A study conducted by MIT researchers has discovered that doctors have trouble accurately diagnosing skin diseases in people with darker skin based solely on images. The study consisted of more than 1,000 dermatologists and general practitioners and the results found that dermatologists accurately diagnosed about 38% of images examined. However, when considering those with darker skin, the accuracy was reduced down to 34%.
The disparity in accuracy between lighter and darker skin can possibly be attributed to dermatology textbooks and training materials featuring mostly lighter skin tones, according to the researchers. This, along with the possibility that some doctors may lack experience in treating patients with darker skin, could be contributing factors to the discrepancy.
The study also revealed that incorporating the assistance of an AI algorithm can enhance the accuracy of diagnoses made by doctors. However, the improvements were notably greater when diagnosing patients with lighter skin.
Darker skin images for diagnosing skin diseases were compiled for the study, which showed that both dermatologists and general practitioners had a decrease in accuracy of about four-percentage points when trying to diagnose darker skin conditions. This drop in accuracy was statistically significant.
A surprising discovery was dermatologists being less likely to refer darker skin images of a specific cancer (cutaneous T-cell lymphoma) for biopsy. Conversely, they were more likely to refer them for biopsy for noncancerous skin conditions.
An AI algorithm built by the researchers was applied to a second set of images and resulted in an overall improvement in both dermatologists’ and general practitioners’ accuracy in diagnosis. However, general practitioners showed greater improvement on images of lighter skin than darker.
The researchers hope their study could encourage medical schools and textbooks to include training on patients with darker skin to bridge the diagnostic gap. Moreover, the findings could help in the future deployment of AI assistance programs in dermatology.
The study highlights the need for empirical evidence to refine policies around dermatology education and the utilization of AI to aid in medical decision-making. Although dermatologists improved their accuracy when diagnosing both light and dark skin using AI, general practitioners showed less improvement when diagnosing darker skin, indicating the need for more experience and training in this particular area.
