Learn to fuse input features for large-deformation registration with differentiable convex-discrete optimisation

Hanna Siebert, Mattias P Heinrich

Keywords: large deformation registration, convex optimisation, end-to-end learning

TL;DR: A differentiable convex discrete optimisation approach that is able to align images with large deformations and is used to learn the fusion of semantic and hand-crafted image features

Abstract: Hybrid methods that combine learning-based features with conventional optimisation have become popular for medical image registration. The ConvexAdam algorithm that ranked first in the comprehensive Learn2Reg registration challenges completely decouples semantic and/or hand-crafted feature extraction from the estimation of the transformation due to the difficulty of differentiating the discrete optimisation step. In this work, we propose a simple extension that enables backpropagation through discrete optimisation and learns to fuse the semantic and hand-crafted features in a supervised setting. We demonstrate state-of-the-art performance on abdominal CT registration.