AI Skin Cancer Tools Show Disparities for Darker Skin Tones

New artificial intelligence tools designed to aid in skin cancer detection are exhibiting notable performance disparities, particularly concerning different skin tones. While these AI applications promise greater accessibility to medical screenings, their effectiveness is proving inconsistent across diverse populations.

AI models function by learning to identify diseases based on visual patterns. Studies indicate that models predominantly trained on images of lighter skin tones lose accuracy when presented with images of darker skin. This issue arises because the AI may incorrectly rely on skin color as a shortcut for diagnosis, rather than focusing on the inherent characteristics of the skin lesion itself.

For instance, conditions like atopic dermatitis can manifest differently on light versus dark skin. AI may reliably identify symptoms on lighter skin but fail to recognize the same condition on darker skin. This discrepancy can lead to suboptimal care for individuals with darker skin tones. This is especially critical as certain skin cancers, like melanoma, are more challenging to detect on pigmented skin.

Researchers have observed that even large language models, such as GPT-4, can misclassify benign moles as malignant melanoma when depicted against darker skin tones. This can result in unnecessary patient anxiety or, conversely, a failure to identify life-threatening cancers.

The skewed nature of AI training data is a primary driver of this bias. Medical image databases and textbooks have historically contained more images of lighter-skinned individuals. To rectify this, more diverse datasets featuring a substantial number of images from patients with various skin tones are necessary. Generative AI techniques can be employed to create synthetic images, augmenting training datasets and preserving patient privacy.