emission imaging. Direct images obtained from the scattering modality are usually quite convoluted, since in essence the obtained image is modulated by transmission, from the source to the scattering point, and by emission, from the scattering point to the detector. Direct images, when acquired in a digital form, can be supplemented by numerical processing to enhance the image quality, manipulate the image display, store an image in a digital form or render an image remotely accessible. For example, in computed radiography ( cr), the image is acquired in a digital form by extracting light energy stored in a phosphorus plate as result of exposure to x-rays, which is then converted to electrical signals for storage and viewing. On the other hand, in digital radiography ( dr), digital images are directly obtained by capturing light from a scintillation screen with the aid of a ccd (charged-coupled device) camera, photodiodes or a plate made of photoconductor material. In both cr and dr, the image attributes are available in a numerical form, but they are representing measurements of radiation intensity, rather than any physical property of the material in the imaged object. They are still, therefore, direct imprints of the measurements, since any subsequent numerical manipulation does not extract any parameters related to physical properties.

In transmission computed tomography ( ct), measurements are gathered in a digital form then processed to reconstruct the radiation attenuation attributes within each pixel of the section of the image. Then numerical manipulation is more than just simple manipulation of the measurements. Image reconstruction takes a set of measurements and from them produces a set of physical parameters (attenuation coefficients). A similar process occurs in computed emission tomography ( spect and pet), where radiation measurements are used to reconstruct images of the intensity (activity) of the introduced radiation source. Though then, the source's attributes are not direct physical parameters; a source's distribution in a medium depends on the physiological function in the case of living beings, the permeability of soil and rocks, and the fluid flow rate in gases and liquids. In scatter tomography, measurements are used to reconstruct images of the density of the medium (or its scattering coefficient), the attenuation coefficient of the incident radiation, the attenuation coefficient of the scattered radiation or all three physical attributes. Therefore, in all forms of reconstructed imaging, measurements are converted into physical attributes. This requires the availability of a measurement model that mathematically describes how a measurement value is related to the sought physical attributes. Such a measurement model enables the calculation of modeled measurements (synthetic data) corresponding to the actual measurements, for a known set of physical attributes, via the so-called forward mapping. Determining the physical attributes then involves the inversion of the forward model, i.e. inverse mapping of actual measurements, so that modeled measurements match, as much as possible, the actual measurements. These forward/inverse mapping processes are the workhorses of image reconstruction and are examined in further detail in 1.4 and 1.5.

1.4. The Forward and Inverse Problems

In order to decipher a parameterdepicting the physical attributes of a medium from a measurement, of the mercury in its capillary tube; see . This can be provided by developing a mathematical model based on the physical , for mercury:

(1.1)

by:

(1.2)

is the diameter of the capillary tube. The physical attribute is determined by invertingEq. (1.2) to obtain:

(1.3)

The inverse problem of Eq. (1.3) is said to be well-posed.

In the above simple example of the mercury thermometer, we have formulated the forward problem by . To arrive at the forward model of Eq. (1.2), we have relied on the physical theory that defines the volume expansion coefficient, Eq. (1.1), of liquids (mercury in this case) and the geometric relationship between a tube's volume and its height. The inverse problem , some calibration . Accordingly, a more accurate forward model that accounts for the glass expansion takes the form:

(1.4)

This model also illustrates that we will have a singular , the inverse problem becomes ill-posed. Fortunately, most liquids have larger volume expansion coefficients than solids, by two orders of magnitude or