RDNet: Region specific iterative deformation with multi-scale attention for medical image registration

Deformable medical image registration is essential for various clinical applications, including diagnosis, treatment planning, and disease monitoring. Although significant progress has been made with pyramid architecture, they often struggle to effectively capture the complex variations in deformation fields at feature maps with different resolutions. However, conventional skip connection designs inadequately address the asymmetric roles of moving and fixed images in deformation estimation, as they treat both images symmetrically without accounting for their distinct contributions to the alignment process. To address these challenges, we present RDNet, a learning-based dual-stream pyramid-based framework incorporating two key components: the Mapping Block (MB) and the Region Specific Layer (RSL). The MB module is carefully integrated into the fixed image skip connections to improve hierarchical feature alignment between the encoder and decoder. The high-level hierarchical semantic gap is efficiently minimized by MB through spatial and channel-wise attention methods, improving feature correspondence and registration accuracy. Additionally, to address the challenges caused by complex variations in the pyramid architecture, we present the RSL module in a multi-scale framework. This incorporation improves the capture of long-range dependencies specific to a region, resulting in more precise deformation estimation and improved registration accuracy while minimizing deformation loss. We conducted comprehensive experiments on two publicly available Brain MRI datasets, OASIS and LPBA40, and one Lung CT dataset to demonstrate that our proposed framework achieves state-of-the-art registration results.