A study led by researchers at the Massachusetts Institute of Technology (MIT) has revealed that physicians’ success at diagnosing skin diseases using images is lower when the subject has darker skin. The study documented the accuracy of over 1,000 dermatologists and general practitioners at diagnosing diseases based on images, and found that while dermatologists correctly identified about 38% of the images, this figure dropped to 34% for pictures illustrating darker skin. A similar trend was observed with general practitioners.
This reduction in accuracy may be due to a lack of experience treating patients with darker skin or inadequate representation of darker skin tones in dermatology education resources, according to the researchers. Notably, this is the first study to highlight the disparity of diagnosis across skin tones.
The research team also found that diagnostic accuracy could be improved with the help of an AI algorithm, though the effects were larger when identifying lighter-skinned patients’ diseases. Assistant professor Matt Groh, one of the study’s authors, emphasized the need for empirical evidence to facilitate changes in dermatology education policies.
A comprehensive database of skin images was used in the study to assess doctors’ diagnostic abilities. The skin conditions represented in these images often present differently on dark and light skin. Dermatologists accurately classified 38% of images, general practitioners only 19%. Both groups’ accuracy dropped by approximately four percentage points when diagnosing conditions on darker skin, a statistically significant reduction.
To further enhance the study, dermatologists and general practitioners were given additional images to analyze, aided by an AI algorithm that the team developed. This algorithm, trained on about 30,000 images, boasted an accuracy rate of around 47%. In testing, the use of this algorithm increased the diagnostic accuracy of both dermatologists and general practitioners. However, general practitioners showed greater improvement when diagnosing on images of lighter skin than darker skin.
The authors hope that their findings will stimulate medical schools to increase training on patients with darker skin and highlight the potential for AI programs in dermatology. This study was financially supported by the MIT Media Lab Consortium and the Harold Horowitz Student Research Fund.
