SGP-SAM: Self-Gated Prompting for Transferring 3D Segment Anything Models to Lesion Segmentation

Researchers Zixuan Tang and Shen Zhao have developed SGP-SAM (Self-Gated Prompting for Segment Anything Models), a novel framework that significantly improves the transfer of 3D SAM-style models to lesion segmentation in medical imaging. The core innovation is the Self-Gated Prompting Module (SGPM), which uses a lightweight multi-channel gating unit to dynamically decide whether a given feature map needs additional multi-scale spatial fusion. Only when needed does it activate a Multi-Scale Feature Fusion Block, enriching spatial context without unnecessary computational overhead—a critical advantage for handling small, irregular lesions that standard SAM models struggle to detect.

To further address the extreme foreground-background imbalance common in 3D medical scans, the team designed a Zoom Loss that combines Dice loss with a voxel-balanced focal loss, up-weighting supervision on lesion regions. In experiments on the Medical Segmentation Decathlon (MSD) Liver Tumor and Brain Tumor (enhancing tumor) datasets, SGP-SAM delivered consistent gains over strong transfer baselines based on SAM-Med3D. On MSD Liver Tumor, it improved mean Dice similarity coefficient (mDice) by 7.3% compared to standard fine-tuning. This work, published on arXiv (2604.22825), offers a practical, computationally efficient path to adapting large segmentation foundation models for high-stakes medical applications.