Reproduce “A Generative AI Approach for Reducing Skin Tone Bias in Skin Cancer Classification” on OpenResearch

About this paper

Areez Muhammed Shabu, Mohammad Samar Ansari, Asra Aslam

Researchers at the University of Sheffield and University of Chester developed a generative AI pipeline, leveraging LoRA-adapted Stable Diffusion, to synthesize dermoscopic images of darker skin tones. This approach reduced skin tone bias and improved the performance of skin lesion segmentation and classification models, particularly on real, unseen images.

Key insights

• LoRA-adapted Stable Diffusion can effectively generate high-fidelity, clinically relevant dermoscopic images for underrepresented skin tones, offering a scalable solution to data scarcity and bias in medical imaging.
• Rigorous evaluation of segmentation performance on a held-out set of *real* images provides compelling evidence that synthetic data augmentation genuinely improves model generalization and captures relevant structural features.
• The proposed generative AI pipeline offers a practical and efficient framework for directly mitigating skin tone bias, fostering more equitable and robust AI diagnostic tools across diverse patient populations.

Summary by alphaXiv.