DETECTING SKIN CANCER USING DEEP LEARNING MODELS :A THOROUGH COMPARATIVE EVALUATION OF THEIR EFFECTIVENESS AND PERFORMANCE

Abstract

Early detection is essential for treating skin cancer, especially melanoma, and for the increased survival rate of patients. This study offers a comparative analysis of three deep learning-based object detection approaches-YOLOv3, YOLOv5, and YOLOv8-for classifying ISIC 2020 dermoscopic microscopy-images into benign or malignant. All the models were trained and evaluated after undergoing standardized preprocessing and hyperparameter settings to ensure the fair comparison. YOLOv8 attained the best performance, with mAP@0.5 of 91.5%, precision of 96.1%, recall of 90.8% at an inference time of 21 ms per image, beating YOLOv3 (mAP@0.5: 76.3%, precision: 78.5%, recall: 72.4%, inference: 45 ms) and YOLOv5 (mAP@0.5: 84.7%, precision: 87.6%, recall: 83.6%, inference: 29 ms). Moreover, YOLOv8 was also economical with respect to resources and required less computational capacity to work with, thereby making it a suitable choice for any resource-scarce environment. A Flask-based custom web program was developed wherein users can upload dermoscopic images, choose an operating model, and obtain real-time diagnostic predictions along with confidence scores (e.g., 93.2% for malignant lesions) and Grad-CAM visualization for interpretability. While there have been advances in this front, however, other issues such as dataset imbalance which would nudge toward predicting benign lesions and the lack of model interpretability still remain .Future scopes include using more explainability tools, such as SHAP, experimenting with a multi-class classification, and ultimately developing a mobile-ready package for easy use and access. Overall, the research highlights the potential of YOLOv8 as a fast and reliable testing instrument for skin cancer detection, especially useful under low-resource or remote health circumstances, significantly moving AI-assisted testing closer to enhancing patient outcomes.