Today we had the pleasure of hosting Dr. Jane Courtney from the School of Electrical Engineering in Dublin for a thought-provoking talk on why AI can fail in image processing and what that means for healthcare applications.
Besides other topics, she focused on a critical issue: skin type diversity in skin lesion datasets. Drawing from her recent review (Alipour, Burke & Courtney, Current Dermatology Reports, 2024), Dr Courtney showed that the datasets used to train AI for skin cancer detection often lack verifiable skin-type diversity. The result are models that work well on lighter skin but consistently underperform on darker skin tones, which is exactly the populations already facing healthcare disparities.
Her message was that if we don’t fix diversity in training data, we bake inequality directly into AI tools. Better reporting, transparent benchmarks, and inclusive dataset design aren’t optional extras – they’re essential for safe, fair, and effective AI.
