Contrastive learning enhances fairness in pathology artificial intelligence systems.

Abstract

AI-enhanced pathology evaluation systems hold significant potential to improve cancer diagnosis but frequently exhibit biases against underrepresented populations due to limited diversity in training data. Here, we present the Fairness-aware Artificial Intelligence Review for Pathology (FAIR-Path), a framework that leverages contrastive learning and weakly supervised machine learning to mitigate bias in AI-based pathology evaluation. In a pan-cancer AI fairness analysis spanning 20 cancer types, we identify significant performance disparities in 29.3% of diagnostic tasks across demographic groups defined by self-reported race, gender, and age. FAIR-Path effectively mitigates 88.5% of these disparities, with external validation showing a 91.1% reduction in performance gaps across 15 independent cohorts. We find that variations in somatic mutation prevalence among populations contribute to these performance disparities. FAIR-Path represents a promising step toward addressing fairness challenges in AI-powered pathology diagnoses and provides a robust framework for mitigating bias in medical AI applications.