AI skin cancer diagnoses risk being less accurate for dark skin – study

Research suggests that AI systems for diagnosing skin cancer could be less accurate in people with darker skin.
AI's potential has led to advances in healthcare. Some studies suggest that image recognition algorithms based on machine learning algorithms could classify skin cancers just as well as human experts.

NHS trusts are exploring AI to assist dermatologists in triaging patients with skin lesions.

Researchers say that more must be done to ensure that the technology is beneficial to all patients. They discovered that very few image databases are available that can be used to "train" AI systems to diagnose skin cancer. The few that have images of people with darker skin are rare.

Dr David Wen, the University of Oxford's first author of this study, stated: "You might have a situation in which the regulatory authorities say that the algorithm was only trained on images from fair-skinned persons, so you can only use it for those individuals. This could result in certain populations being excluded for algorithms that are approved to be used for clinical purposes.

"Alternatively, regulators may be more flexible and say that you can use it [on any patient]", but algorithms might not work as well on patients who don't have enough images to train.

The team stated that this could lead to other complications, such as the risk of unnecessary surgery or missing treatment for cancers.

Wen and his colleagues published their findings in Lancet Digital Health. They report on 21 open-access datasets that contained skin cancer images. 14 of these datasets included the country of origin. 11 of these datasets contained images from Europe, North America, and Oceania.

The 21 datasets did not include information about the skin color or ethnicity of the people photographed. The team noted that this means that it is difficult to create general algorithms based on these data.

Only 2,436 images out of the total of 106.950 images in the 21 databases were found to have skin type information. Only 10 of the 21 images contained skin types that were recorded by the team included people with brown skin, and one image was from someone with dark brown or black skin.

1,585 images only contained data on ethnicity, and not information on skin type. The team reported that no images were taken from people with an African, African Caribbean or South Asian heritage.

They add that "Coupled to the geographic origins of the datasets, there was massive over-representation skin lesion images from darkerskinned people."

Wen stated that the omissions were unlikely to have been intentional, but that standards are needed to ensure important information such as ethnicity and skin type is reported with images. The datasets used in the development of AI systems should be representative of the population where it will be used.

Charlotte Proby, a professor of dermatology at University of Dundee and spokesperson for British Skin Foundation, said that the findings are concerning.