Skin Cancer Detection using Vision Transformer with GNN and Explainable AI (Grad-CAM)

Skin cancer is a serious disease that affects millions of people worldwide. Better treatment results depend on early detection. Convolutional neural networks (CNNs), which are frequently used to classify skin cancer, are one example of how deep learning has enhanced medical image analysis. CNNs are less successful at identifying global patterns in medical images because they primarily concentrate on local features. Though they process image patches independently and frequently overlook spatial relationships, Vision Transformers (ViTs) help in capturing long-range dependencies through self-attention mechanisms. In order to improve classification, ViT features are refined by integrating Graph Neural Networks (GNNs) to model spatial connections between patches. A ViT-GNN hybrid model for skin cancer detection is presented in this work, in which patch-based features are extracted by ViTs and enhanced by GNNs that capture spatial relationships. Using cross-entropy loss and the AdamW optimizer, the model is trained on a balanced dataset of benign and malignant lesions. To enhance generalization, data augmentation is used. Gradient-weighted Class Activation Mapping (Grad-CAM) identifies key areas impacting predictions to guarantee transparency. Accuracy, precision, recall, F1-score, and a confusion matrix are used to assess performance. The ViT-GNN model performs better than standalone ViTs and CNNs, according to the results, with increased accuracy and dependability. This method enhances model interpretability and classification performance by fusing explainable AI techniques, transformer-based feature extraction, and graph-based spatial learning. The proposed method has strong potential for early and accurate skin cancer detection, making it a valuable tool for clinical applications.