melSeg: An Adaptation of Segment Anything Model for Skin Lesion Segmentation

Skin cancer, particularly melanoma, is a major global health concern. Numerous digital analysis methods are available for detecting malignant skin lesions; however, limited accuracy in lesion identification can reduce their effectiveness in providing reliable diagnoses. This study presents a deep learning approach to develop an image segmentation model, melSeg, which accurately segments skin lesion areas and enhances diagnostic accuracy. We employ the state-of-the-art Segment Anything Model (SAM), originally designed for general-purpose segmentation tasks. To adapt SAM for skin lesion segmentation, we incorporate a feature pyramid network (FPN) to capture lesion features at multiple scales during fine-tuning. The model was rigorously evaluated on the HAM10000 dataset, a publicly available resource containing 10,015 dermatoscopic images of skin lesions annotated with corresponding lesion area masks. On this dataset, our adapted SAM model achieved an Intersection over Union (IoU) score of 89.29% and a Dice score of 94.19%, marking a substantial improvement over contemporary state-of-the-art methods in skin lesion segmentation. These results underscore the potential of our model in early skin cancer detection, offering more accurate diagnostic tools to aid in combating skin cancer.