Keywords: heatmap regression, large deformation, graph convolutions, supervised registration learning
Abstract: We propose a novel concept for supervised learning of image registration for large deformations. Based on ideas from discrete graphical models in image registration, we design a network architecture that learns to predict discrete heatmaps for the relative displacement of a number of sparse keypoints between two scans. Graph convolutions are used to model a globally smooth transformation and deformable convolutions are used to learn suitable features representations and a similarity metric to estimate sparse displacements based on the volumetric scans in an end-to-end manner. Experimental validation for weakly-supervised label-driven registration demonstrates an improvement of 10% points of overlap accuracy compared to a state-of-the-art deep learning approach.
Code Of Conduct: I have read and accept the code of conduct.
Remove If Rejected: Remove submission from public view if paper is rejected.