Mutualreg: Mutual Learning for Unsupervised Medical Image Registration

Abstract

Recently, self-training strategies have shown outstanding performance in the unsupervised medical image registration field. These strategies use their own network to generate pseudo-displacement fields (PFs) to supervise network training. However, limited diversity and accuracy of these PFs hinder their effectiveness. To address these limitations, we propose a novel mutual learning registration paradigm (MutualReg), where knowledge is distilled mutually between teacher and student networks for alternate improvement via recursive training. This involves two fundamental challenges: 1) how to generate more diverse and accurate PFs; and 2) how to effectively integrate knowledge distillation from the teacher network and learning from the student network. For the former, we employ a different and powerful teacher network thanks to the decoupling nature of MutualReg. For the latter, we introduce a Voxel-wise Reliability Criterion (VRC) module to retain reliable voxel locations of knowledge distillation. In the abdominal CT registration task, MutualReg outperforms state-of-the-art competitors, demonstrating its effectiveness. Code is available from https://github.com/PerceptionComputingLab/MutualReg/.