Centrale OMA MIA -MVA 2017-2018 - · 1 1 Statistic on Riemannian manifolds and Lie groups...
Transcript of Centrale OMA MIA -MVA 2017-2018 - · 1 1 Statistic on Riemannian manifolds and Lie groups...
1
1
Statistic on Riemannian manifolds and Lie groups
Computing with diffusion tensor images
X. Pennec
Epione team
2004, route des Lucioles B.P. 93
06902 Sophia Antipolis Cedex
https://team.inria.fr/epione/
Statistical computing on manifolds and data assimilation:
from medical images to anatomical and physiological models
http://www-sop.inria.fr/asclepios/cours/MVA/Module2/index.htm
Centrale OMA MIA - MVA 2017-2018
MVA 01-2018
EPIONE
2
Statistical computing on manifolds and data assimilation: from medical images to anatomical and physiological models
Medical Imaging – Module 2 - MVA 2017-2018
Thu Jan 11: Introduction to Medical Image Analysis and Processing [XP]
Thu Jan 18: Riemannian manifolds and Lie groups [XP]
Thu Jan 25: Biomechanical Modelling [HD]
Thu Feb 1: Statistics & Computing with diffusion tensor images [XP]
Thu Feb 8: Data assimilation methods for physiological modeling [HD]
Thu Mar 1: Diffeomorphic transformations for image registration [XP]
Thu Mar 8: Statistics on deformations for computational anatomy [XP]
Thu Mar 15: Personalization of tumor growth models [XP]
Thu Mar 22: Exam [HD,XP]
MVA 01-2018
3
Bases of Algorithms in Riemannian Manifolds
Operation Euclidean space Riemannian
Subtraction
Addition
Distance
Gradient descent )( ttt xCxx
)(yLogxy xxyxy
xyyx ),(distx
xyyx ),(dist
)(xyExpy x
))( ( txt xCExpxt
xyxy
Reformulate algorithms with expx and logx
Vector -> Bi-point (no more equivalence classes)
Exponential map (Normal coordinate system): Expx = geodesic shooting parameterized by the initial tangent
Logx = unfolding the manifold in the tangent space along geodesics
Geodesics = straight lines with Euclidean distance
Local global domain: star-shaped, limited by the cut-locus
Covers all the manifold if geodesically complete
MVA 01-2018
2
4
Statistical Computing on Manifolds for Computational Anatomy
Simple Statistics on Riemannian Manifolds
Mean, Covariance, Parametric distributions / tests
Statistics on the spine
Manifold-Valued Image Processing
Introduction to diffusion tensor imaging
Interpolation, filtering, extrapolation of tensors
Other metrics and applications
MVA 01-2018
5
Statistics on Riemannian manifolds
The geometric framework
Simple statistical tools Mean, Covariance, Parametric distributions / tests
Statistics on the spine
Application examples Evaluation of registration performances
Rigid body transformations
MVA 01-2018
6
Basic probabilities and statistics
Measure: random vector x of pdf
Approximation:
• Mean:
• Covariance:
Propagation:
Noise model: additive, Gaussian...
Principal component analysis
Statistical distance: Mahalanobis and
dzzpz ).(. ) E(x xx
)x( xxΣx , ~
)(zpx
T)x).(x(E xxxx
x
..x
, x)(Thh
h ~ h xxΣxy
2
MVA 01-2018
3
Random variable in a Riemannian Manifold
Intrinsic pdf of x
For every set H
∈
Lebesgue’s measure
Uniform Riemannian Mesure det
Expectation of an observable in M
(variance) : . , ,
log (information) : log log
(mean) :
8MVA 01-2018
9
Fréchet expectation (1944)
Minimizing the variance
Existence
Finite Variance at one point
Variational characterization: exponential barycenter (P(C)=0)
The case of points: classical expectation
Other central primitives
),dist(E argmin 2xx yy M
Ε
0)().(.xxE 0 )(grad 2 M
M zdzpy xx xx
0xE x xxΕ
1
),dist(E argmin xx yy M
Ε
[Oller & Corcuera 95, Battacharya & Patrangenaru 2002, Pennec, JMIV06, NSIP’99 ]
MVA 01-2018
Riemannian center of mass,Fréchet or Karcher mean?
Name conventions in geometric statistics Fréchet mean: global minima [Fréchet 1944]
Local Fréchet or Karcher mean: local minima [Karcher 1977, Kendall 1990]
Exponential barycenters: critical points [Emery 1991, Oller & Corcuera 1995]
All are sometimes called Riemannian center of mass [Groove & Karcher 1973-1976] under uniqueness assumptions
Concentration assumption: Existence and uniqueness[Karcher 77 / Buser & Karcher 1981 / Kendall 90 / Afsari 10 / Le 11]
Support of distribution in a regular geodesic ball U(r) with radius ∗ min , / ( upper bound on sect. Curv. on M with convention
that ∞ 0)
Empirical Fréchet mean (finite sample): Almost surely unique [Arnaudon & Miclo 2013]
10MVA 01-2018
4
11
A gradient descent (Gauss-Newton) algorithm
Vector space
Manifold
vHvvfxfvxf fTT
..2
1.)()(
fHvvxx ftt
. with )1(
1
),()()())((exp2
1 vvHvfxfvf fx
iin
yx2
yE 2 (y)2 xx
IdH 2 2 x
xyE with )(expx x1 vvtt
Geodesic marching
MVA 01-2018
Algorithms to compute the mean
Karcher flow (gradient descent)
exp E y ∑ log
Usual algorithm with 1 can diverge on SPD matrices [Bini & Iannazzo, Linear Algebra Appl., 438:4, 2013]
Convergence for non-negative curvature (p-means) [Afsari, Tron and Vidal, SICON 2013]
Inductive / incremental weighted means
exp log
On negatively curved spaces [Sturm 2003], BHV centroid [Billera, Holmes, Vogtmann, 2001]
On non-positive spaces[G. Cheng, J. Ho, H. Salehian, B. C. Vemuri 2016]
Stochastic algorithm [Arnaudon & Miclo, Stoch. Processes and App. 124, 2014]
12MVA 01-2018
13
First Statistical Tools: Moments
Covariance (PCA) [higher moments]
Principal component analysis
Tangent-PCA:principal modes of the covariance
Principal Geodesic Analysis (PGA) [Fletcher 2004,Sommer 2014]
M
M )().(.x.xx.xE TT
zdzpzz xxx xx
[Oller & Corcuera 95, Battacharya & Patrangenaru 2002, Pennec, JMIV06, NSIP’99 ]
MVA 01-2018
5
PCA vs PGA
PCA find the subspace that best explains the variance
Maximize the squared distance to the mean
Generative model: Gaussian
PGA (Fletcher, Joshi, Lu, Pizer, MMBIA 2004) find a low dimensional sub-manifold generated by geodesic
subspaces that best explain measurements
Minimize the squared Riemannian distance from the measurements to that sub-manifold (no closed form)
Generative model: Implicit uniform distribution within the subspace
Gaussian distribution in the vertical space
Different models in curved spaces (no Pythagore thm)
14MVA 01-2018
15
Example with 3D rotations
Principal chart:
Distance:
Frechet mean:
nr . :ectorrotation v
2)1(
121 ),dist( rrRR
Centered chart:
mean = barycenter
i
2 ),(distmin arg 3
iSOR
RRR
MVA 01-2018
16
Distributions for parametric tests
Uniform density: maximal entropy knowing X
Generalization of the Gaussian density: Stochastic heat kernel p(x,y,t) [complex time dependency] Wrapped Gaussian [Infinite series difficult to compute] Maximal entropy knowing the mean and the covariance
Mahalanobis D2 distance / test:
Any distribution:
Gaussian:
2/x..xexp.)(
TxΓxkyN
)Vol(/)(Ind)( Xzzp Xx
rOk n /1.)det(.2 32/12/ Σ
rO / Ric3
1)1( ΣΓ
yx..yx)y( )1(2 xxx
t
n)(E 2 xx
rOn /)()( 322 xx
[ Pennec, JMIV06, NSIP’99 ]
MVA 01-2018
6
17
Gaussian on the circle
Exponential chart:
Gaussian: truncated standard Gaussian
[. ; .] rrrx
standard Gaussian(Ricci curvature → 0)
uniform pdf with
(compact manifolds)
Dirac
:r
:
:03/).( 22 r
MVA 01-2018
19
Computing on manifolds: a summary
The Riemannian metric easily gives Intrinsic measure and probability density functions
Expectation of a function from M into R (variance, entropy)
Integral or sum in M: minimize an intrinsic functional Fréchet / Karcher mean: minimize the variance
Filtering, convolution: weighted means
Gaussian distribution: maximize the conditional entropy
The exponential chart corrects for the curvature at the
reference point Gradient descent: geodesic walking
Covariance and higher order moments
Laplace Beltrami for free
[ Pennec, NSIP’99, JMIV 2006, Pennec et al, IJCV 66(1) 2006, Arsigny, PhD 2006]
MVA 01-2018
20
Statistical Analysis of the Scoliotic Spine
Database 307 Scoliotic patients from the Montreal’s
Sainte-Justine Hospital.
3D Geometry from multi-planar X-rays
Mean Main translation variability is axial (growth?)
Main rotation var. around anterior-posterior axis
PCA of the Covariance 4 first variation modes have clinical meaning
[ J. Boisvert, X. Pennec, N. Ayache, H. Labelle, F. Cheriet,, ISBI’06 ]
MVA 01-2018
7
21
Statistical Analysis of the Scoliotic Spine
• Mode 1: King’s class I or III
• Mode 2: King’s class I, II, III
• Mode 3: King’s class IV + V
• Mode 4: King’s class V (+II)
MVA 01-2018
32
Per - Operative US Image
Typical Registration Resultwith Bivariate Correlation Ratio
Pre - Operative MR Image
Registered
Acquisition of images : L. & D. Auer, M. Rudolf
axial
coronal sagittal
axial
coronal sagittal
MVA 01-2018
33
US Intensity MR Intensity and Gradient
MVA 01-2018
8
34
Accuracy Evaluation (Consistency)
222/
2 2 USMRUSMRloop
MVA 01-2018
35
Validation using Bronze Standard
Best explanation of the observations (ML) : LSQ criterion
Robust Fréchet mean
Robust initialization and Newton gradient descent
Grid scheduling for efficiency
Result
221
221
2 ),,(min),( TTTTd
transrotjiT ,,,
ij ijij TTdC )ˆ,(2
[ T. Glatard & al, MICCAI 2006,
Int. Journal of HPC Apps, 2006 ]
MVA 01-2018
36
Data (per-operative US) 2 pre-op MR (0.9 x 0.9 x 1.1 mm)
3 per-op US (0.63 and 0.95 mm)
3 loops
Robustness and precision
Consistency of BCR
Results on per-operative patient images
Success var rot (deg) var trans (mm)MI 29% 0.53 0.25CR 90% 0.45 0.17
BCR 85% 0.39 0.11
var rot (deg) var trans (mm) var test (mm)Multiple MR 0.06 0.06 0.10
Loop 2.22 0.82 2.33MR/US 1.57 0.58 1.65
[Roche et al, TMI 20(10), 2001 ]
[Pennec et al, Multi-Sensor Image Fusion, Chap. 4, CRC Press, 2005]
MVA 01-2018
9
40
Statistical Computing on Manifolds for Computational Anatomy
Simple Statistics on Riemannian Manifolds
Manifold-Valued Image Processing
Introduction to diffusion tensor imaging
Interpolation, filtering, extrapolation of tensors
Other metrics and applications
MVA 01-2018
41
Introduction to diffusion tensor imaging (DTI)
MVA 01-2018
42
Introduction to diffusion tensor imaging (DTI)
Classical MRI: white mater, grey matter, CSF
Diffusion MRI: MR technique born in the mid-80ies (Basser, LeBihan).
in vivo imaging of the white mater architecture
set of nervous fibers (axons): information highways of the brain!
MVA 01-2018
10
43
Anatomy of a diffusion MRI
T2 image
+
6 diffusion weighted images
MVA 01-2018
44
Stejskal & Tanner equation
MVA 01-2018
45
Visualization using ellipsoids
PL
S
RA
I
MVA 01-2018
11
46
Diffusion Tensor ImagingCovariance of the Brownian motion of water
-> Architecture of axonal fibers
Symmetric positive definite matrices
Cone: Convex operations are stable mean, interpolation
Null eigenvalues are reachable in a finite time : not physical!
More complex operations are not PDEs, gradient descent…
SPD-valued image processing
Clinical images: Very noisy data
Robust estimation
Filtering, regularization
Interpolation / extrapolation Diffusion Tensor Field(slice of a 3D volume)
Intrinsic computing on Manifold-valued images?
MVA 01-2018
48
Statistical Computing on Manifolds for Computational Anatomy
Simple Statistics on Riemannian Manifolds
Manifold-Valued Image Processing
Introduction to diffusion tensor imaging
Interpolation, filtering, extrapolation of tensors
Other metrics and applications
Metric and Affine Geometric Settings for Lie Groups
Analysis of Longitudinal Deformations
MVA 01-2018
50
GL(n)-Invariant Metrics on SPD matrices
Action of the Linear Group GLn
Invariant metric
Isotropy group at the identity: Rotations All rotationally invariant scalar products:
Geodesics at Id
Exponential map
Log map
Distance
-1/n)( )Tr().Tr( Tr| 212121 WWWWWW Tdef
Id
Id
def
AA
tt WWAAWAAWWWt 2
2/11
2/12121 ,||
TAAA ..
2/12/12/12/1 )..exp()( Exp
22/12/12 )..log(|),(Id
dist
2/12/12/12/1 )..log()( Log
)exp()(, tWtWId
MVA 01-2018
12
51
GL(n)-Invariant Metrics on Tensors
Space of Gaussian distributions (Information geometry) (=0) Fisher information metric [Burbea & Rao J. Multivar Anal 12 1982,
Skovgaard Scand J. Stat 11 1984, Calvo & Oller Stat & Dec. 9 1991] Tensor segmentation [Lenglet RR04 & JMIV 25(3) 2006]
Affine-invariant metrics for DTI processing (=0) [Pennec, Fillard, Ayache, IJCV 66(1), Jan 2006 / INRIA RR-5255, 2004] PGA on tensors [Fletcher & Joshi CVMIA04, SigPro 87(2) 2007]
Geometric means (=0) Covariance matrices in computer vision [Forstner TechReport 1999] Math. properties [Moakher SIAM J. Matrix Anal App 26(3) 2004] Geodesic Anisotropy [Batchelor MRM 53 2005]
Homogeneous Embedding (=-1/(n+1) ) [Lovric & Min-Oo, J. Multivar Anal 74(1), 2000]
-1/n)( Tr .Tr21112
WWWW
MVA 01-2018
52
Bi-linear, tri-linear or spline interpolation?
Tensor interpolationGeodesic walking in 1D )(exp)( 211
tt
2),( )(min)( ii distxwxRephrase as weighted mean
RiemannianEuclidean
RiemannianEuclidean
MVA 01-2018
53
Gaussian filtering: Gaussian weighted mean
n
iii distxxGx
1
2),( )(min)(
Raw Coefficients =2 Riemann =2
MVA 01-2018
13
54
PDE for filtering and diffusion
Harmonic regularization
Gradient = manifold Laplacian
Trivial intrinsic numerical schemes with exponential maps!
Geodesic marching
Anisotropic diffusion Perona-Malik 90 / Gerig 92
Robust functions
2
2)1(
,, ,,
2 )()( )( uO
u
uxxx
ui
zyxi zyxiii
dxxC
x
2
)()()(
))((exp)( )(1 xCx xt t
u uxuw xxwx )( )()(
)(
dxx
x
2
)()()(Reg
[ Pennec, Fillard, Arsigny, IJCV 66(1), 2005, ISBI 2006]
MVA 01-2018
55
Anisotropic filtering
Initial
Noisy
Recovered
MVA 01-2018
56
Anisotropic filtering
Raw Riemann Gaussian Riemann anisotropic
)/exp()( with )( )()(ˆ 22
)(ttwxxwx
uuxu
2)1(2 /)()( )( uuxxx uuuu
MVA 01-2018
14
57
Extrapolation by Diffusion
Diffusion =0.01Original tensors Diffusion =
2
)(1
2 )(2
)),(()(2
1)(
x
n
iii xdxxdistxxGC
))(()()())((1
xxxxGxCn
iii
MVA 01-2018
58
Statistical Computing on Manifolds for Computational Anatomy
Simple Statistics on Riemannian Manifolds
Manifold-Valued Image Processing
Introduction to diffusion tensor imaging
Interpolation, filtering, extrapolation of tensors
Other metrics and applications
Metric and Affine Geometric Settings for Lie Groups
Analysis of Longitudinal Deformations
MVA 01-2018
Log Euclidean Metric on Tensors
Exp/Log: global diffeomorphism Tensors/sym. matrices
Vector space structure carried from the tangent space to
the manifold
Log. product
Log scalar product
Bi-invariant metric
Properties
Invariance by the action of similarity transformations only
Very simple algorithmic framework
2121 loglogexp
logexp
2
212
21 loglog, dist
[Arsigny, MICCAI 2005 & MRM 56(2), 2006]
59MVA 01-2018
15
60
Riemannian Frameworks on tensors
Affine-invariant Metric (Curved space – Hadamard)
Dot product
Geodesics
Distance
Log-Euclidean similarity invariant metric (vector space)
Transport Euclidean structure through matrix exponential
Dot product
Geodesics
Distance
[ Arsigny, Pennec, Fillard, Ayache, SIAM’06, MRM’06 ]
[ Pennec, Fillard, Ayache, IJCV 66(1), 2006, Lenglet JMIV’06, etc]
2/12/12/12/1 )...exp().( ttExp
22/12/12
2
)..log(|),(distL
2
212
21 loglog,dist
))log(.)exp(log().( ttExp
IdAA
TT WVAWAAVAWVT
2/12/12/12/1 |||
IdWVWV )log(|)log(|
MVA 01-2018
Log Euclidean vs Affine invariant
Means are geometric (vs arithmetic for Euclidean)
Log Euclidean slightly more anisotropic
Speedup ratio: 7 (aniso. filtering) to >50 (interp.)
Euclidean Affine invariantLog-Euclidean
61MVA 01-2018
Log Euclidean vs Affine invariant
Real DTI images: anisotropic filtering
Difference is not significant
Speedup of a factor 7 for Log-Euclidean
Original Euclidean Log-Euclidean Diff. LE/affine (x100)
62MVA 01-2018
16
63
Some other metrics on tensors
Square root metrics Cholesky [Wang Vemuri et al, IPMI’03, TMI 23(8) 2004.]
Size and shape space [Dryden, Koloydenko & Zhou, 2008]
Power Euclidean [Dryden & Pennec, arXiv 2011]
Non Riemannian distances J-Divergence [Wang & Vemuri, TMI 24(10), 2005]
Geodesic Loxodromes [Kindlmann et al. MICCAI 2007]
4th order tensors [Gosh, Descoteau & Deriche MICCAI’08]
MVA 01-2018
64
Euclidean metric
Choleski metric
Log-Euclidean metric
Geodesic shooting in tensors spaces
Undefined (non positive)
MVA 01-2018
65
DTI Estimation from DWI
Baseline + at least 6 DWI
iTii DgbgSS exp0
+
Stejskal & Tanner diffusion equation
Diffusion Tensor Field
Noise is Gaussian in complex DW signal Rician = amplitude of complex
Gaussian LSQ on Rician noise = bias for low
SNRs [Sijbers, TMI 1998]
Estimation / Regularization on complex DWI: Anisotropic diffusion on Choleski factors
[Wang & Vemuri, TMI’04]
Estimation with a Rician noise Smoothing DWI before estimation
[Basu & Fletcher, MICCAI 2006] ML (MMSE) [Aja-Fernández et al, TMI 2008] MAP with log-Euclidean prior
[Fillard et al., ISBI 2006, TMI 2007]
17
66
Rician MAP estimation with Riemannian spatial prior
ML Rician MAP RicianStandard
Estimated tensors
FA
[ Fillard, Arsigny, Pennec, Ayache ISBI’06, TMI 26(11) 2007 ]
dxxdxSS
ISSS
MAPx
N
i
iiiii .)( .ˆ)(
2
)(ˆexp
ˆlog
2
)(1
202
22
2
MVA 01-2018
67
Clinical DTI of the spinal cord: fiber tracking
MAP RicianStandard
[ Fillard, Arsigny, Pennec, Ayache ISBI’06, TMI 26(11) 2007 ]
MVA 01-2018
T1 + Activation map + fibers
Corpus callosum + cingulum
Corticospinal tract and thalamo cortical connections
Courtesy of P. Fillardusing MedINRIA
68MVA 01-2018
18
69MVA 01-2018
Database 7 canine hearts from JHU
Anatomical MRI and DTI
Freely available at http://www-sop.inria.fr/asclepios/data/heart
•Average cardiac structure
•Variability of fibers, sheets
A Statistical Atlas of the Cardiac Fiber Structure[ J.M. Peyrat, et al., MICCAI’06, TMI 26(11), 2007]
70MVA 01-2018
Database 7 canine hearts from JHU
Anatomical MRI and DTI
Method Normalization based on aMRIs
Log-Euclidean statistics of Tensors
A Statistical Atlas of the Cardiac Fiber Structure[ J.M. Peyrat, et al., MICCAI’06, TMI 26(11), 2007]
71
Norm covariance
Eigenvalues covariance (1st, 2nd, 3rd)
Eigenvectors orientation covariance (around 1st, 2nd, 3rd)
MVA 01-2018
19
Diffusion model of the human heart10 human ex vivo hearts (CREATIS-LRMN, Lyon, France)
Classified as healthy (controlling weight, septal thickness, pathology examination)
Acquired on 1.5T MR Avento Siemens bipolar echo planar imaging, 4 repetitions, 12 gradients
Volume size: 128×128×52, 2 mm resolution
72
[ H. Lombaert Statistical Analysis of the Human Cardiac Fiber Architecture from DT-
MRI, ISMRM 2011, FIMH 2011]
Fiber tractography in the left ventricle
Helix angle highly correlated to the transmural distance
MVA 01-2018
A Statistical Atlas of the Cardiac Fiber Structure
73
[ R. Mollero, M.M Rohé, et al, to appear in FIMH 2015]
10 human ex vivo hearts (CREATIS-LRMN, France)
Classified as healthy (controlling weight, septal thickness, pathology examination)
Acquired on 1.5T MR Avento Siemens bipolar echo planar imaging, 4 repetitions, 12
gradients
Volume size: 128×128×52, 2 mm resolution
MVA 01-2018