1 / 30

Statistics on Diffeomorphisms in a Log-Euclidean Framework

Mathematical Foundations of Computational Anatomy ( MFCA-2006 ), Copenhagen, October 1st, 2006. Satellite workshop of MICCAI’06. Statistics on Diffeomorphisms in a Log-Euclidean Framework. Vincent Arsigny ¹ ,Olivier Commowick ¹ ² , Xavier Pennec ¹ , Nicholas Ayache ¹ .

dorit
Download Presentation

Statistics on Diffeomorphisms in a Log-Euclidean Framework

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Mathematical Foundations of Computational Anatomy (MFCA-2006), Copenhagen, October 1st, 2006. Satellite workshop of MICCAI’06. Statistics on Diffeomorphisms in a Log-Euclidean Framework Vincent Arsigny¹ ,Olivier Commowick¹ ², Xavier Pennec¹, Nicholas Ayache¹. ¹ Research Team ASCLEPIOS, INRIA Sophia, France.² DOSISoft SA, Cachan, France.

  2. Why Statistics on Diffeomorphisms? • Linked to non-rigid registration: • Comparison of algorithms • Introducing constraints[Pennec, MFCA, MICCAI’05], [Commowick, MICCAI’05] • Registration-based morphometry[Lepore, MFCA & MICCAI’06] Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  3. Statistics on Diffeomorphisms • Euclidean statistics:[Charpiat et al., ICCV’05], [Rueckert et al., TMI, 03] • Simple: vectorial on displacement fields (or B-Spline parameters) • Not consistent with invertibility • Space of “initial momentum”[Vaillant et al., NeuroIm, 04] • Remarkable framework of Trouvé et al., widely used • Hard to use for general diffeos (vs. landmarks) Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  4. Log-Euclidean Framework • Idea: • Simple processing • Consistency with group structure(e.g., inversion-invariance) • Previous work: finite-dimensional case Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  5. Outline • Presentation • Finite-Dimensional Case • Case of Diffeomorphisms • Numerical Algorithms • Experimental Results • Conclusions Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  6. Tensor Processing • In recent years : • Need to process symmetric positive-definite matrices (“tensors”) in various contexts • Deformation tensors (e.g., in registration results) • Diffusion tensors (i.e., DT-MRI) • Metric tensors, etc. • Need: • Consistency with manifold and algebraic structures. • Simplicity desirable. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  7. Log-Euclidean Framework • References: [Arsigny, MRM, 06][Arsigny, SIAM, 06], patent pending. • Idea: one-to-one correspondence with symmetric matrices, via matrix logarithm. • Simply process tensors via their (vectorial) logarithm! Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  8. ³ ´ 2 2 ( ) ( ( ) ( ) ) d l l S S T S S i t ¡ s r a c e o g o g = 1 2 1 2 ; : Theoretical Properties • Inversion-invariance • Similarity-invariance, for example with (Frobenius): • No Euclidean defect, exactly as in the affine-invariant case. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  9. Ã ! N X ( ) ( ) l E S S w e x p w o g = L E i i i i ; : i 1 = Log-Euclidean Mean • Log-Euclidean Fréchet meangeneralizes the geometric mean: • Affine-invariant case: implicit equation and iterative solving (20 times slower). Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  10. MedINRIA Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  11. And Linear Transformations? • References: [Arsigny, WBIR’06], [Commowick, ISBI’06], [Alexa, SIGGRAPH’02]. • Idea: linearize geometrical transformationsclose enough to identity via matrix logarithm. • Simply process transformations via their (vectorial) logarithms! • E.g., fuse local linear transformations into global invertible deformations. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  12. Examples: Polyaffine Transformations Fusing two translations Fusing two rotations

  13. Theoretical Properties • Restriction: to data whose logarithm is well-defined (e.g., no negative determinant allowed). • Inversion-invariance • Log-Euclidean mean is: • Affine-invariant (i.e., by affine change of coordinate system) • A geometric mean (determinant is geometric mean of data) Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  14. General Finite-Dimensional Case • References: [Arsigny, PhD, 06] • Data: logarithm must be well-defined(ok near the identity). • Properties: • Inversion-invariance • Log-Euclidean mean: invariant w.r.t. action of adjoint representation. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  15. Outline • Presentation • Finite-Dimensional Case • Case of Diffeomorphisms • Numerical Algorithms • Experimental Results • Conclusions Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  16. ( ) V _ x x = . Generalization to Diffeomorphisms • Diffeomorphisms belong to an infinite-dimensional Lie groups. • Logarithm of a diffeomorphism is a smooth vector field. • Exponential of a smooth vector field V(x): integration during 1 unit of time of the ODE: Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  17. Correspondence between Vector fields and Diffeomorphisms exp log Vector field Diffeomorphism Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  18. ( ) ( ) 8 @ d V V D I i 0 0 ' : e x p e e x p = = V ; . . . Technical Difficulty • Is the exponential locally diffeomorphic? • We have: • Infinite-dimensional case: not sufficient. • For general diffeomorphisms (very large space): not true. • For Banach-Lie groups: true. • Group of A. Trouvé: very close to a Banach-Lie group. Thus excellent candidate. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  19. Outline • Presentation • Finite-Dimensional Case • Case of Diffeomorphisms • Numerical Algorithms • Experimental Results • Conclusions Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  20. ( ( ) ) V t e x p R t 2 : ( ) ( = ) ( = ) V V V 2 2 e x p e x p e x p = : . General Principle • Idea: take advantage of algebraic properties of exp and log. • In particular:is a one-parameter subgroup. • E.g., → Direct generalization of numerical matrix algorithms. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  21. N N N N ¡ ¡ ( ) M 2 2 2 2 e x p : Scaling and Squaring Method Matrix case Choose normalization Compute Square recursively N times Vector field case Choose normalization Compute flow at time Compose recursively N times Deformations double at each recursive step. Vector field Diffeomorphism Numerical precision so far: 0.3% on average. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  22. Scaling and Squaring Method Fusion of two rotations (N=6).

  23. N N N 2 2 2 N N 2 2 Inverse Scaling and Squaring Diffeomorphism case Choose normalization Compute recursively N square roots (gradient descent). Multiply by final displacements Matrix case Choose normalization Compute recursively N square roots. Multiply by final matrix. Numerical precision so far: 3% on average. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006 Inverse Scaling and Squaring Method

  24. Outline • Presentation • Finite-Dimensional Case • Case of Diffeomorphisms • Numerical Algorithms • Experimental Results • Conclusions Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  25. Experimental Setup • Data set: 9 T1 MR images (3D) • Atlas-to-subject registration with 256x256x60 artificial T1 MR image (the ‘atlas’, from the Brainweb) • Robust affine registration followed by non-rigid registration of [Stefanescu, MedIA,04] guaranteeing invertibility of deformations. • → Computation of Euclidean and Log-Euclidean mean deformations. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  26. Experimental Results • Idea: L-E Mean deformation Jacobians Amplitude of def. Euclidean vs. Log-Euclidean • Largest deformations: ventricles, bigger in subjects than atlas. • Euclidean and Log-Euclidean quite close, except in regions of large deformations (then up to 30% of difference).

  27. Outline • Presentation • Finite-Dimensional Case • Case of Diffeomorphisms • Numerical Algorithms • Experimental Results • Conclusions Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  28. Conclusions • Log-Euclidean framework for diffeomorphisms: simple in spite of infinite dimensions. • Nice properties: e.g.,inversion-invariance (compatible with “inverse-consistency”) • Vectorial statistics thus directly generalized to diffeomorphisms. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  29. Perspectives • Addressing technical/mathematical issues • Better numerical algorithms for exp and log, more adapted to geometrical deformations (vs. matrices) • Challenge: finding efficient way of injecting global statistics on deformations in registration algorithms. Vincent Arsigny et al., Log-Euclidean Statistics, MFCA-2006

  30. Any questions? Thank you for your attention!

More Related