AI technology could help improve doctors’ accuracy when diagnosing skin diseases in individuals with darker skin, scientists from the Massachusetts Institute of Technology have found. The study involved over 1,000 dermatologists and general practitioners, with dermatologists correctly identifying roughly 38% of skin conditions seen in images, but this figure dropped to 34% where patients had darker skin. General practitioners, who were less successful in general, also saw a similar decrease in accuracy when it came to darker skin tones. It was also revealed that textbooks typically displayed individuals with lighter skins, potentially contributing to the problem.
The study’s lead author, Matt Groh, who is also an assistant professor at Northwestern University, stated that the lack of experience and understanding when dealing with darker skin might worsen the performance of physicians. According to him, the research provides the empirical evidence required to influence policy change around dermatology education. Several years ago, a different study found that facial recognition technology was less effective at identifying the gender of people with darker skin. This was the foundation for the current research, which investigated if AI and human physicians shared this problem, and if the diagnostic disparity could be rectified.
364 images of 46 skin conditions across various skin shades were compiled from different sources to test the doctors’ judgment. Eight inflammatory illnesses, like atopic dermatitis, Lyme disease, and secondary syphilis, as well as a rare type of cancer known as cutaneous T-cell lymphoma were represented in the majority of these images. These often present differently on lighter and darker skin. The study’s final group consisted of 389 dermatologists, 116 dermatology residents, 459 general practitioners, and 154 other physicians. The participants were asked to diagnose 10 diseases from images and decide if they should be referred for biopsy.
Unsurprisingly, dermatologists surpassed general physicians in the diagnostic process, with an accuracy rate of 38%, compared to 19% for general practitioners. However, both groups showed a decline in accuracy for images of darker skin.
The researchers offered additional images to the doctors and allowed them to use an AI algorithm to assist with their diagnoses. This algorithm was trained on around 30,000 images to identify one of the nine categories including the eight most represented diseases and an additional “other” category. The algorithm proved to be equally effective on both light and dark skin and managed to boost accuracy rates to 60% for dermatologists and 47% for general physicians. Interestingly, doctors were likelier to consider the higher-accuracy algorithm after a few successful diagnoses but typically disregarded incorrect AI suggestions.
