Explainable YOLO models for early skin cancer detection using hair artifact removal and VGG16 guided annotation

Skin cancer is among the most common and potentially lethal diseases globally, emphasizing the need for early and accurate detection. Deep learning, particularly object detection models, has shown significant promise in dermatological image analysis. This study introduces a YOLO, a deep learning framework for automatic identification and classification of skin lesions. The proposed methodology incorporates preprocessing techniques such as image resizing, normalization, and noise removal using hair strand removal techniques to enhance lesion visibility. Data augmentation is utilized to expand the dataset and enhance model generalization. Feature extraction using VGG16 is utilized for accurate annotation, followed by bounding box generation for object detection. We evaluate multiple versions of YOLO model, including YOLOv5, YOLOv6, YOLOv8-N, YOLOv8-S, YOLOv9-S, and YOLOv10-S, by training and validating them on preprocessed skin lesion dataset. Among these, YOLOv8-S achieves the best performance across key metrics: precision (78%), recall (88.8%), F1-score (83%), and mAP-50 (90.1%), making it the most effective model for skin cancer classification. To enhance model explainability, Grad-CAM visualization is utilized to highlight the image areas affecting the model’s classification decisiveness. The results demonstrate that combining advanced preprocessing, deep learning-based classification and explainability techniques significantly improves performance, offering a robust tool to help dermatologists in the premature detection and classification of skin cancer.