Timely glaucoma mapping with AI: efficient, accurate and fast | MedNet

Easily detect glaucoma with a ‘smart’ algorithm; music in the future or will it come true soon? In the case of ophthalmologist Dr. Hans Lemij (Rotterdam Eye Hospital), the latter is true. He and his colleagues developed a glaucoma detection model based on artificial intelligence and explains why this is such an effective and important method.^1,2 “The years of blindness and vision loss associated with glaucoma can be so frustrating for patients. So prevention is a high priority and with our model it is possible. “

Glaucoma is a common condition and its prevalence will only increase in the coming years. Estimates indicate that there will be 112 million glaucoma patients worldwide in 2040. In the Netherlands, approximately 150,000-300,000 people have glaucoma. The effects of the situation can be catastrophic. “Blindness (full or partial) is a major risk factor from glaucoma,” Lemij said. “This means that timely detection and treatment of this eye disease is essential. The sooner glaucoma is detected, the better.” He explains that although glaucoma is very treatable, eventually about 10% of people with glaucoma will become blind in both eyes and 25% in 1 eye. “These percentages do not increase but the glaucoma later found. Another problem we encounter is that the number of ophthalmologists is not keeping up with the increasing number of glaucoma cases. In short: there is not enough manpower to diagnose and treat all patients in a timely manner.”

Lost diagnosis

It is therefore important to detect glaucoma early, but it is not always easy. “We know from Australian research that around half of people with glaucoma were aware of this. What is special, however, is that everyone who knew him had been to an ophthalmologist or optometrist in the previous year who had missed a glaucoma diagnosis. To be honest, I don’t think the situation is very different in the Netherlands. If professionals miss the diagnosis, other methods may be needed to diagnose glaucoma.” Lemij says that fundus photography is suitable for this, especially in combination with artificial intelligence (AI). “Chinese research groups were in particular has been working on this for a while. It actually turned out that fundus pictures can be accurately assessed for the presence or absence of glaucoma. However, this only applied to eyes Chinese. If the same method was applied to non-Chinese eyes, this method was very inferior.”

Classification of fundus images

For the application outside of China, the AI ​​system had to be taught to recognize glaucoma in non-Chinese eyes using a multi-ethnic set of fundus images. In an American company that takes a lot of fundus images for screening for diabetic retinopathy, Lemij and his colleagues found a partner that could provide this imaging material. “These images were so good that we assumed they could also detect glaucoma.” The images had to be classified first before they could be included. to the AI ​​model. However, it was impossible to evaluate the number of fundus images – more than 100,000 – by one person. We have trained several optometrists and ophthalmologists at home to do this correctly of the 91 applicants, 30 remained after the mandatory test. In the end, only 20 people were good enough; they reviewed the pictures.”

AI model development

With this extensive set of fundus images that would form the basis for the AI ​​model, a ‘challenge’ was then organized on the advice of the University of Amsterdam, with which Lemij and his colleagues were working. “By pitting different AI developers against each other in a competition, the company that could best develop this model would come out.” The enthusiasm was great; 351 companies from 51 countries participated. The winner developed a model that achieved a sensitivity of 86% with a specificity of 95%.² “That specificity has to be high,” Lemij explains. “There has to be a clear distinction between false and normal; after all, health care is overrun with false positives ; an undesirable situation.”

A population-based study of glaucoma

So the AI ​​model appeared to be very capable of detecting glaucoma based on fundus images.^1,2 “That gave us a lot of confidence. What we really want to do now is that detection with this AI model will lead to a population study like it has been for breast and colon cancer for some time,” said Lemij . However, such a step is not easy. For example, the Ministry of Health, Welfare and Sport (VWS) sets 20 criteria for population screening that must be met. Among other things, large-scale research must be shown to be based on the latest science and practice and to be cost-effective. Lemij: “We are currently doing a cost-effectiveness study together with Erasmus University. Another important point is that the research used works well with the Dutch population. That may seem obvious, but our model was tested on images taken from the American population, which is very ethnically diverse. I definitely think we will pass this test, but we have to show it. ”

Other uses of AI in ophthalmology

In addition to glaucoma, there are 2 other important common eye conditions in which AI plays a role: diabetic retinopathy and macular degeneration. The use of AI in macular degeneration is still in its infancy. This is more developed in diabetes. “With the help of cameras that can detect retinopathy, it is possible to screen a large number of people. There are already cameras with AI that detect glaucoma, diabetic retinopathy and macular degeneration. However, it is not happening or not enough. ” Funding is the biggest hurdle here. “There is structured funding from health insurers for diabetic retinopathy because this screening is part of the routine examination for patients with diabetes. This is not available (yet) for macular degeneration and glaucoma.” Lemij believes that the reimbursement process, the unknown and the complexity of the certification procedure (‘Every time a change is made to the procedure, everything has to be certified again ‘) are important reasons why people are wary of using AI to screen the population. “The technology and infrastructure is already in place.

A new challenge will soon be organized to develop/use software that can show the cause of glaucoma in glaucoma patients (the so-called ‘explanatory’ AI). Interested parties may contact Dr. Hans Lemij via H.Lemij@oogziekenhuis.nl.

References:

1. Lemij HG, De Vente C, Sánchez CI, et al. Features of a large labeled data set for training artificial intelligence for glaucoma screening with Fundus images. Ophthalmol Sci. 2023; 17; 3:100300.
2. De Vente C, Vermeer KA, Jaccard N, et al. AIROGS: Artificial intelligence for the robust glaucoma screening challenge. IEEE Trans-Med Imaging. 2024; 43:542-57.