GLiM-SAM: Global–local Mamba-enhanced SAM for 3D epilepsy lesion segmentation

Accurate localization of epilepsy foci is critical for surgical decision-making in the preoperative evaluation of drug-resistant epilepsy. However, epilepsy lesions are typically small, have blurred boundaries, and exhibit low contrast with surrounding tissues on MRI scans, posing significant challenges for segmentation. Although the Segment Anything Model has shown potential in medical image segmentation, it still struggles with the fine anatomical structures and complex spatial dependencies inherent in epilepsy images. To address these challenges, we propose the GLiM-SAM model, which integrates a Mamba adapter to enhance the global–local interaction mechanism of SAM. This model combines the pre-trained SAM-Med3D with a lightweight Mamba adapter and employs a two-stage hierarchical learning strategy. Experimental results show that GLiM-SAM requires fewer than 20M trainable parameters and outperforms existing methods on two clinical datasets. On the PUFH dataset, GLiM-SAM achieves a Dice coefficient of 75.59% and an HD95 of 6.26, exceeding the best traditional model, MedNeXt, by 24.29% in Dice while reducing HD95 by 21.52. It also outperforms state-of-the-art methods, with improvements of 2.1% in Dice and reduction of 0.56 in HD95. On the UCLH dataset, GLiM-SAM attains a Dice score of 78.84%, consistently outperforming all baselines. These results indicate that GLiM-SAM provides more accurate lesion segmentation and improved boundary delineation, supporting reliable epilepsy diagnosis and treatment.