DermDiT framework addresses the critical issue of skin color bias in medical AI systems by generating diverse, realistic dermoscopic images using vision-language models.

Key Innovations:

• Leverages multimodal text-image learning to create balanced training datasets
• Utilizes prompting techniques to guide AI toward generating inclusive diagnostic images
• Specifically targets bias mitigation for sensitive attributes like skin color
• Creates a pathway to more equitable AI diagnostic tools in dermatology

Why It Matters: Biased diagnostic performance across different skin types poses serious equity concerns in healthcare. This approach offers a scalable, technical solution to address representation gaps without requiring massive new data collection efforts.

Prompting Medical Vision-Language Models to Mitigate Diagnosis Bias by Generating Realistic Dermoscopic Images