BiasMitigateGan: SYNTHESIZING FAIR TRAINING DATA FOR DERMATOLOGY AI USING DIFFUSION MODELS

AI systems used in dermatology which have been trained using open databases, frequently show bias toward people with lighter skin tones (types I-III).When a community does not have enough trained practitioners, patients from that area may not get accurate or safe diagnoses which raises important questions about equal treatment and patient safety.This paper tells about BiasMitigateGAN, a generative approach built to generate dermatoscopic images equal in representation among several Fitzpatrick skin types.In our approach, we combine (1) conditional diffusion modeling to regulate image generation with skin tone and disease in mind and (2) distributionaware latent resampling to more clearly expose the less common disease-skin type combinations.On dermatology-oriented datasets, including Fitzpatrick17k, BiasMitigateGAN makes sure to treat groups equally using a special fairness loss.Both evaluation results show that our way of classifying melanoma helps close diagnostic gaps reaching 92% accuracy for individuals with dark flat moles (FST V-VI) and achieving a FID score 12.8, so it overperforms standard diffusion models.The research indicates that fairness-by-design generative models can support equal treatment and promote fair AI use in dermatology and other similar areas.