Key Takeaways
- The atlas catalogued 1.1 million images across 29 public dermatology datasets
- Only 11.0% of images show Fitzpatrick skin types V–VI, and only 2.3% are paediatric
- Informational novelty has plateaued even as the number of images grows exponentially
The datasets teaching dermatology AI barely include dark skin — or children
Modern dermatology increasingly runs on image datasets — the photo libraries used to train diagnostic tools and AI systems. A study published November 21 in Nature Communications mapped that library worldwide, and found it badly skewed.
The team catalogued 1.1 million images across 29 public datasets. Only 11.0% of the images show Fitzpatrick skin types V–VI — the darkest skin tones. And only 2.3% are paediatric images, even though skin disease is a major part of childhood illness.
The researchers also found that while the number of images has grown exponentially, the informational novelty has plateaued — the libraries keep adding more of the same kinds of images rather than filling the gaps.
This study mapped the data itself; it did not test how any specific AI model performs on different skin tones, and no model-accuracy figures can be drawn from it. But the implication is structural: a diagnostic tool can only learn what its training library shows it, and the library underrepresents exactly the patients most at risk of being missed — people with dark skin, and children.
The Twikup Take
Before asking whether dermatology AI is biased, this atlas asks a simpler question: what has it been shown? 1.1 million images sounds like a lot — until you learn that nearly nine in ten show lighter skin, and barely one in forty shows a child.
Image: Illustrative image.
