Computational Anatomy & Physiology M2 BioComp 2013-2014 http://www-sop.inria.fr/teams/asclepios/cours/CBB-2013-2014/
X. Pennec
Medical Image registration
Asclepios team 2004, route des Lucioles B.P. 93 06902 Sophia Antipolis Cedex
http://www-sop.inria.fr/asclepios 1
Goals of Registration A dual problem
Find the point y of image J which is corresponding (homologous) to each points x of image I.
Determine the best transformation T that superimposes homologous points
T
yk xk
J I
2
Principal Applications
Fusion of multimodal images
Temporal evolution of a pathology
Inter-subject comparisons
Superposition of an atlas
3
1 Fusion of Multimodal Images
MRI PET X-scan
Histology US Visible Man
4
Temporal Evolution
coronal coronal
Time 1 Time 2
sagittal axial sagittal axial
5
Inter-Subject comparison
6
2 Registration to an Atlas
Voxel Man
7
Augmented reality Brigham & Women ’s Hospital
E. Grimson
8
Classes of problems vs. applications
Temporal evolution Intra Subject – Monomodal
Multimodal image fusion Intra Subject - Multimodal
Inter-subject comparison Inter Subject - Monomodal
Superposition on an atlas Inter Subject - Multimodal
Intra Subject: Rigid or deformable
Inter Subject: deformable
9
3 Intuitive Example How to register these two images?
10
Feature-based/Image-based approach
Feature detection (here, points of high curvature) 2 Measure: for instance S(T) T (xk ) y k k
11
Feature-based/Image-based approach
Feature detection (here, points of high curvature) 2 Measure: for instance S(T) T (xk ) y k k
12
4 Feature-based/Image-based approach
No segmentation! 2 S(T ) (i j ) Interpolation: j J(T(x )) Measure: e.g. k k k k k
T
jk ik xk T(xk )
13
Feature-based/Image-based approach
No segmentation! 2 Measure: e.g. S(T ) (ik jk ) k
T1 =Id
14
Feature-based/Image-based approach
No segmentation! 2 Partial overlap Measure: e.g. S(T ) (ik jk ) k
T2
15
5 Classes of Transformations T
Rigid (displacement)
Similarities
Affine (projective for 2D / 3D)
Polynomials
Splines
Free-form deformations
16
Classes of Transformations T Rigid: T (x) Rx t
Rotation and translation
6 parameters: (R: 3; t: 3)
invariants: distances (isometry), frame orientation, curvatures, angles, straight lines
Similarities: T (x) s.Rx t
Add a global scale factor
7 parameters
invariants: ratio of distances, orientation, angles, straight lines
17
Classes of Transformations T
Affine: T (x) Bx t 3x3 matrix B
12 parameters: (B: 9; t: 3)
invariants: straight lines, parallelism
Quadratic: T (x) xt Ax Bx t Add a symmetric 3x3 matrix A
18 parameters (A: 6; B: 9; t: 3)
invariants: do not preserve straight lines any more
18
6 Classes of Transformations T Splines:
Local polynomials of degree d, with a global continuity of degree C(d-1).
number of parameters: depend on the number of control points (knots)
locally affine: simplified version
Free form transformations: T (x) x u(x)
a vector u(x) is attached to each point x
parameters: at most 3 times the number of voxels
regularization: constrain to homeomorphisms (diffeomorphisms)
19
Classification of registration problems
Type of transformation
Parametric Rigid (displacement), similarity, affine, projective
Deformables Polynomial, spline, free-form deformations Type of acquisition
Monomodal
Multimodal Homology of observed objects
Intra-subject (generally a well posed problem)
Inter-subject (one-to-one correspondences, regularization ?)
20
Mathematical Formulation of registration (Brown, 1992)
Registration: Given two datasets (images) I and J, find the geometric transformation T that « best » aligns the physically homologous points (voxels)
Tˆ arg max S(I, J,T ) TΤ
Optimization algorithm Similarity measure
Transformation space (rigid, affine, elastic,…)
21 21
7 Geometric methods Extract geometric features
Invariant by the chosen transformations
Points
Segments
Frames
Given two sets of features, registration consists in:
Feature identification (similarity): Match homologous features
Localization: Estimate the transformation T
Algorithms
Interpretation trees
Alignement
Geometric Hashing
ICP
22
Artificial markers
Stereotactic frame
Invasive
External markers
Motion Short time use
23
Anatomical markers
Find geometric invariants to characterize a small number of singular points on anatomical surfaces
Generalization of edges and corner points to differentiable surfaces
24
8 Iterative closest point (ICP)
One global criterion, C(A,T) dist(T(P), A(P )) 2 i i i alternatively minimized over
* 2 Step 1: matches A Argmin dist(T(Pi ), A(Pi )) A i * 2 Step 2: transformation T Argmin dist(T(Pi ), A(Pi )) T i Positive and decreasing criterion: convergence Robustness w.r.t. outliers: robust distances
2 (x) || x || (standard mean ) 2 2 2 dist(x, y) (|| x y || ) (x) min || x || , (saturated mean ) (x) || x || (median)
25
1. Optimization for matches
Qj Pi
26
2. Optimization for T
Qj Pi
27
9 1-bis Optimization for matches
Qj Pi
28
2-bis Optimization for T
Qj Pi
29
Least squares
Problem:
We have n matched points ( x i , y i ) and we want to minimize: 2 C(A,t) yi A xi t C i 2(yi A xi t) 0 t i
topt y A x
In barycentric coordinates ~x x x i i ~ ~ 2 ~ C(A,topt ) C(A) yi Axi yi yi y i
30
10 Least squares affine transformation
1 ~ ~ 2 t t C(A) yi Axi Tr( yy ) Tr(A xx A ) 2Tr(A. yx ) N i
1 ~ ~ t 1 ~ ~ t t 1 ~ ~ t xx xi xi , yy yi yi , yx xy xi yi N i N i N i 1 Optimum: V C(A) lim C(A V ) C(A) 0 V 0 t t V C(A) 2Tr(A. xx .V ) Tr( yx .V ) 0 =Tr(A.Bt) Froebenius dot product on the space of matrices: B,V 0 V B 0
(1) V C(A) 0 A.xx K A yx .xx
31
Augmented reality guided radio-frequency tumor ablation Current operative setup at IRCAD (Strasbourg, France)
Per-operative CT “guidance”
Respiratory gating Marker based 3D/2D rigid registration
S. Nicolau, X.Pennec, A. Garcia,L. Soler, N. Ayache
32
Increase 3D/2D registration accuracy: A new Extended Projective Point Criterion
M N 2 Pl(T*M ) ml Standard criterion: i i l1 i1
image space minimization (ISPPC)
noise only on 2D data
Complete statistical assumptions + ML estimation
Gaussian noise on 2D and 3D data
Hidden variables Mi (exact 3D positions) 2 ~ 2 M N l ~ l N P (T * M i ) mi M i M i 2 2 l1 i1 2 2D i1 2 3D
33
11 Registration for Augmented reality
34
Whole loop accuracy evaluation Stereoscopic HD video acquisition Left monitor with AR Right monitor with AR & needle tracking & needle tracking
Real scene: phantom + needle
35
Whole loop accuracy evaluation
Left monitor with AR Right monitor with AR & needle tracking & needle tracking
Endoscopic control
Real scene: phantom + needle
36
12 Whole loop accuracy evaluation
37
Liver puncture guidance using augmented reality S. Nicolau, IRCAD / INRIA 3D (CT) / 2D (Video) registration
2D-3D EM-ICP on fiducial markers
Certified accuracy in real time Validation
Bronze standard (no gold-standard)
Phantom in the operating room (2 mm)
10 Patient (passive mode): < 5mm (apnea)
[ S. Nicolau, PhD’04 MICCAI05, ECCV04,, IS4TM03, Comp. Anim. & Virtual World 2005 ]
39
Geometric methods
Advantages
automatic (no initialization, except ICP)
fast (after the feature segmentation)
robust to outliers
Drawbacks
Requires a high resolution and a low noise
Invariant features: Rigid, monomodal, same subject
Possible exception: skull images
41
13 Intensity-based methods
No geometric feature extraction
Advantages:
Noisy images and/or low resolution
Multimodal images
Drawbacks:
All voxels must be taken into account
42
Compare multi-modal images?
Which similarity criterion ?
Many available criteria:
SSD, Correlation, Mutual Information...?
Variable costs and performances
Where is the optimum ?
Maintz & Viergever, Survey of Registration Methods, Medical Image Analysis 1997
44
Course overview
Introduction
Multimodal Intensity-based Registration
Classification of existing measures
A general framework
Deformable intensity-based Registration
Some extensions
Conclusion
45
14 Classification of existing measures Assumed relationship
Intensity conservation Intensity o ’image I
Intensity of image J Adapted measures
Sum of Square Differences 2 S(T ) (ik jk ) Sum of Absolute value of Differences k Measures of intensity differences (Buzug, 97)
Interpolation: jk J(T(xk ))
46
Classification of existing measures Assumed relationship
Affine Intensity of image I
Intensity of image J Adapted measures
Correlation coefficient
1 IJ (T ) (ik I )( jk J ) n I J k
47
Classification of existing measures Assumed relationship
Functional Intensity of image I
Intensity of image J Adapted measures Woods’ criterion (1993) Woods’ variants (Ardekani, 95; Alpert, 96; Nikou, 97) Correlation ratio (Roche, 98) Var[E(I | J (T ))] 2 Var(I)
48
15 Classification of existing measures Assumed relationship
Statistical Intensity of image I
Intensity of image J Adapted measures Joint Entropy (Hill, 95; Collignon, 95) Mutual Information (Collignon, 95; Viola, 95) Normalized Mutual Information (Studholme, 98) P(i, j) MI (I, J ) H (I) H (J ) H (I, J ) P(i, j)log ij P(i)P( j)
49
Course overview
Introduction
Multimodal Intensity-based Registration
Classification of existing measures
A general framework
Deformable intensity-based Registration
Some extensions
Conclusion
51
A general framework
A. Roche proposed a unifying maximum likelihood framework
Physical and statistical modeling of the image acquisition process
Create a hierarchy of criteria
A. Roche, G. Malandain and N.A. : Unifying maximum likelihood approaches in medical image registration. International Journal of Imaging Systems and Technology : Special Issue on 3D Imaging 11(1), 71-80, 2000.
• Based on pioneer works of (Costa et al, 1993), (Viola, 1995), (Leventon & Grimson, 1998), (Bansal et al, 1998)
52
16 Principle of the method [A. Roche]
Generative model of images
– probabilistic – parametric T I Spatial Acquisition Image I S transfo. model Anatomical ScèneScene S prior J Acquisition Image J model Maximum likelihood Inference
max P(I, J |T, I , J , S ) T , I , J ,S Auxiliary variables: q
53
Generic model
Scene: discrete random field (segmentation)
Assumptions: – Spatial independence – Stationarity
P(S) (sl ) l
54
Generic model
Images: noisy measures of the scene
i f (s ) s s(T (x )) k k k k k jl g(sl ) l
Assumptions on and : – white (spatial indep.) – Stationary P(I, J | S) G (i f (s )) ε k k – Gaussian k G ( j g(s )) – Additive η l l – Independent of each other l
55
17 Likelihood function
Likelihood = joint law of images P(I, J ) P(I, J | S)P(S)dS
L(T, ) with ( , f , g, ε , η )
Log likelihood
P (i , j ) log L(T, ) log k k logP (i ) logP ( j ) k l kA P (ik )P ( jk ) k l (T, ) ( ) K with P (i, j) G (i f )G ( j g ) p ε p η p p1
56
Example: X-Scan / MRI rigid registration
The estimation of allows the a posteriori estimation of the scene
57
Limitations
In general, the maximization w.r.t. is costly (EM segmentation algorithm)
With additional assumptions on , one can approach the solution analytically
58
18 Particular case 1 Assumptions
f g ik jk k | | | | T Spatial Gaussian Image J transfo noise Image I
Maximum Likelihood
c 2 max L(T, ) nlog (ik jk ) Ecte 2 n k SSD
59
Particular case 2 Assumptions
g injective ik f ( jk ) k , f constant by part | | | | T f Spatial Intensity Gaussian Image J Transfo. function noise Image I
Maximum Likelihood
c 2 max L(T, ) nlog min (ik f ( jk )) Ecte 2 f n k Squared fit error
60
Particular case 2
Limit case: correlation ratio (CR)
Squared fit error
2 min (ik f ( jk )) f 2 (T) k I|J n Var(I) Different normalization of the likelihood criteria
Equivalence ML / CR
62
19 Generalization of the Correlation Ratio
Robust metric
min (I f (J T )) gene (T) f : M-estimator of scale I|J min (I ) (Rousseeuw, 87) cte
Space in which one seeks the function f
polynomials, splines, ... Intensity in in Intensity I Intensity in J
63
Particular case 3
Assumptions
Known
Spatial Gaussian transfo. noise Image I Discrete field Scene S
Gaussian Image J noise Unknown number of classes Known
64
Particular case 3
Maximum Likelihood
1 Pˆ(i, j) G (i i )G ( j j ) (Parzen) ε k η k n k
Pˆ(i , j ) max L(T, P) log k k corrective terms P ˆ ˆ k P(ik )P( jk ) n Mutual Informatio n
65
20 Choice of a criterion
Choosing a criterion imposes to deeply understand the physical image acquisition process
Whenever several models are known, choosing the one with the smallest number of degrees of freedom increase the robustness of the approach.
66
Roboscope
MR coordinate system US coordinate system
TMR/US TMR0 / MR1
Robot / patient coordinate system MR 1 US 1 over time over MR 0 with surgical plan
US 2 Virtual MR 2 deformation
• Rigid matching tools for MR Brain • MR / US registration • US tracking of brain US n deformations Virtual MR n
67
Manipulator
Steady Hand Motion Compensation Active Motion Constraints
QuickTime™ and a Intel Indeo® Video 5.0 decompressor are needed to see this picture.
Courtesy B. Davies & S. Starkie
68
21 Manipulator
Steady Hand Motion Compensation Active Motion Constraints
QuickTime™ and a Intel Indeo® Video 5.0 decompressor are needed to see this picture.
Courtesy B. Davies & S. Starkie
69
MR-US Images
Pre - Operative MR Image Per - Operative US Image axial axial
coronal sagittal coronal sagittal Acquisition of images : L. & D. Auer, M. Rudolf
70
Ultrasound image / MRI registration
Elementary principles of US imagery
Irf (x) Z.u (x) (x) x Z Logarithmic u compression
I(x) Alog Z.u (x) B (x)
US MRI Grad MRI 71
22 Ultrasound image / MRI registration
Assumption: acoustic impedance is a function of the MR signal (denote by J)
Z(x) g(J (x)) Z(x) g(J (x))J (x)
Relation between US and MR signals
I(x) f J (x), J.u (x) (x)
In practice, the influence of I(x) f J(x), J (x) (x) orientation is neglected
72
Bivariate Correlation Ratio
I function of 2 variables
I f (J,| J |)
2 iterated stages Dependency Hypothesis Robust polynomial approx. of f
Estimation of T:
ˆ ˆ 2 T arg min ( I(xk ) f (J (T (xk ), J (T (xk ) ) ) T k A. Roche, X. Pennec, G. Malandain, and N. A. Rigid Registration of 3D Ultrasound with MR Images: a New Approach Combining Intensity and Gradient Information. IEEE Transactions on Medical Imaging, 20(10):1038--1049, October 2001.
73
Typical Registration Result with Bivariate Correlation Ratio
Pre - Operative MR Image Per - Operative US Image axial axial
Registered
coronal sagittal coronal sagittal Acquisition of images : L. & D. Auer, M. Rudolf
74
23 US Intensity MR Intensity and Gradient
75
77
24