Keywords: Locally Rotation Invariant, Convolutional Neural Network, Segmentation
TL;DR: We proposed a novel design for locally rotation invariant CNNs which is evaluated on a cellular nuclei segmentation task.
Abstract: Locally Rotation Invariant (LRI) operators have shown great potential to robustly identify biomedical textures where discriminative patterns appear at random positions and orientations. We build LRI operators through the local projection of the image on circular harmonics followed by the computation of the bispectrum, which is LRI by design. This formulation allows to avoid the discretization of the orientations and does not require any criterion to locally align the descriptors. This operator is used in a convolutional layer resulting in LRI Convolutional Neural Networks (LRI CNN). To evaluate the relevance of this approach, we used it to segment cellular nuclei in histopathological images. We compared the proposed bispectral LRI layer against a standard convolutional layer in a U-Net architecture. While they performed equally in terms of F-score, the LRI CNN provided more robust segmentation with respect to orientation, even when rotational data augmentation was used. This robustness is essential when the relevant pattern may vary in orientation, which is often the case in medical images. Keywords: Locally Rotation Invariance, Convolutional Network, Deep Learning, Segmentation, Bispectrum.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Paper Type: methodological development
Primary Subject Area: Segmentation
Secondary Subject Area: Application: Histopathology
Confidentiality And Author Instructions: I read the call for papers and author instructions. I acknowledge that exceeding the page limit and/or altering the latex template can result in desk rejection.
Code And Data: https://github.com/voreille/2d_bispectrum_cnn https://monuseg.grand-challenge.org/Data/