PEFT-MedSAM: Efficient Fine-Tuning of Medical Foundation Models for Explainable Skin Lesion Segmentation
Researchers have developed PEFT-MedSAM, a parameter-efficient fine-tuning method for the Medical Segment Anything Model (MedSAM) to improve the segmentation of skin lesions in dermoscopic images. This technique freezes the pre-trained image and prompt encoders while training only the lightweight mask decoder. Experiments on the ISIC 2018 dataset demonstrated PEFT-MedSAM achieved a dice coefficient of 0.9411, outperforming both a fully trained U-Net baseline and zero-shot MedSAM inference. Further validation on the PH2 dataset yielded a dice coefficient of 0.9467, with Grad-CAM explainability confirming the model's accurate classification of skin lesion regions. AI
IMPACT This research could lead to earlier and more accurate detection of melanomas through improved automated skin lesion segmentation.