Sanjeev Sharma and Akhtar Husain, MJPRU University, India

Machine learning models have shown remarkable precision in skin lesion classification, but their acceptance in clinical settings is still restricted due to insufficient interpretability and consistency with medical reasoning. This study presents a framework that combines Gradient-weighted Class Activation Mapping (Grad-CAM) with knowledge from dermatology to improve model transparency and confidence.Grad-CAM is employed to produce visual explanations, which are systematically evaluated against recognized clinical features to determine alignment.The method put forward is assessed on benchmark dermoscopic datasets, indicating that the integration of medical priors enhances both interpretability and diagnostic dependability. The accuracy of the BCC/non-BCC classification was 91%. Regarding clinically meaningful XAI results, 98.9% accuracy was attained in identifying clinician-relevant BCC patterns. The average Grad-CAM normalized value for the human-assisted seprated clinical characteristics in the Clinically-inspired Visual XAI findings is 0.57.

Keywords