Exploring local rotation invariance in 3D CNNs with steerable filters

Vincent Andrearczyk, Julien Fageot, Valentin Oreiller, Xavier Montet, Adrien Depeursinge
Proceedings of The 2nd International Conference on Medical Imaging with Deep Learning, PMLR 102:15-26, 2019.

Abstract

Locally Rotation Invariant (LRI) image analysis was shown to be fundamental in many applications and in particular in medical imaging where local structures of tissues occur at arbitrary rotations. LRI constituted the cornerstone of several breakthroughs in texture analysis, including Local Binary Patterns (LBP), Maximum Response 8 (MR8) and steerable filterbanks. Whereas globally rotation invariant Convolutional Neural Networks (CNN) were recently proposed, LRI was very little investigated in the context of deep learning. We use trainable 3D steerable filters in CNNs in order to obtain LRI with directional sensitivity, i.e. non-isotropic filters. Pooling across orientation channels after the first convolution layer releases the constraint on finite rotation groups as assumed in several recent works. Steerable filters are used to achieve a fine and efficient sampling of 3D rotations. We only convolve the input volume with a set of Spherical Harmonics (SHs) modulated by trainable radial supports and directly steer the responses, resulting in a drastic reduction of trainable parameters and of convolution operations, as well as avoiding approximations due to interpolation of rotated kernels. The proposed method is evaluated and compared to standard CNNs on 3D texture datasets including synthetic volumes with rotated patterns and pulmonary nodule classification in CT. The results show the importance of LRI in CNNs and the need for a fine rotation sampling.

Cite this Paper

BibTeX
@InProceedings{pmlr-v102-andrearczyk19a, title = {Exploring local rotation invariance in 3D CNNs with steerable filters}, author = {Andrearczyk, Vincent and Fageot, Julien and Oreiller, Valentin and Montet, Xavier and Depeursinge, Adrien}, booktitle = {Proceedings of The 2nd International Conference on Medical Imaging with Deep Learning}, pages = {15--26}, year = {2019}, editor = {Cardoso, M. Jorge and Feragen, Aasa and Glocker, Ben and Konukoglu, Ender and Oguz, Ipek and Unal, Gozde and Vercauteren, Tom}, volume = {102}, series = {Proceedings of Machine Learning Research}, month = {08--10 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v102/andrearczyk19a/andrearczyk19a.pdf}, url = {https://proceedings.mlr.press/v102/andrearczyk19a.html}, abstract = {Locally Rotation Invariant (LRI) image analysis was shown to be fundamental in many applications and in particular in medical imaging where local structures of tissues occur at arbitrary rotations. LRI constituted the cornerstone of several breakthroughs in texture analysis, including Local Binary Patterns (LBP), Maximum Response 8 (MR8) and steerable filterbanks. Whereas globally rotation invariant Convolutional Neural Networks (CNN) were recently proposed, LRI was very little investigated in the context of deep learning. We use trainable 3D steerable filters in CNNs in order to obtain LRI with directional sensitivity, i.e. non-isotropic filters. Pooling across orientation channels after the first convolution layer releases the constraint on finite rotation groups as assumed in several recent works. Steerable filters are used to achieve a fine and efficient sampling of 3D rotations. We only convolve the input volume with a set of Spherical Harmonics (SHs) modulated by trainable radial supports and directly steer the responses, resulting in a drastic reduction of trainable parameters and of convolution operations, as well as avoiding approximations due to interpolation of rotated kernels. The proposed method is evaluated and compared to standard CNNs on 3D texture datasets including synthetic volumes with rotated patterns and pulmonary nodule classification in CT. The results show the importance of LRI in CNNs and the need for a fine rotation sampling.} }
Endnote
%0 Conference Paper %T Exploring local rotation invariance in 3D CNNs with steerable filters %A Vincent Andrearczyk %A Julien Fageot %A Valentin Oreiller %A Xavier Montet %A Adrien Depeursinge %B Proceedings of The 2nd International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2019 %E M. Jorge Cardoso %E Aasa Feragen %E Ben Glocker %E Ender Konukoglu %E Ipek Oguz %E Gozde Unal %E Tom Vercauteren %F pmlr-v102-andrearczyk19a %I PMLR %P 15--26 %U https://proceedings.mlr.press/v102/andrearczyk19a.html %V 102 %X Locally Rotation Invariant (LRI) image analysis was shown to be fundamental in many applications and in particular in medical imaging where local structures of tissues occur at arbitrary rotations. LRI constituted the cornerstone of several breakthroughs in texture analysis, including Local Binary Patterns (LBP), Maximum Response 8 (MR8) and steerable filterbanks. Whereas globally rotation invariant Convolutional Neural Networks (CNN) were recently proposed, LRI was very little investigated in the context of deep learning. We use trainable 3D steerable filters in CNNs in order to obtain LRI with directional sensitivity, i.e. non-isotropic filters. Pooling across orientation channels after the first convolution layer releases the constraint on finite rotation groups as assumed in several recent works. Steerable filters are used to achieve a fine and efficient sampling of 3D rotations. We only convolve the input volume with a set of Spherical Harmonics (SHs) modulated by trainable radial supports and directly steer the responses, resulting in a drastic reduction of trainable parameters and of convolution operations, as well as avoiding approximations due to interpolation of rotated kernels. The proposed method is evaluated and compared to standard CNNs on 3D texture datasets including synthetic volumes with rotated patterns and pulmonary nodule classification in CT. The results show the importance of LRI in CNNs and the need for a fine rotation sampling.
APA
Andrearczyk, V., Fageot, J., Oreiller, V., Montet, X. & Depeursinge, A.. (2019). Exploring local rotation invariance in 3D CNNs with steerable filters. Proceedings of The 2nd International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 102:15-26 Available from https://proceedings.mlr.press/v102/andrearczyk19a.html.

Related Material