Enhancing Skin Cancer Diagnosis with XAI: Integrating Grad-CAM, SHAP, and LIME for Model Interpretability

The promise of AI in medical diagnostics seems to be much brighter. So far, the adoption is crippled by the opacity of model predictions, especially in sensitive domains such as skin cancer detection. In this work, we utilize the HAM10000 dataset to develop a more robust AI framework that emphasize explainability alongside predictive accuracy. Although CNNs have dominated skin cancer classification tasks for a long time, we explore whether ViTs are truly superior in making image-based predictions. ViTs provide much better contextual understanding by means of self-attention mechanisms. Closing the interpretability gap, we apply Grad-CAM to the ViTs, SHAP to XGBoost, and LIME on EfficientNet-B3 while providing a plethora of explainable AI tools. The methods produce visual and quantitative explanations of the decision-making process of the models, so clinicians can gain insight into how features are relevant in computing predictions. Our approach therefore combines state-of-the-art modelling methods with techniques in XAI toward equipping clinicians with a transparent and reliable decision-support system, which will help increase trust and improve diagnostic workflows.