The study, published Tuesday in The Lancet-Digital Healthcare, examined 21 freely accessible skin condition image data sets. These data sets contain a total of more than 100,000 pictures. Only more than 1,400 of these images have information about the patient’s race, and only 2236 have information about the skin color. This lack of data limits the ability of researchers to find deviations in algorithms trained on images. And this algorithm is likely to be biased. Among the images with skin color information, only 11 are from the two deepest categories of patients in the Fitzpatrick Skin Scale, which classifies skin color. There are no images of patients from African, African-Caribbean, or South Asian backgrounds.
These conclusions are similar to a study published in September, which also found that most data sets used to train dermatological algorithms have no information about race or skin color. The study examined the data behind 70 studies that developed or tested algorithms and found that only 7 described the skin types in the images used.
Roxana Daneshjou, a clinical scholar in dermatology at Stanford University and the author of a paper published in September, said: “What we have seen from the few papers that report the distribution of skin color is that those papers do show that dark skin is underrepresented.” The paper analyzed many of the same data sets as the new research of The Lancet and reached similar conclusions.
When the images in the data set are public, researchers can go to see which skin tones seem to exist. But this may be difficult, because the photo may not exactly match the skin tone in real life. “The ideal situation is to notice the skin tone during clinical visits,” Daneshjou said. Then, the image of the patient’s skin problem can be tagged before entering the database.
If there are no labels on the images, researchers cannot check the algorithms to see if they use a dataset with enough examples of people with different skin types.
It is important to examine these image sets carefully, because they are often used to build algorithms to help doctors diagnose patients’ skin conditions, some of which – such as skin cancer – are more dangerous if they are not detected early. If algorithms are only trained or tested on light-colored skin, they will not be as accurate for others. “Research shows that a procedure that only trains images of people with light skin types may not be as accurate for people with dark skin, and vice versa,” said David Wen, a co-author of the new paper and a researcher at the University of Oxford. Say.
New images can always be added to public data sets, and researchers hope to see more examples of dark skin conditions. Increasing the transparency and clarity of data sets will help researchers track the progress of more diverse image sets, which may lead to fairer AI tools. Daneshjou said: “I want to see more open data and more carefully labeled data.”