Fairness-aware data subset selection for skin cancer detection using AI

Skin cancer remains a serious public health concern in the United States, with over 200,000 new cases and approximately 13,000 deaths reported in 2024. Melanoma leads to disproportionatelyhigh mortality among people with darker skin tones, often due to delayed detection and limited access to accurate diagnostic tools.Artificial Intelligence (AI) and Deep Learning (DL) models are increasingly used to support early detection of skin cancer. However, many of these systems underperform darker skin tones,especially those classified as FST V–VI under the Fitzpatrick Skin Tone classification. This performance gap raises concerns that AI and DL technologies may inadvertently reinforce existing health inequities. While recent fairness research has emphasized algorithmic adjustments, such efforts frequently overlook the foundational role of data quality and representation.This dissertation introduces two complementary frameworks that address fairness through intelligent data selection. The first, FAIR-SCAN (Fairness and Accuracy through Intelligent Data Ranking), estimates how much each image contributes to both model accuracy and fairness, then selects subsets of data that balance these goals. The second, FAIR-SCOPE (Fairness-Aware Subset Selection using Clustering and Observed Phenotype Equity), extends this concept by grouping images according to skin phenotype and uses a coreset methodology to ensure equitable representation across clusters during training. Using publicly available dermatology datasets, both approaches improve accuracy and reduce bias in model performance across skin tones while reducing training data requirements by half or more.By placing data representation at the center of fairness efforts, this research offers a new path for mitigating racial and phenotypic disparities in AI-driven healthcare. Together, FAIR-SCAN and FAIR-SCOPE advance the development of diagnostic systems that are not only accurate and efficient but also equitable and socially responsible, supporting the broader goal of achievingfairness and trust in medical AI.