A research team at Seoul National University Bundang Hospital (SNUBH) has developed an artificial intelligence (AI) program that can diagnose 134 different types of skin diseases, the hospital said Monday.

Professor Na Jung-im

"The 134 diseases include most of the common skin diseases, and this is the first time AI has been announced that can diagnose more than 100 skin diseases," the team said in a press release. "The diagnosis number is far larger than that of the one developed by a multinational pharmaceutical company, which diagnosed only 26 disease groups."

With the competitiveness of local AI technology falling behind other countries, the team has managed to prove that Korea has a world-class level in the field of AI-based skin disease research, the team added.

Because the lesions of skin diseases are diverse in appearance, it was difficult to apply conventional diagnostic AI to an actual clinical setting as it could stop at detecting a limited number of conditions, it said. Also, such AI software mainly provided simple classification, such as identifying the malignancy of skin tumors.

"For example, showing atopic dermatitis pictures to conventional AI software, trained to distinguish between benign and malignant skin tumors, sometimes produced misleading results as the platform misdiagnosed it as malignant," the team said. "Even though a disease is easily distinguishable to even non-experts’ eyes, the existing software could not correctly diagnose it if not properly trained."

To develop AI that can classify and diagnose more skin diseases, the team, led by Professor Na Jung-im, trained its AI software with 220,000 Asian and Western skin lesions using a convolutional neural network (CNN).

The developed AI model, while not being as accurate as a dermatologist, was able to diagnose skin cancer as accurately as medical residents, and suggest primary treatment methods such as antibiotics as well as classifying 134 skin diseases.

When 26 residents and 21 dermatologists used the developed software for diagnosing 3,501 image data for skin cancer, the sensitivity of the diagnosis significantly increased from 77.4 percent without the AI platform to 86.8 percent, with the AI platform.

Also, when the team used the software with 23 non-medical people, the sensitivity increased from 47.6 percent to 87.5 percent with the help of the AI.

"Unlike previous studies that simply compared AI's diagnostic capabilities with doctors, this study showed that the AI's diagnostic capabilities could double the detection rate of skin cancer for non-medical personal," the team said. "Among the doctors, AI, and AI-assisted doctors, the AI-assisted doctor group was found to have the highest diagnostic ability."

Therefore, AI support is the most efficient way to diagnose skin diseases, the team added.

Professor Na said, "Although the focus and composition of the picture affect the accuracy of, we can supplement this problem by human intelligence. Consequently, medical staff can help diagnose skin disease more accurately and quickly with the help of our platform."

The team expects that AI and doctors will complement rather than replace each other, and become a useful assistant in improving the doctor's ability to make a more accurate diagnosis, Na added.

Na said the team plans to conduct additional studies to see if these algorithms can become commercially available. "If so, it will help patients visit dermatologists early, as the general public could check for skin cancer in real-time using a smartphone or PC without special equipment," Na said.

Journal of Investigative Dermatology published the results of the study in its latest edition.

Copyright © KBR Unauthorized reproduction, redistribution prohibited