Brain image registration optimization method via SAM-Med3D multi-scale feature migration

Aiming at the problems of insufficient anatomical structure constraints and limited feature expression ability in medical image registration, this paper proposes a registration optimization method based on SAM-Med3D and dynamic large kernel convolution. A fixed SAM-Med3D encoder was used to extract multi-scale anatomical prior (32×32×32 to 8×8×8 resolution), and a dynamic large kernel Convolution module (DLK) was used to capture long-range spatial dependencies. A cross-attention mechanism was designed to achieve hierarchical fusion of anatomical features and local details. Innovative introduction of lightweight channel attention adapter, complete feature space mapping with few parameters, reduce the computational overhead. Experiments on OASIS and LPBA40 data sets show that the DICE coefficient of this method is improved to 0.763±0.023 (baseline 0.725±0.125) in registration, and the normalized Jacobian determinate (NJD) is stable below 0.06, meeting the clinical safety threshold. This study verified the feasibility of feature transfer of visual pre-training model, and opened up a new paradigm for multi-modal medical analysis, which is expected to provide a reliable registration tool for brain disease diagnosis and surgical navigation, and has important clinical application potential.