This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Biomedical Image Analysis
Statistical and Variational Methods
Ideal for classroom use and self-study, this book explains the implementation of the mosteffective modern methods in image analysis, covering segmentation, registration, andvisualization, and focusing on the key theories, algorithms and applications that haveemerged from recent progress in computer vision, imaging, and computational biomed-ical science.
* Structured around five core building blocks – signals, systems, image formation, andmodality; stochastic models; computational geometry; level-set methods; and toolsand CAD models – it provides a solid overview of the field.
* Mathematical and statistical topics are presented in a straightforward manner, enablingthe reader to gain a deep understanding of the subject without becoming entangled inmathematical complexities.
* Theory is connected to practical examples in X-ray, ultrasound, nuclear medicine, MRIand CT imaging, removing the abstract nature of the models and assisting readerunderstanding, whilst computer simulations, online course slides, and a solutionmanual provide a complete instructor package.
Aly A. Farag is Professor of Electrical and Computer Engineering, and the foundingDirector of the Computer Vision and Image Processing Laboratory, at the University ofLouisville. His research interests center around object modeling with biomedical appli-cations, and his more recent biomedical inventions have led to the development ofimproved methods for tubular object modeling, virtual colonoscopies, lung noduledetection and classification based on CT scans, real-time monitoring of vital signs fromthermal imaging, and image-based reconstruction of the human jaw. He is a Fellow of theIEEE.
“This is a comprehensive book on the topic of biomedical image analysis. It covers bothstatistical and variational approaches as well as some of the foundations of imageacquisition. The individual chapters and sections also include practical examples, mean-ingful exercises, and computer labs. The book is an outstanding and thorough introduc-tion to the field of biomedical image analysis and is suitable both for classroom use andself-study. Awell curated bibliography provides starting points for additional study.”
University Printing House, Cambridge CB2 8BS, United Kingdom
Cambridge University Press is part of the University of Cambridge.
It furthers the University’s mission by disseminating knowledge in the pursuit ofeducation, learning and research at the highest international levels of excellence.
www.cambridge.orgInformation on this title: www.cambridge.org/9780521196796
This publication is in copyright. Subject to statutory exceptionand to the provisions of relevant collective licensing agreements,no reproduction of any part may take place without the writtenpermission of Cambridge University Press.
First published 2014
Printed in the United Kingdom by TJ International Ltd. Padstow Cornwall
A catalog record for this publication is available from the British Library
Library of Congress Cataloging in Publication dataFarag, Aly A., author.Biomedical image analysis / Aly A. Farag.
Additional resources for this publication at www.cambridge.org/farag
Cambridge University Press has no responsibility for the persistence or accuracy ofURLs for external or third-party internet websites referred to in this publication,and does not guarantee that any content on such websites is, or will remain,accurate or appropriate.
3.6 Magnetic resonance imaging 613.6.1 Signal induction 633.6.2 Relaxation processes 643.6.3 Pulse sequences 653.6.4 Spatial encoding 673.6.5 Tissue contrast in MRI 683.6.6 Components of an MRI system 69
4.2.1 Sample space Ω 804.2.2 Field (algebra) σF 804.2.3 Probability measure P 80
4.3 Random variables 834.3.1 Basic concepts 834.3.2 Properties of the CDF and the PDF of a random variable 874.3.3 The conditional distribution 884.3.4 Statistical expectation 904.3.5 Functions of a random variable 94
4.4 Two random variables 964.4.1 Statistical expectation in two dimensions 994.4.2 Functions of two random variables 1004.4.3 Two functions of two random variables 102
4.5 Simulation of random variables 1034.6 Summary 1044.7 Computer laboratory 1044.8 Exercises 105References 106
5 Random processes 107
5.1 Definition and general concepts 1075.1.1 Description of random processes 1095.1.2 Classification of a random process 1105.1.3 Continuity of a random process 1125.1.4 The Kolmogorov consistency conditions 113
5.2 Distribution functions for a random process 1135.2.1 Definitions 1135.2.2 First- and second-order probability distribution functions 114
5.3 Some properties of a random process 1185.3.1 Stationarity 1185.3.2 The autocorrelation function 1185.3.3 The autocovariance function 1205.3.4 The cross-correlation function 1205.3.5 Time average 1215.3.6 The power spectrum of a random process 1235.3.7 Cross-spectral density 1255.3.8 Power spectral density of discrete-parameter random process 125
5.4 Linear systems with random inputs 1265.5 Two-dimensional random processes 1285.6 Exercises 128References 130
6 Basics of random fields 131
6.1 Introduction 1316.2 Graphical models 1366.3 Markov system 1396.4 Hidden Markov model 1406.5 Markov random field 1416.6 Gibbs model 1436.7 Markov–Gibbs random field models 145
6.7.1 Auto-models 1466.7.2 Aura-based GRF model 1476.7.3 Other models 148
9.2.1 The Harris detector 2179.2.2 The SUSAN corner detector 2199.2.3 Harris–Laplace and Harris–affine corner detectors 2199.2.4 Blob detectors 2219.2.5 Region detectors 223
9.3 Comparative evaluation of interest points 2259.3.1 Multi-scale representations 2259.3.2 Scale-space representation 2329.3.3 Scale-space and feature detection 2339.3.4 Differential singularities and feature detection 234
9.4 Local descriptors 2359.4.1 Scale-invariant feature transform (SIFT) 2359.4.2 Case study: Descriptors of small-size lung nodules
in chest CT 2389.4.3 Extensions to the SIFT algorithms 2399.4.4 Speeded-up robust features (SURF) 2419.4.5 Multi-resolution local binary pattern (LBP) 2419.4.6 Image stitching 245
9.5 Three-dimensional local invariant feature descriptors 2579.5.1 Interest point detection 2579.5.2 3D descriptor building 2619.5.3 Descriptor matching 264
10.1 Calculus of variation and Euler equation 27510.1.1 Euler–Lagrange equation for one independent variable 27610.1.2 Euler–Lagrange equation for multiple independent variables 27710.1.3 Euler–Lagrange and the gradient descent flow 278
10.2 Curve/surface evolution via classical deformable models 27910.2.1 Curves and planar differential geometry 27910.2.2 Geometry of surfaces 28010.2.3 Geodesic curvature 28110.2.4 Principal curvatures 28110.2.5 Planar curves and surface normal 28110.2.6 Curve/surface evolution as a variational problem 28210.2.7 Discretization and numerical simulation of snakes 283
10.3 Level sets 28410.3.1 Implicit representation and the evolution PDE 28410.3.2 Level-set calculus 286
10.4 Numerical methods for level sets 28710.4.1 Conservation law and weak solutions 28710.4.2 Entropy condition and viscosity solutions 28810.4.3 Upwind direction and discontinuous solutions 28810.4.4 The Eulerian formulation and the hyperbolic conservation law 289
10.5 Numerical algorithm 29010.5.1 Need for reinitialization and the distance function 29110.5.2 Front evolution without reinitialization 292
12.4.1 Shape registration 32412.5 Shape-based segmentation 32412.6 Curve/surface modeling by level sets 32612.7 Variational model for evolution-based region statistics 32812.8 Examples and evaluation 329
12.8.1 Performance on images and volumes 32912.8.2 Validation experiment on a real phantom 33112.8.3 Blood vessel extraction 332
12.9 Clinical example: lung nodule segmentation 33412.9.1 Variational approach for nodule segmentation 33612.9.2 Shape alignment 33712.9.3 Level-set segmentation with shape prior 33912.9.4 Some results 33912.9.5 Extensions 341
13.1 Introduction 34513.2 Basic concepts and definitions 346
13.2.1 Components of the registration transformation 34913.2.2 Choice of transformation 35213.2.3 Similarity measures 353
13.3 Surface registration by the ICP algorithm 35513.3.1 Mathematical preliminaries 35513.3.2 The ICP algorithm 359
13.4 Global image registration via mutual information 36613.4.1 Imaging model 36913.4.2 Basics of information theory 37113.4.3 Registration metric 37513.4.4 Mutual information registration 377
14.2.1 Parametric representations 38914.2.2 Landmark-based representation 39014.2.3 Medial axes representation 39114.2.4 Implicit representation using the vector distance function 39214.2.5 Implicit representation using distance transform 392
14.3 Global registration of shapes in implicit spaces 39414.3.1 Global matching of shapes 39414.3.2 VDF-based dissimilarity measure 39714.3.3 SDF-based dissimilarity measure 39814.3.4 Examples 400
14.4 Local shape registration 40314.4.1 Local alignment 40514.4.2 Gradient descent flows and numerical implementation 408
15.2 Statistical shape models 41715.2.1 Construction of statistical shape model using PCA 41915.2.2 Fitting a model to new points 42115.2.3 Statistical modeling of structures 42215.2.4 Modeling shape variations 424
15.3 Statistical appearance models 42815.3.1 Image warping 42815.3.2 One-dimensional thin-plate splines 42915.3.3 N-dimensional thin-plate splines 42915.3.4 Statistical appearance model construction using PCA 43115.3.5 Combined appearance models 433
15.4 Analysis of lung nodules in low-dose CT (LDCT) scans 43615.4.1 Lung nodules in low-dose CT 437
15.5 Appearance-based approach for complete human jaw reconstruction 44115.5.1 Jaw prior models 44415.5.2 Model-based shape and albedo recovery 44515.5.3 Sample results 446
15.6 Summary 448References 448Appendix 15.1 Pseudocodes and MATLAB realizations 450
About two decades ago, I worked with the University of Louisville College of Engineeringand the Office of the Vice President for Research to establish the Computer Vision andImage Processing Laboratory (CVIP Lab – www.cvip.uofl.edu) as a multidisciplinaryenvironment for research, teaching, and training in computational image analysis. Overthe years, the CVIP Lab has been home to researchers in engineering, medicine, dentistry,mathematics, and psychology who are interested in imaging. The support of the Universityof Louisville administration and colleagues at various units literally made the CVIP Lab aplace that I miss whenever I am away from it, even for an enjoyable vacation.
At the CVIP Lab we pushed agendas for imaging research, from basics to applications,and in the process established immediate and auxiliary but essential infrastructure. Amongthe auxiliary infrastructure has been high-speed networking to link the University to whathas become known as Internet 2, an initiative funded by the National Science Foundation,and to link the main campus (Belknap) to the Health Science Campus (HSC) a few milesaway. The auxiliary infrastructure included supercomputers and immersive visualization.The essential hardware included high-end computing and graphics workstations, objectscanners, and various laboratory benches for electronic design and testing. The laboratoryhas been visited by researchers, potential engineering students, faculty candidates inengineering, dentistry and medicine; its research activities have been showcased onnational and local media, and the University President (John Shumaker) and the Deanof Engineering (Thomas Hanley) recorded advertisements there as the University pushedto promote biomedical research, and to establish a biomedical engineering department,during 1996–2002. Today the CVIP Lab is well recognized by colleagues elsewhere.Research at the laboratory has been funded by the National Science Foundation (NSF), theNational Institutes of Health (NIH), the Department of Defense (DoD), the Department ofHomeland Security (DHS), Norton and Jewish Hospitals, and various government andindustrial organizations.
At the CVIP Lab I have had the privilege and pleasure of coaching some of the mostbrilliant students from around the world, and have supervised and hosted a large numberof postdoctoral researchers and researchers who have spent sabbaticals and short visitshere. The laboratory has three main areas of focus: computer vision, biomedical imaging,and biometrics. Students and researchers at the laboratory have worked on theoretical,algorithmic and practical domains of these three focal areas. A number of courses,seminars, and presentations have been created over the years to train researchers andpromote research at the laboratory. This book is one such result of the activities at the
CVIP Lab. It offers both basic background and sample research problems on the subjectof biomedical image analysis.
As the CVIP Lab has been the nest of so many brilliant researchers and students withwhom I have worked with over the years, I will list only a few who have made an impacton me. First and foremost is Dr. Darrel Chenoweth, the Chairman of the ECEDepartmentfrom 1994 to 2004. Darrel gave me unconditional support and encouragement; no wordsare enough to thank him for his impact on me. Thomas Hanley, Dean of J.B. SpeedSchool of Engineering during 1992–2004, was a visionary who gave me freedom to thinkand never hesitated to provide support. Without him and Darrel, the CVIP Lab wouldhave not been established. Dr. Nancy Martin, Vice President for Research during 1996–2006, kept the CVIP Lab on her radar and provided support whenever asked. FormerDean Mickey Wilhelm and current Dean Neville Pinto have maintained this trend ofsupport, as did Dr. James Graham, ECE Department Chairman during 2006–2013.
More than 50 colleagues have collaborated with me at the CVIP Lab, so I will justmention a few: Dr. Christopher Shields, former Chairman of Neurological Surgery,Dr. Allan Farman, Professor of Dental Radiology, Dr. Thomas Starr, Associate Deanfor Research, Dr. Manuel Casanova, Professor of Psychiatry, Dr. Edward Essock,Professor of Psychology and Brain Sciences, and Dr. Robert Falk, Director of MedicalImaging at Jewish Hospital, have been longstanding collaborators and friends withwhom I have enjoyed working, and their support has been crucial to whatever I haveachieved at the CVIP Lab.
Three scientists and engineers of the highest caliber and professionalism – CharlesSites, Mike Miller and Salwa Elshazly –made me anxious to come to the CVIP Lab; andif I go away, I can trust that it is safe in their hands. Chuck is a visionary; he and I wroteproposals that brought high-speed networking, computing, visualization and autono-mous robotics to the University of Louisville. We have worked together since 1996.Salwa agreed to assist me at the Laboratory in 1997 and for 15 years has providedintellectual support and coordination of efforts that have been crucial for success at thelaboratory. Mike joined the laboratory in 2006 after two decades of work in the industry.He showed unbounded dedication and rare talent in almost every aspect of computingand circuit design, and has handled the university regulations for biomedical data. Mikehas been my right hand in student advising, documentation of research, and communi-cations with the university as well as funding agencies. Many other technical staff havehelped the CVIP Lab; too many to list, but I appreciate all their efforts and assistance.
As I mentioned before, I have been privileged to coach some of the most brilliantpeople from around the world for their Ph.D. andMasters research. They have worked onmy funded projects and have excelled in executing the research plans and expandingthem to frontiers that I could not imagine, or perform, alone. I owe a great deal ofappreciation and thanks to each of them. In this book, I must acknowledge the following:Dr. Mohamed Sabry, Dr. Hossam Abdelmunim, Dr. Asem Ali, Dr. Rachid Fami,Dr. Shireen Elabian, Dr. Amal Farag, Dr. Ham Rara, Dr. Melih Aslan, Mr. AhmedShably, Dr. Mostafa Abdelrahman, Ms. Marwa Ismail, and Dr. Aly Abdelrahim. These12 individuals have had a direct impact on this book and I owe them my deepestappreciation. In particular, Dr. Shireen Elhabian and Dr. Ahmed Shalby have shown
tenacity, intelligence and dedication in assisting me throughout the preparation of thisbook; I remain very grateful to both of them.
Funding from various organizations and support of the University of Louisville aregratefully acknowledged, as is a long stream of local, national and international collab-orators with whom I have had the honor and pleasure to collaborate and interact overthree decades.
I must state the obvious: all errors and mishaps in the book are mine. I shall be gratefulfor any hints from readers that might assist in improving the text in revised prints or neweditions. Together with Cambridge University Press, I have a website for auxiliarymaterial including teaching aids, newer homework problems and laboratories, solutionsto problems, and codes for the implementations in the book.
Michelle Carey and Elizabeth Horne at Cambridge University Press have providedencouragement throughout this project, Lindsay Nightingale provided a most skilledreview of the manuscript, and Christina Sarigiannidou managed the book production.They were very patient with me and worked around my schedule, despite my endlessobligations to the CVIP Lab and derailment by circumstances, not the least of which hasbeen the engagement of mymind and soul with the events in my beloved home country ofEgypt. Since 2011, countless hours of thought have been spent engaging with mycompatriots, family members and officials, pushing for the common good and towardspeaceful democratic changes in a country that the entire world wishes to see peaceful andprosperous. I thank Michelle, Elizabeth, Lindsay and Christina for their help, and repeatmy highest appreciation to the personmost deserving of thanks and appreciation, my dearwife, collaborator, and friend, Salwa A. Elshazly.
The following conventions are used throughout the document.
Symbol Description
x Point in 2D or 3D Cartesian spaceℝ Set of real numbersΩ An open bounded subset of ℝΩ\ Ω0 Complement of Ω0 in Ω∈ Element of⊆ Subset of|·| Absolute value in ℝ
||·|| Euclidean norm in a vector spaceC A curve or family of curves in R
2
ϕ Level set functionΦS Implicit representation of a given shape S~ϕ Implicit representation of a shape priorL(.) Labeling functionA Rigid or affine transformation in R
n
S Scale matrixR Rotation matrixθ Rotation angleT Translation vectoru ¼ ðuiÞ1 ≤ i ≤ n Displacement field in R
n
F Speed functionV Vector distance functiondiv(·) Divergence of a vector fieldD(·) Dissimilarity measureDMl(·) Mutual information dissimilarity measureDSSD(·) Sum of squared differences dissimilarity measureR(·) Regularization termM Manifold in R
δ(·) Dirac functionH(·) Heaviside functionαi Weight of shape energyD Space dimension, degrees of freedomW(.) Transform on R
D
t Translation vectorL(.) Linear transform on R
D
L Matrix form for a linear transform L(.) on R3 with matrix elements ℓij
s Scaling parameter in the case of uniform scalingsx; sy; sz Scaling parameters in the case of non-uniform scalingS Matrix form for a scale transform on R
3 with scale parameter(s) s∈R in thecase of uniform scaling and sx; sy; sz ∈R in the case of non-uniform scaling
tx; ty; tz Translation parametersT Matrix form for a translation transform on R
3 with translation parameterstx; ty; tz ∈R
Rx;α Matrix form for a rotation transform onℝ2\ℝ3 about the x-axis by angle ofrotation α
(α, β, γ) Euler rotation anglesE(α, β, γ) Euler transformv (Eigen)vector in R
D
A Square transformation matrix defined on RD�D
λ Eigenvalue of matrix Aj: j Matrix determinantr,q A quaternionq0; q1; q2; q3 Elements of a quaternion qVq An ordinary vector defining the complex part of a quaternion qqx; qy; qz The imaginary parts of a quaternion qi; j; k Basis defining the imaginary parts of a quaternionR Orthogonal matrix defining the multiplication of two quaternionsR Same as R, except that the lower right-hand 3×3 submatrix is transposedq� Conjugate of a quaternion qI 4×4 identity matrixM Model shape represented by a set of points fmigNm Number of points in the model shapeP Model shape represented by a set of points fpigNp Number of points in the scene shapeR(.) Rotation operatord(p, m) Euclidean distance between two points p;m∈R
3
d(pi, M) Euclidean distance between a scene point pi and the model point set MY Closest point in the model set which yields the minimum distanceC The closest point operatorei Residual error for each point pairE Sum of squares of the residual error for each point pair
μP Centroids/origins of the shape defined by the point set Ppʹi Centered (zero-means) pointstʹ Translation vector of the centered pointsSy, Sp Sums of the squares of the distances between points and their centroidsS The S-matrix whose elements are sums of products of coordinates
measured in the scene shape with coordinates measured in the modelshape
pðxkÞ Probability of the outcome xkI(xk) Information measure of the outcome xkH(X) Entropy of the random variable XH(X,Y) Joint entropy of the random variables X and Ypðxi; yjÞ Joint probability density function of the two random variables, i.e. the
probability of having both outcomes xi and yi occur togetherHðX jY ¼ yjÞ Conditional entropy of X given Y = yjHðX jY Þ Conditional entropy of X given YI(X,Y) Mutual information between X and YpX ðxÞ Marginal probability mass function of the random variable XpXY ðx; yÞ Joint probability mass function of the random variables X and YTΘ A transformation with registration parameters ΘR Random variable denoting the intensities observed in the reference volumeF Random variable denoting the intensities observed in the floating volumeTΘF Random variable denoting the intensities observed in the transformed