1 / 36

Fast and Simple Calculus on Tensors in the Log-Euclidean Framework

8th International Conference on M edical I mage C omputing and C omputer A ssisted I ntervention, Oct 26 to 30, 2005. Fast and Simple Calculus on Tensors in the Log-Euclidean Framework. Vincent Arsigny, Pierre Fillard, Xavier Pennec, Nicholas Ayache.

Télécharger la présentation

Fast and Simple Calculus on Tensors in the 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. 8th International Conference on Medical Image Computing and Computer Assisted Intervention, Oct 26 to 30, 2005. Fast and Simple Calculus on Tensors in the Log-Euclidean Framework Vincent Arsigny, Pierre Fillard, Xavier Pennec, Nicholas Ayache. Research Project/Team EPIDAURE/ASCLEPIOSINRIA, Sophia-Antipolis, France.

  2. What are ‘tensors’? • In general: all multilinear applications. • In this talk: symmetric positive-definite matrices. • Typically : covariance matrices. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  3. Diffusion Tensor MRI • Diffusion-weightedMR images • Diffusion Tensor: local covariance of diffusion [Basser, 94]. • Generalizationof vector processing tools (filtering, statistics, etc.) to tensors? Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  4. Outline • Presentation • Euclidean and Affine-Invariant Calculus • Log-Euclidean Framework • Experimental Results • Conclusions and Perspectives Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  5. Euclidean calculus • DTs: 3x3 symmetric matrices, thus belong to a vector space. • Simple, but: • unphysical negative eigenvalues appear • ‘swelling effect’: more diffusion than originally. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  6. Remedies in the literature • First family: • process features from tensors • propagate processing to tensors. • Example: regularization • dominant directions of diffusion [Coulon, IPMI’01] • orientations and eigenvalues separately [Tschumperlé, IJCV, 02, Chefd’hotel JMIV, 04]. • Drawback: some information left behind. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  7. Remedies in the literature • Second family: specialized procedures • Affine-invariant means [Wang, TMI, 05] • Anisotropic interpolation [Castagno-Moraga, MICCAI’04] • Etc. • Drawback: lack of general framework. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  8. A general solution: Riemannian geometry • Powerful framework for curved spaces. • Statistics[Pennec, JMIV, 98],PDEs[Pennec, IJCV, 05]. • Riemannian arithmetic mean: ‘Fréchet mean’. • Basic tool:differentiable distance between tensors. http://www.alumni.ca/~wupa4p0/ Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  9. Choice of distance? • Relevant/natural invariance properties. • In 2004: affine-invariant metrics [Fletcher, CVAMIA’04, Lenglet, JMIV, 05, Moakher, SIMAX, 05, Pennec, IJCV, 05]. • invariance w.r.t. any affine change of coordinate system. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  10. 1 1 ¡ ¡ ( ) k ( ) k d l S S S S S i 2 2 t s o g = 1 2 2 1 1 ; : : : Affine-invariant metrics • Excellent theoretical properties: • no 'swelling effect' • non-positive eigenvalues at infinity • High computational cost: many algebraic operations Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  11. Outline • Presentation • Euclidean and Affine-Invariant Calculus • Log-Euclidean Framework • Experimental Results • Conclusions and Perspectives Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  12. A novel vector space structure • Surprise: a vector space structurefor tensors! • Idea: one-to-one correspondence with symmetric matrices, via matrix logarithm. • More details: [Arsigny, INRIA RR-5584, 2005].French patent pending. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  13. ( ( ) ( ) ) l l S S S S ( ( ) ) ¸ ¸ l S S ¯ + ~ e x p o g o g = e x p o g 1 2 1 2 = 1 : A novel vector space structure • Tensors: Lie group with 'logarithmic multiplication': • Tensors: vector space with 'logarithmic scalar multiplication': Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  14. ( ) k ( ) ( ) k d l l S S S S i t ¡ s o g o g = 1 2 1 2 ; : ¯ ¯ ~ ; Log-Euclidean Distances • Log-Euclidean metrics: • Euclidean metricsfor vector space structure • Bi-invariant Riemannian metricsfor Lie group structure Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  15. 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 • Similarity-invariance, for example with (Frobenius): • No Euclidean defect, exactly as in the affine-invariant case. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  16. 1 2 3 Conversion Tensor/Vectorwith Matrix Logarithm Euclidean Processing on logarithms (filtering, statistics…) Conversion Vector/Tensorwith Matrix Exponential Log-Euclidean framework in practice • Existing Euclidean algorithms readily recycled! Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  17. Ã ! N X ( ) ( ) l E S S w e x p w o g = L E i i i i ; : i 1 = Example: computing the mean • Closed formfor Log-Euclidean Fréchet mean: • Affine-invariant case: implicit equation and iterative solving (20 times slower). Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  18. Outline • Presentation • Euclidean and Affine-Invariant Calculus • Log-Euclidean Framework • Experimental Results • Conclusions and Perspectives Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  19. Affine-invariant 11\Euclidean Log-Euclidean Interpolation • Typical example of bilinear interpolation on synthetic data: Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  20. Original slice Euclidean case Log-Euclidean case Interpolation on real DT-MRI • Reconstruction by bilinear interpolation of slice in mid-sagital plane: Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  21. [b] [a] [c] [d] Regularization of tensors Data: clinical DT image128x128x30 • [a] Raw data • [b] Euclidean reg. • [c] Log-Eucl. reg. • [d] Log-Eucl. vs.affine-inv. (x100!) Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  22. Outline • Presentation • Euclidean and Affine-Invariant Calculus • Log-Euclidean Framework • Experimental Results • Conclusions and Perspectives Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  23. Conclusions • Log-Euclidean Riemannian framework: • Riemannian excellent properties. • Euclidean speed and simplicity • Existing vector algorithms readily recycled. • More applications: • Joint estimation and smoothing for DTI:[Fillard, INRIA RR-5607, 2005]. • Statistical priors in non-linear registration[Pennec, MICCAI’05, Post. II-943], [Commowick, Post. II-927]. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  24. Perspectives • Evaluation/validation (phantoms...).Which metric for which application? • Diffusion tensors(statistics, interpolation, estimation, registration…) • Variability tensors[Fillard, IPMI’05](models of anatomical varibility) • Structure tensors[Fillard, DSSCV’05](classical image processing) • Metric tensors[Allauzet, INRIA RR-4759, 2003](anisotropic mesh adaptation for PDE solving) • Extension of Log-Euclidean framework to: • Generalized diffusion tensors [Özarslan, MRM, 2003] • Q-balls [Tuch, MRM, 2004]. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  25. Thank you for your attention! Any questions?

  26. FA Gradient Regularization of tensors • Effect of anisotropicregularization on FractionalAnisotropy (FA)and gradient: Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  27. Orginal data Data+noise Euclidean result Log-Euclidean res. Regularization of tensors • Anisotropic regularization on synthetic data: Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  28. Log-Euclidean vs. affine-invariant • Very little differences • On DT images, Log-Euclidean advantages are: • simplicity: Euclidean computations on logarithms! • faster computations: computations at least 4 times faster in all situations. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  29. ( ( ) ) ( ( ) ) E E T S T S < r a c e r a c e A I L E ( ) ( ) h 6 E E S S w e n e v e r = A I L E Log-Euclidean vs. affine-invariant • Small difference: larger anisotropy in Log-Euclidean results. • (Theoretical) reason: inequality between the 'traces' of the Log-Euclidean and affine-invariant means: Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  30. 1 1 1 1 ³ ´ ¡ ¡ ( ( ) ( ) ( ) ) ( ) l l S S l S S S S S 2 2 2 2 1 t t t ¡ + e x p o g o g e x p o g 1 2 2 1 1 1 1 : : : : : : : Geodesics • Log-Euclidean case: • Affine-invariant case: Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  31. Tensor Space Homogenous ManifoldStructure Vector SpaceStructure Algebraicstructures Metrics on Tensors Invariant metric Euclidean metric Affine-invariant metrics Log-Euclidean metrics Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  32. Log-Euclidean vs. affine-invariant • with DT images, very similar results. Identical sometimes. • Reason: associated means are two different generalizations of the geometric mean. • In both cases determinants are interpolated geometrically. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  33. 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 S S ; ! 7 Log-Euclidean metrics • Invariance properties: • Lie group bi-invariance • Similarity-invariance, for example with (Frobenius): • Invariance of the mean w.r.t. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  34. Variability along sulci on the cortex and their extrapolation. Variability tensors • [Fillard, IPMI'05] Anatomical variability: local covariance matrix of displacement w.r.t. an average anatomy. Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  35. [Alauzet, RR-4981], GAMMA project. Application to fluid mechanics. Use of Tensors • Generation of adapted meshes in numerical analysis for faster PDE solving(SMASH project): Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

  36. Interpolated tensors Interpolated tensors Interpolated volumes Defects of Euclidean Calculus • Typical 'swelling effect' in interpolation: • In DT-MRI: physically unacceptable ! Vincent Arsigny et al., Log-Euclidean Framework, MICCAI'05

More Related