Doctors struggle to diagnose skin disease accurately in patients with darker skin tones, according to new research from MIT. The study found that dermatologists correctly diagnosed images of skin disease in just 34% of cases involving darker skin, compared with 38% involving lighter skin. The researchers also examined the effect of AI assistance on diagnostic success rates. This led to a notable improvement, though it was more significant for patients with lighter skin. Lead author, Matt Groh, believes inadequate representation of darker skin tones in dermatology resources and a lack of experience in treating patients with darker skin may contribute to this disparity. He suggests empirical evidence is needed to influence potential changes in dermatology education policies.
In a previous MIT study, facial analysis programs were found to make more mistakes with darker skin tones. This led to the new research, which looked to assess the same in medical diagnostics. Various images were used for the research, with a particular focus on inflammatory skin diseases. These images were presented to nearly 1,200 doctors, who were asked to provide their diagnosis and whether they would suggest a biopsy. Dermatologists had a higher accuracy rate of 38%, compared to 19% for general practitioners.
The AI algorithm developed by the researchers was trained on around 30,000 images. It managed an overall accuracy rate of 47% and helped increase diagnostic success amongst dermatologists to 60% and GPs to 47%. Importantly, both of the AI models were equally accurate across different skin tones.
The researchers hope this study will encourage better representation of darker skin tones in medical textbooks and guide the evolution of AI assistance programs in dermatology. The study was funded by the MIT Media Lab Consortium and the Harold Horowitz Student Research Fund.
