A Practical Deep Learning Framework for Multi-Class Skin Cancer Detection Using U-Net Segmentation and ResNet50 Transfer Learning

Existing artificial intelligence (AI)-based diagnostic systems often lack interpretability, clinical decision support, and accessibility for practical deployment, despite the importance of early skin cancer detection in improving patient outcomes. This paper proposes an integrated deep learning framework for multi-class skin cancer detection that combines lesion segmentation using U-Net with transfer-learning classification of seven skin-lesion classes using ResNet50. The framework links classification results with segmented lesion regions to improve interpretability and uses Gradient-weighted Class Activation Mapping (Grad-CAM) to provide visual explanations of model predictions, thereby improving transparency and clinician confidence in the recommendations. It is implemented as a lightweight web-based application for remote, resource-efficient skin-lesion screening without costly local hardware. Experimental results on the HAM10000 dataset indicate that the framework outperforms baseline deep learning models in segmentation and classification while providing interpretable predictions and computationally efficient computer-aided dermoscopic screening. The proposed system offers an explainable decision-support approach for preliminary skin cancer assessment and may support earlier diagnosis in resource-limited clinical settings.