Unsupervised Deformable Image Registration Revisited: Enhancing Performance with Registration-Specific Designs

Hengjie Liu, Dan Ruan, Ke Sheng

Keywords: Deformable Image Registration, Unsupervised Learning, Benchmark, Ablation Study

TL;DR: We pinpoint the true contributors to good unsupervised learning-based deformable image registration.

Abstract: Deformable image registration (DIR) is ill-posed. Many registration-specific designs and regularizations, whose rationale carries across classic optimization methods to deep-learning-based (DL) frameworks, are crucial to registration performance. This paper presents a comprehensive “ablation” type study to pinpoint the key drivers for unsupervised monomodal DL-DIR. We conducted controlled experiments and benchmarked performance against state-of-the-art methods. Our findings highlight the benefits of multi-resolution pyramids, local correlation, and inverse-consistency constraints, and demonstrate that simple network architectures can achieve strong performance—even with far less training data. The code will be publicly available at: https://github.com/HengjieLiu/Unsupervised-DL-DIR-Revisited.

Submission Number: 21