Abstract: Non-rigid 3D medical image registration, which establishes dense anatomical correspondences between scans, is a critical step in clinical diagnosis and surgical planning. However, existing methods struggle to address three key challenges simultaneously: capturing complex anatomical deformations, balancing global alignment with local detail preservation, and maintaining computational efficiency.To overcome these limitations, we introduce MRAF-Net, a Multi-stage Recursive Alignment and Fusion network designed for high-precision registration with minimal computational overhead. MRAFNet adopts a multi-stage encoder-decoder scheme. It begins with a Hierarchical Residual Encoder (HR-Encoder), which incorporates a Hybrid Anti-Aliasing Downsampling (HAAD) to extract multi-scale features while preserving structural integrity and suppressing aliasing artifacts during downsampling. Within the decoder, the core Progressive Alignment Module (PAM) recursively propagates and refines deformation fields across scales. The PAM integrates a Context-Aware Feature Fusion Module (CAFM) to model spatial-channel correspondences between images, and a lightweight Spatial Residual Flow Head (SRFH) for parameter-efficient local deformation prediction. Experiments on LPBA40, OASIS, and IXI datasets demonstrate that MRAF-Net generates anatomically plausible deformation fields with superior topological integrity. Comprehensive evaluations against the strong baseline TransMorph reveal that MRAF-Net achieves a 6.7% higher Dice score (DSC), reduces the 95% Hausdorff Distance (HD95) by 38.1%, and lowers the Average Symmetric Surface Distance (ASSD) by 41.9% on the LPBA40 dataset, while requiring 60% fewer parameters (18.7 M vs 46.5 M). These results confirm the method's efficiency and accuracy for high-resolution clinical deployment.

Key words: Cross-scale feature fusion')">Cross-scale feature fusion, Cascaded deformation registration, Weakly supervised learning, Sub-voxel registration