MedSAM Guider: A lightweight guidance module to mitigate prompt sensitivity in segmentation foundation models

抄録

Medical segmentation models such as MedSAM have shown good zero-shot capabilities on high-contrast modalities such as MRI and CT, yet remain highly sensitive to prompt accuracy. Even minor inaccuracies in bounding boxes or shape annotations can significantly degrade the quality of predicted masks. This limitation poses a major challenge in real-world clinical environments, where time pressure and anatomical complexity frequently lead to imperfect prompts. To address this, we propose a lightweight “MedSAM Guider” module that adaptively refines imprecise prompts without altering MedSAM’s underlying structure. By interpreting and correcting bounding boxes or noisy inputs, the module offers a steadier “guiding hand” for large foundation models such as MedSAM. Preliminary experiments suggest that our guidance approach consistently improves segmentation accuracy for challenging clinical scenarios. Future work will extend this solution to a wider range of modalities and anatomies, further validating its clinical potential for prompt-based medical image segmentation.