Deep Learning–Based Detection of Skin Lesions Using CNNs and Grad-CAM Visualization

Abstract

This Early detection of skin cancer, particularly melanoma, plays a vital role in improving patient survival. However, dermoscopic diagnosis is often subjective and depends heavily on clinical expertise. This paper presents an explainable hybrid deep learning framework for automated skin lesion classification. The proposed method integrates EfficientNetB0 as a convolutional feature extractor with a dense layer and an RBF-kernel Support Vector Machine (SVM) for final classification, aiming to improve generalization on limited and imbalanced datasets. The model was trained and evaluated using the ISIC 2020 dermoscopic image dataset with a stratified train–validation split. To enhance transparency and clinical trust, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize discriminative regions influencing model predictions. Experimental results demonstrate high classification accuracy and robust performance on unseen images, while Grad-CAM visualizations highlight clinically relevant lesion areas. These findings indicate that the proposed hybrid CNN–SVM approach provides an effective and interpretable solution for computer-aided skin lesion analysis and has strong potential for clinical decision support.