AI skin cancer detection tools are getting better – but only for people with light skin
Melanoma and other skin conditions can look different in people with darker versus lighter skin – but image databases include many more examples from people with lighter skin. SeventyFour/iStock via Getty Images Imagine you’re getting out of the shower one morning and you notice a mole on your thigh that you’ve never seen before.
New artificial intelligence (AI) tools are being developed to help detect skin cancer, including melanoma, a serious form of skin cancer. These tools include smartphone apps for home use and software for doctors. They analyze images of moles or skin lesions to identify potential signs of cancer. AI models work by learning patterns from large datasets of medical images, associating visual features with specific diseases. The goal is to provide accessible medical expertise, especially in areas with few dermatologists, potentially saving lives by enabling early detection. However, current AI tools have a significant limitation: they perform well on light skin but struggle with darker skin tones.
AI models rely on pattern recognition, learning to associate visual features with diseases by analyzing thousands of images. However, they can mistakenly focus on irrelevant details, such as the color of surrounding skin, rather than the lesion itself. For example, researchers trained an AI model on images of moles on light skin and then digitally darkened the skin in the images. The AI's accuracy in diagnosing melanoma dropped sharply, even though the lesion itself had not changed. This happens because the AI uses skin color as a shortcut, leading to misdiagnoses. Conditions like atopic dermatitis (a chronic skin disease) appear differently on darker skin (gray or violet) compared to light skin (pink), further reducing AI's reliability for darker skin tones.
The bias in AI skin cancer detection tools has serious consequences for patients with darker skin. These patients are already more likely to be diagnosed with melanoma at later stages, which reduces survival rates. An inaccurate AI tool worsens this disparity by providing unreliable results specifically for darker skin, leading to poorer care. For instance, a study showed that OpenAI’s GPT-4, when presented with an image of a benign mole on darkened skin, incorrectly classified it as malignant melanoma. This could cause unnecessary panic or, conversely, overlook actual cancer. The risk is higher for home users relying on AI chatbots like ChatGPT for medical advice without professional oversight.
The bias in AI tools stems from the datasets used to train them. Medical image databases and dermatology textbooks have historically included far more images of light skin than darker skin. This underrepresentation means AI models are not adequately trained to recognize diseases on darker skin. For example, melanoma is harder to spot on pigmented skin, and without diverse training data, AI tools fail to learn the correct diagnostic features. Addressing this requires collecting more images of darker skin conditions, but ethical and privacy concerns make gathering such data challenging. Without diverse datasets, AI tools remain blind to how diseases appear on people of color.
One proposed solution is using generative AI to create synthetic medical images of skin conditions on darker skin tones. This approach could bypass privacy issues by generating large datasets without using real patient images. Researchers have shown that AI models trained on synthetic images can perform as well as those trained on real images. However, this method has risks. Generative AI may produce images that do not accurately reflect real-world conditions, leading to flawed training data. For example, synthetic images might not capture the true characteristics of melanoma on darker skin, rendering the AI tool ineffective. Ultimately, the only reliable solution is to build inclusive, representative datasets of real images from people with darker skin tones.
AI skin cancer detection tools are already being used in some clinics and app stores, but researchers and regulators are pushing for stricter testing to ensure accuracy across all skin tones. Studies and initiatives, such as those by the FDA and Colorado’s AI medical device bias audit, aim to address these disparities before widespread deployment. Eliminating color-based bias in AI is critical not only for fairness but also to ensure these tools work effectively for everyone. Without diverse training data and rigorous testing, AI tools risk exacerbating existing healthcare inequalities, particularly for patients with darker skin who already face higher risks of late-stage melanoma diagnoses.
