DADWMorph: A Global-Local Collaborative Network for Deformable Medical Image Registration

Deformable image registration enables spatial alignment across sequential scans for longitudinal disease monitoring, multi-modal fusion in treatment planning, and atlas-based segmentation. Accurate registration supports automated workflows in clinical imaging informatics by ensuring spatial correspondence across heterogeneous imaging acquisitions. Current Transformer-based registration methods struggle to establish reliable long-range anatomical correspondences during large deformations. Moreover, they are susceptible to redundant feature interference that compromises fine-grained local detail preservation. To address these challenges, we developed DADWMorph, a deep learning framework incorporating two dedicated components. The dilation-aggregator block (DAB) employs aggregator attention with dilated convolution to capture long-range spatial dependencies and establish robust anatomical correspondences, while the density-weighted block (DWB) integrates spatial density function-weighted convolution to enhance structure-specific feature extraction while suppressing redundant edge responses. On the complex brain IXI dataset, DADWMorph achieved improved registration accuracy (DSC = 76.05%, HD95 = 3 mm) with 58% reduction in folding artifacts compared to baseline methods on average. To validate generalizability, we further evaluated the framework on cross-modality abdomen MR-CT registration involving large anatomical deformations and heterogeneous image characteristics, where it attained the highest accuracy (DSC = 69.77%, HD95 = 9.8 mm) with low computational cost. DADWMorph demonstrates robust performance across diverse anatomical structures and imaging modalities. This work demonstrates potential clinical value for applications in atlas-based brain parcellation, radiation therapy dose mapping, and multi-modal image fusion in medical imaging informatics.