New research suggests that AI systems being developed to diagnose skin cancer may be less accurate for people with dark skin.
According to the latest study, there are only a few images of darker skin in datasets where skin tone information is accessible, thus algorithms constructed using these datasets may not be as accurate for persons who aren't white.
David Wen, the study's first author from the University of Oxford, and colleagues detail how they found 21 open-access datasets for skin cancer photos, 14 of which specified their place of origin, in the journal Lancet Digital Health. Only images from Europe, North America, and Oceania were featured in 11 of them.
Few of the 21 datasets recorded the ethnicity or skin color of the people photographed, which means it's uncertain how generalizable algorithms based on them would be, according to the team.
Only 2,436 photos out of a total of 106,950 in the 21 databases had skin type documented, according to the researchers. Only 10 of these photographs were from people who were recorded as having brown skin, and one was from someone who was recorded as having dark brown or black skin.
Only 1,585 photos contained ethnicity information instead of, or in addition to, skin type information.
"No images were from individuals with an African, African-Caribbean or South Asian background," the team reported.
Wen believes the exclusions are unlikely to be intentional, but that guidelines are needed to guarantee that vital information, such as ethnicity or skin color, is presented alongside photos. The authors go on to say that datasets utilized to develop AI systems should reflect the populations that will be served by the technology.
It's crucial to assess these image sets because they're frequently used to create algorithms that help doctors diagnose patients with skin disorders, some of which are more harmful if not discovered early, such as skin malignancies. The algorithms will not be as accurate for everyone else if they have only been trained or tested on light skin.
Researchers want to see more examples of conditions on darker skin, and new photographs can constantly be submitted to public datasets. Furthermore, increasing the datasets' transparency and clarity would aid researchers in tracking progress toward more varied image sets, which could lead to more equitable AI technologies.
