Advancing Equity in AI-Based Skin Cancer Diagnosis

Abstract

This project addresses disparities in skin cancer detection models, which often underperform on patients with darker skin tones, contributing to inequities in healthcare diagnosis. Standard ResNet-based models, even when fine-tuned on publicly available dermatology datasets, show clear gaps in performance across skin-tone subgroups. To improve representation and generalization, we applied targeted preprocessing techniques and balanced sampling strategies to enhance representation of minority skin tones. Building on these approaches, we incorporated self-supervised learning using the DINO training process on unlabeled dermatology images to learn more domain-aligned feature representations. We experimented with partial fine-tuning, domain-specific pretraining, and group-aware evaluation metrics to enhance fairness and reduce performance disparities under domain shift. Overall, our work demonstrates the importance of equitable training practices, advanced representation learning, and careful dataset curation in developing reliable, generalizable, and fair AI-driven skin cancer diagnostic models.