AI-Based Skin Disease Diagnosis with Non-Dermoscopic Images: Tackling Bias and Data Limitations

AI-based skin disease diagnosis holds significant potential for improving healthcare equity but remains challenged by fairness concerns, particularly in underrepresented populations. This study addresses these issues using a real-world dataset from an Italian hospital, which suffers from limited diversity in skin tones and disease classes, as well as non-dermoscopic, low-quality images captured under inconsistent conditions. These factors contribute to classification bias and hinder existing fairness mitigation strategies. We propose a novel two-stage pipeline that combines (1) targeted data augmentation using DreamBooth fine-tuned Stable Diffusion to generate synthetic images for darker skin tones, and (2) disease classification using a Swin Transformer model. Our results show improved fairness metrics and balanced performance across skin tone groups, demonstrating the effectiveness of synthetic data in reducing dermatological AI bias.