Gül, SevdaAydın, Bekir MuratAkgün, DevrimKara, Rabia ÖztaşÇetinel, Gökçen2025-02-132025-02-132024Gül, S., Aydın, B. M., Akgün, D., Kara, R. Ö., Çetinel, G. (2024). Skin lesion segmentation with semantic SAM: Pros and cons. Selcuk University Journal of Engineering Sciences, 23 (3), 77-84.2757-8828https://hdl.handle.net/20.500.12395/54399The Segment Anything Model (SAM), introduced in April 2023, has gained prominence for its ability to generalize across various image segmentation tasks. This study evaluates SAM's performance on skin lesion segmentation using both public (3463 images) and private (773 images) dermoscopy image datasets, the latter collected with ethical approval from *** University Training and Research Hospital. The segmentation performance was assessed using Intersection over Union (IoU) and Dice metrics, achieving Dice scores of 0.6598 (IoU: 0.5865) for the private database and 0.6513 (IoU: 0.5624) for the public database. A post-processing step was applied to refine the segmentation results, enhancing SAM's ability to delineate lesion boundaries. However, while SAM demonstrated strong generalization, its performance on low-contrast and irregularly shaped lesions indicates the need for further adaptation. This paper highlights SAM’s potential in medical image segmentation while outlining its limitations, especially in specialized tasks like skin lesion analysis.eninfo:eu-repo/semantics/openAccessSkin Lesion SegmentationSegment Anything ModelDeep LearningCilt Lezyonu SegmentasyonuHer Şeyi Segmentleme ModeliDerin ÖğrenmeArticle2337784