Keywords: Supervoxel segmentation, inter-patient registration, noisy labels
Abstract: Deep learning has the potential to substantially improve inter-subject alignment for shape and atlas analysis. So far most highly accurate supervised approaches require dense manual annotations and complex multi-level architectures but may still be susceptible to label bias. We present a radically different approach for learning to estimate large deformations without expert supervision. Instead of regressing displacements, we train a 3D DeepLab network to predict automatic supervoxel segmentations. To enable consistent supervoxel labels, we use the warping field of a conventional approach and increase the accuracy by sampling multiple complementary over-segmentations. We experimentally demonstrate that 1) our deformable supervoxels are less sensitive to large initial misalignment and can combine linear and nonlinear registration and 2) using this self-supervised classification loss is more robust to noisy ground truth and leads to better convergence than direct regression as supervision.
Paper Type: methodological development
Primary Subject Area: Image Registration
Secondary Subject Area: Learning with Noisy Labels and Limited Data
Paper Status: original work, not submitted yet
Source Code Url: https://github.com/multimodallearning/slic_reg
Data Set Url: https://github.com/multimodallearning/slic_reg
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.
TL;DR: DL-based registration so far used either regression or metric/label-supervision, here we propose deformable supervoxel segmentation for improved learning.