Novel Multi-Scale Architecture for Medical Image Registration

Abstract

Medical image registration is an integral component of many medical image analysis pipelines. While registration has conventionally been carried out using optimization techniques, there is growing interest in the application of deep learning to medical image registration. Deep learning based image registration (DLIR) methods have shown mixed results; they are competitive with optimization-based methods for some small-displacement datasets, but struggle to match the performance of optimization-based methods in large displacement settings. This work explores what architectural features can improve network generalization by adopting tried and tested approaches from optical flow literature.