FairGen uses AI diffusion to create demographically balanced medical images, cutting bias by up to 95.9%

Medical AI models often inherit demographic biases from imbalanced training data, where certain groups are underrepresented due to healthcare access gaps or differential disease prevalence. To address this, a team of researchers led by Zhimin Li et al. developed FairGen, a fairness-aware diffusion framework that generates demographically balanced medical images while preserving pathology-relevant features. Unlike standard generative models, FairGen embeds physician-aligned preferences into the diffusion process, ensuring that synthesized images not only look realistic but also cover minority subgroups that are typically sparse in clinical datasets.

FairGen was evaluated across three domains: dermatology, chest radiography, and brain MRI. It achieved fairness improvements of 95.9% for skin images, 80.0% for chest X-rays, and 35.2% for brain MRIs—all while maintaining competitive diagnostic accuracy compared to models trained on original data. Clinician-facing reviews and external validation on independent cohorts confirmed that these gains are not limited to standard fidelity metrics. The paper, accepted at npj Digital Medicine, demonstrates a practical path toward equitable AI in medical imaging without sacrificing clinical utility.