AI Skin Cancer Tools Show Racial Bias in Diagnostic Accuracy

Artificial intelligence tools designed to detect skin cancer are proliferating in dermatology clinics and smartphone apps, but research reveals a critical flaw: they work substantially better for patients with light skin than for those with darker complexions.

The diagnostic gap stems from how these AI models learn. Rather than focusing solely on the characteristics of suspicious lesions, the systems use surrounding skin color as a pattern-matching shortcut. When researchers at the University of Tennessee digitally darkened the skin tone in medical images while leaving the actual lesion unchanged, the AI's diagnostic accuracy dropped sharply.

Why it matters

Patients of color already face worse outcomes in melanoma care, with diagnoses typically occurring at more advanced stages and lower survival rates. AI tools that perform poorly on darker skin threaten to widen existing health disparities at scale, especially as these technologies deploy in clinics and consumer apps worldwide without adequate testing across diverse populations.

The training data problem

The bias traces directly to the image databases used to train these models. Historical medical photo libraries and dermatology textbooks have predominantly featured lighter skin tones, reflecting clinical norms developed around white patients. When an AI system never learns what melanoma looks like on dark skin during training, it cannot reliably identify it in practice.

Researcher Mohamed Akrout demonstrated the real-world risk by testing GPT-4 with images of benign moles. When the surrounding skin was digitally darkened while the mole remained identical, the AI misclassified the harmless spot as malignant melanoma. The reverse problem also occurs: life-threatening cancers on darker skin can be overlooked entirely.

Conditions like atopic dermatitis illustrate the challenge. The inflammatory disease causes pink discoloration on light skin but appears gray or violet on darker skin. AI models reliably flag the pink presentation but often miss the darker manifestations.

Synthetic data offers limited help

Some researchers have explored using generative AI to create synthetic medical images of diverse skin tones, bypassing patient privacy concerns. Akrout's team showed that models trained entirely on synthetic images can match the performance of those trained on real photos.

However, this approach carries risk. Generative AI may produce visually convincing images that don't accurately represent how diseases actually manifest in real patients of color. Training diagnostic tools on these flawed synthetic datasets could create systems that appear diverse on paper but remain functionally blind to actual patient presentations.

The path forward

Researchers and regulators are pushing for mandatory testing across all skin tones before AI dermatology tools receive widespread deployment. The medical AI field faces a fundamental choice: build truly representative image collections from diverse patient populations, or continue deploying tools that work reliably only for some patients.

Eliminating color-based bias isn't about fairness alone—it's the baseline requirement for these tools to function as medical devices. Without inclusive training data, AI skin cancer detection remains a technology that could save lives for some while missing deadly cancers in others.

These findings were first reported by Mohamed Akrout, assistant professor of electrical engineering and computer science at the University of Tennessee, writing in The Conversation.