Region-Adaptive Attention and Edge-Guided Alignment for Medical Image Registration

Medical image registration is essential for tasks such as surgical navigation and disease diagnosis, aiming to spatially align medical images for consistent anatomical interpretation. In brain MR images, structural complexity and intensity heterogeneity lead to significant variation in registration difficulty across regions. However, most existing methods adopt uniform processing strategies, which often result in inaccurate alignment in regions with complex anatomy or blurred boundaries. To address these challenges, we propose a novel registration method that integrates a Region-Adaptive Attention (RAA) module and an Edge-Guided Alignment (EGA) module. Specifically, the RAA module extracts multi-scale features and employs a structure-aware gating mechanism to dynamically modulate the network's focus according to regional registration difficulty. This enables the network to emphasize difficult areas such as low-contrast or highly deformed regions. The EGA module incorporates edge-aware information and leverages an attention-based fusion strategy to enhance the registration precision near anatomical boundaries. The outputs of both modules are fused and decoded to generate a dense deformation field. Our method is evaluated on three public datasets: OASIS, IXI, and LPBA40. Compared to ten state-of-the-art registration techniques, our approach demonstrates distinct advantage, achieving DSC coefficients of 0.902, 0.792, and 0.705, respectively. These results validate the effectiveness of the proposed framework in addressing regional registration difficulties and improving boundary accuracy, offering a robust solution for high-precision MR image alignment.