AI-driven race prediction from mammographic images: anatomical insights for AI model's bias mitigation

Bias in datasets and AI models have emerged as a significant challenge in the medical AI community, particularly due to the underrepresentation of certain racial groups and the overrepresentation of others, leading to performance degradation in real clinical practice. This underscores the critical need for further and deeper analysis of racial differences to address these issues effectively. Despite its importance, research on imaging race differences is scarce. This study explores race-based anatomical differences in mammographic images among Asian and White women populations. We extracted mammographic images of 1043 Asian and 986 White individuals from the EMBED dataset, ensuring that the development dataset was matched for age and BIRADS breast density. For independent testing, we extracted mammograms from 157 Asian and 673 White women from the University of Pittsburgh Medical Center (UPMC). We found that mammograms from these groups have notable anatomical differences. Specifically, we developed a Vision Transformer (ViT) model to predict race (Asian and White) in mammographic images, achieving the test AUC of 0.82 from the development and 0.74 from external validation for original images. We also used each cohort's binary breast density (representing fibroglandular tissue distribution) and binary breast area (representing breast size and shape) masks to analyze spatial breast density distribution, breast volume, and shape differences between races. Our classifier predicted race based on binary dense masks, yielding a result of 0.70/0.71 for AUC (development/external testing datasets); for binary breast area masks, the classifier achieved 0.60/0.70 (development/external testing datasets). We found that the main anatomical difference between Asian and White races is the breast size followed by breast shape. This result can be used to address racial bias between Asian and White races to ensure equitable performance of AI models in clinical settings.