Viewpoint Equivariance: Are CapsNets Any Good for Hippocampal Segmentation?
- Online Publication Date: 21 Jun 2021
- DOI: N/A
- Keywords: Hippocampal Subfields, Segmentation, Capsules, Deep Learning
Full PDF here.
In the recent years, research has shown that hippocampal subfields (HSF, fig. 1) are crucially implied in learning and memory [5].
Thus, an accurate delineation of the HSF is of critical interest. If current methodologies allow the segmentation of the HSF, traditional models aren’t robust to some natural transformations like rotations. As an example, those transformations may occur in a substantial part of the population with incomplete hippocampal inversions which are currently described in Temporal Lobe Epilepsy. Therefore, faster and more robust models are needed to be trustfully used in clinical domains.
In order to fulfill the need of an accurate and robust automated segmentation of the HSF, we propose an extension of Capsule Networks [2] to 3D segmentation tasks. Capsules exhibit built-in equivariances to transformations like rotations, local skewness, or thickness.
We hypothesise that Capsules will be able to correctly segment the HSF at an expert-like accuracy.