Abstract
Background: Melanoma, the deadliest form of skin cancer, demands accurate and timely diagnosis to improve patient survival rates. However, traditional diagnostic approaches rely heavily on subjective clinical interpretations, leading to inconsistencies and diagnostic errors. Methods: This study proposes a robust YOLOv8-based deep learning framework for real-time melanoma detection and segmentation. A multi-dataset training strategy integrating the ISIC 2020, HAM10000, and PH2 datasets was employed to enhance generalizability across diverse clinical conditions. Preprocessing techniques, including adaptive contrast enhancement and artifact removal, were utilized, while advanced augmentation strategies such as CutMix and Mosaic were applied to enhance lesion diversity. The YOLOv8 architecture unified lesion detection and segmentation tasks into a single inference pass, significantly enhancing computational efficiency. Results: Experimental evaluation demonstrated state-of-the-art performance, achieving a mean Average Precision (mAP@0.5) of 98.6%, a Dice Coefficient of 0.92, and an Intersection over Union (IoU) score of 0.88. These results surpass conventional segmentation models including U-Net, DeepLabV3+, Mask R-CNN, SwinUNet, and Segment Anything Model (SAM). Moreover, the proposed framework demonstrated real-time inference speeds of 12.5 ms per image, making it highly suitable for clinical deployment and mobile health applications. Conclusions: The YOLOv8-based framework effectively addresses the limitations of existing diagnostic methods by integrating detection and segmentation tasks, achieving high accuracy and computational efficiency. This study highlights the importance of multi-dataset training for robust generalization and recommends the integration of explainable AI techniques to enhance clinical trust and interpretability.