Leveraging Multi-scale Cycle‑consistency for Point‑based Deformable Lung CT Registration

Nahyuk Lee, Doohyun Park, Sungjoo Lim, Taemin Lee

Keywords: Multi-scale Cycle Consistency, Point-based Deformable Lung CT Registration

TL;DR: We propose a point-based lung CT registration framework that leverages multi-scale cycle consistency to improve the geometric coherence and invertibility of deformation prediction.

Abstract: We propose a new point-based deformable lung CT registration network that integrates multi-scale cycle consistency into bidirectional, hierarchical deformable registration. Our method jointly estimates forward and backward flows while enforcing consistency across resolutions to learn anatomically coherent and invertible mappings. Experiments on the Lung250M benchmark demonstrate improved registration and robustness over unidirectional methods, establishing our approach as an effective and efficient registration baseline.

Submission Number: 44