You Only Look at Patches: A Patch-wise Framework for 3D Unsupervised Medical Image Registration

Lihao Liu, Zhening Huang, Pietro Lio, Carola-Bibiane Schönlieb, Angelica I Aviles-Rivero

Keywords: Medical Image Registration, Unsupervised Learning, Patchwise Contrastive Learning，Patchwise Stitching

Abstract: Medical image registration is a fundamental task for a wide range of clinical procedures. Automatic systems have been developed for image registration, where the majority of solutions are supervised techniques. However, those techniques rely on a large and well-representative corpus of ground truth, which is a strong assumption in the medical domain. To address this challenge, we propose a novel unified unsupervised framework for image registration and segmentation. The highlight of our framework is that patch-based representation is key for performance gain. We first propose a patch-based contrastive strategy that enforces locality conditions and richer feature representation. Secondly, we propose a patch stitching strategy to eliminate artifacts. We demonstrate, through our experiments, that our technique outperforms current state-of-the-art unsupervised techniques.