<p>Deep brain stimulation (DBS) involves implanting electrodes into specific brain regions to modulate aberrant neural activity patterns, alleviating patient symptoms. The globus pallidus internus (GPi) serves as a critical target for DBS treatment, where accurate localization is crucial for optimizing electrode placement. The difficulty of fusing multi-modal medical data and the inadequate adaptability of large medical vision models in specific domains increases the complexity of GPi segmentation. In this study, we propose to enhance the multi-modal fusion with Mamba in SAM-Med3D, termed MF-SAM, for automatic GPi segmentation in multi-modal imaging to reduce the labor-intensive imaging labeling workload for DBS surgical planning. We integrate Mamba with transformer to enhance the image encoder’s capability to capture long-range spatial dependencies and design a Mamba-based embedding fusion module that merges mask image embeddings generated from each point sampling with those produced by the image encoder, significantly improving segmentation accuracy. Our experiments demonstrate that the proposed method achieved Dice coefficient (DSC) of 88.9% and Intersection over Union (IoU) of 80.2% for GPi segmentation, outperforming state-of-the-art models. Therefore, the efficacy of our approach in GPi segmentation provides robust support for the diagnosis and treatment of Parkinson’s disease. Source code will be available upon acceptance.</p>