Learn to Fuse Input Features for Large-Deformation Registration with Differentiable Convex-Discrete Optimisation

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.