Researchers at Seoul National University Hospital said that they have developed an XAI (explainable artificial intelligence), which can diagnose glaucoma.
Glaucoma is a leading ophthalmic disease that causes blindness. It can progress without specific symptoms, and if patients miss their treatment period, recovery may be impossible, while the cost of treatment can also increase exponentially. Therefore, early treatment based on accurate diagnosis is essential for the disease.
The research team, led by Professors Park Ki-ho and Park Sang-min, trained the AI for the diagnosis of glaucoma by analyzing 6,000 sheets of fundus images, obtained from the patient cohort treated at Seoul National University Hospital Health Promotion Center, three times.
Afterward, to explain the diagnosis of glaucoma, the team applied an adversarial explanation methodology to the completed AI.
The adversarial explanation is a method that uses the prediction deep learning model (DLM) to make counterexamples as a means to explain the DLM’s understands of glaucoma-specific features.
By using this method, doctors can confirm that the glaucoma findings are added or emphasized in the adversarial example, thereby confirming the understanding of AI's glaucoma.
Conversely, if the AI generates an adversarial example that is normal, the doctors can confirm that the actual glaucoma findings are removed.
As a result, the team confirmed that the glaucoma findings changed through an adversarial example.
As a result of a survey conducted on ophthalmologists, the team found that using an adversarial example helped doctors better understand the AI compared to conventional methods.
“The AI explanatory methodology developed this time has higher explanatory power than the existing Heatmap method, which only showed the location of lesions. It also shows the medical findings that led to the diagnosis of glaucoma,” Professor Park Ki-ho said.
Professor Park Sang-min said, “The new explainable AI technology solves the black box problem of existing AI programs and increases reliability.”
Therefore, it is a vital source technology for the future authentication and usability of AI medical devices, he added.
The journal, Ophthalmology, published the results of the study.