Skin Cancer Detection via Deep Learning with Explainability and Fairness Analysis

Abstract

Abstract— Skin cancer is one of the most common cancers worldwide, and early detection is critical for improving survival rates. Deep learning has shown promising performance in automated skin lesion classification; however, two barriers limit its clinical adoption: the black-box nature of predictions and performance disparities across different skin tones. In this study, we propose a deep learning–based framework that integrates lesion classification, interpretability, and fairness analysis. We fine-tune a ResNet-50 backbone for the binary classification of dermatoscopic images, apply Grad-CAM to generate visual explanations of model predictions, and evaluate fairness across skin tone subgroups using performance and fairness-specific metrics. Lightweight bias mitigation strategies, including oversampling and weighted loss, are explored to reduce disparities. This integrated approach addresses not only accuracy but also transparency and Fairness across subgroups.