Descriptor-based Structural Similarity and Neural ODEs for Multimodal Diffeomorphic Registration

Abstract

This work proposes a novel learning-based method to address multimodal diffeomorphic registration. Traditional algorithms that address dense registration use numerical optimization solvers on intensity-based similarity metrics, so they work best in the monomodal setting. This work tackles this task by modality-agnostic descriptors which encode structural self-similarity and a Neural Ordinary Differential Equation (Neural ODE) encoding the dynamics of the estimated registration, achieving state-of-the-art smoothness and registration accuracy.