Data-Centric AI for Fair and Robust Dermatology

This ongoing research project focuses on data-centric artificial intelligence for dermatology, aiming to improve the fairness, reliability, and real-world robustness of automated skin-disease analysis. Conducted in collaboration with MedirAI, a Canada-based medical-AI company, the work addresses a key limitation of current dermatology AI: models are trained on large but highly imbalanced datasets, where common conditions dominate while rare diseases and darker skin tones remain under-represented. The research follows two complementary directions. The first develops generative data-augmentation methods, centered on GAN and diffusion-based image generation, to synthesize clinically valid dermatology images that are controllable by condition, skin tone, and Fitzpatrick type, closing the gaps left by scarce real data. The second develops skin-lesion segmentation methods that accurately outline lesions despite surrounding clutter and remain robust on smartphone and edge-device images, trained on a dataset in which every mask is explicitly validated. Together, they form a general-purpose framework for equitable, deployment-ready dermatological image analysis across diverse patient populations.