Toward Interpretable Skin Cancer Diagnosis: A Hybrid EfficientNet-Graph Attention Network with Grad-CAM Explainability

Skin lesion classification remains a challenging task because benign and malignant lesions vary substantially in morphology and appearance. This study aims to develop an accurate and interpretable deep learning framework for automated binary classification of skin lesions, explicitly modeling contextual dependencies among spatially distributed visual features. We propose a hybrid architecture combining EfficientNet-B0 with a Graph Attention Network (GAT), in which a pretrained EfficientNet-B0 backbone extracts discriminative representations from dermoscopic images, while the GAT models interactions among spatial features derived from the convolutional feature map. The refined representations are globally aggregated and subsequently classified into benign and malignant categories. To further enhance interpretability, Grad-CAM is employed to visualize the regions contributing most strongly to the model's predictions. Experiments on the publicly available "Skin Cancer: Malignant vs. Benign" dataset show that the proposed model achieves 91.21% accuracy, 91.12% precision, 91.17% recall, and a 91.14% F1-score on the independent test set, surpassing the baseline EfficientNet-B0 without GAT, alternative attention mechanisms, evaluated convolutional backbones, and results reported in related literature. Grad-CAM visualizations further confirm that the model focuses predominantly on discriminative lesion regions. These findings demonstrate that integrating graph-based attention with convolutional feature extraction enhances discriminative capability and generalization while providing clinically meaningful visual explanations for skin lesion classification.