Implicit Bias: Sex Differences in Synthetic Brain MRIs from Generative AI
Keywords: Sex Differences, Generative AIs, and Synthetic Brain MRI
TL;DR: Generative AI could be prone to misrepresenting sex differences, highlighting the need for evaluation pipelines to ensure accurate and fair sex differences in synthesis.
Abstract: While research on sex differences has garnered significant attention, the evaluation of fairness and sex representation in generative AIs, particularly medical imaging generative models, remains largely unexplored. This study examines sex-specific patterns in 33 brain lesion volumes on synthetic T1-weighted brain MRIs produced by a state-of-the-art generative model. We analyze two distinct cases to explore different aspects of the data. For each analysis, we control the generation process across 125 condition cases with repeated sampling 30 times, resulting in a total of 7,500 synthetic brain datasets. Compared to the reference values, the model consistently generates larger brain and atlas lesion volumes for males (e.g., 7.235% σ) and smaller volumes for females (e.g., −6.361% σ) with significant differences (p < 0.001). Normalized volumes also reveal considerable sex differences for certain subcortical areas, with Cohen’s d values often exceeding 1 (e.g., 1.6520 in the left lateral ventricle), indicating distinct distribution patterns between sexes. However, except for a few subcortical areas, most regions do not exhibit significant differences, as generally reported in the literature. These findings suggest that the synthetic model may over-amplify sex differences compared to real brain data. In this study, we highlight the potential misrepresentation of sex differences in generative AI and emphasize the need for evaluation frameworks to ensure accurate and fair representation in synthetic medical data.
Submission Number: 59
Loading