A team of CNU researchers succeeded in developing an AI system that can diagnose 134 skin diseases.

It has achieved improvement in actual clinical circumstances compared to existing AI devices that only allow for a simple dichotomy of whether a tumor is benign or malignant while skin lesions generally have a wide variety of appearances.

Professor Park Il-woo (Department of Radiology) and his research team accumulated and input more than 220,000 photos of skin lesions in Asians and Westerners to the system using a deep-learning algorithm called the Convolutional Neural Network. The artificial intelligence model not only succeeded in the diagnosis of skin cancer as precisely as resident doctors in dermatology and provided suggestions on primary treatment methods such as prescription of antibiotics, but also successfully conducted the categorization of skin cancer into 134 dermatological diseases.

For the skin cancer diagnosis performance tests, 21 dermatologists and 26 residents diagnosed 3,501 photos. As a result, the diagnostic sensitivity and specificity of the specialists and residents recorded the highest score when they diagnosed skin cancer with AI’s help rather than when they diagnosed it alone. In addition, regarding the skin cancer diagnosis of 23 non-medical patients, the sensitivity was almost doubled.

Another test measured diagnosis capabilities by comparing the success rate of skin disease recognition of AI-only, physician-only, and AI-assisted physician methods. Among these, AI-assisted physicians were found to have the highest diagnostic abilities in both skin cancer diagnosis, treatment method presentation, and the identification of 134 diseases. Therefore, AI-supported doctors turned out to be the most effective in diagnosing skin diseases.

Professor Park said, "In medicine, AI has shown that it can play a complementary role rather than substitute for doctors, and that AI can play the role of helper to improve the diagnostic ability of medical staff."

The research team released a beta version that is available on PCs and smartphones on modelderm.com so that researchers can test and exchange opinions.