RADMorph: A region-aware deformable registration network for organ-specific thoracic CT alignment

Deformable image registration plays an important role in radiation therapy, supporting applications such as contour propagation, treatment planning, and motion analysis. However, accurate registration of lung CT images remains challenging in the presence of large respiratory-induced deformations. In this study, we propose RADMorph, a deep learning framework that explicitly models region-dependent motion to improve registration robustness and anatomical consistency under large deformations. RADMorph comprises two key components. The Adaptive Region Feature Encoder (ARFE) learns region-specific motion representations through feature clustering and region-conditioned filtering, while the Region-Centered Displacement Modulator (RCDM) estimates deformation fields by integrating local image similarity with inter-region distance-based modulation. We evaluated RADMorph on multi-center datasets acquired under free-breathing and deep-inspiration breath-hold conditions, as well as on four-dimensional CT datasets. Across these datasets, RADMorph showed competitive overall performance compared with convolutional and transformer-based registration methods, with improvements varying across geometric and intensity-based metrics. The proposed method improved contour propagation and boundary localization for clinically relevant structures in most evaluated settings while reducing global deformation folding. Quantitative evaluations further demonstrated accurate voxel-level correspondence across the evaluated datasets, supporting the potential use of RADMorph in applications that rely on spatial correspondence, such as dose mapping and adaptive radiotherapy. RADMorph integrates region-aware representation learning with distance-modulated deformation estimation to provide a region-informed approach for modeling complex respiratory motion. These characteristics may facilitate more reliable contour transfer and quantitative analysis of dynamic thoracic imaging.