How Artificial Intelligence can aid Eye Testing

Artificial intelligence (AI) has evolved into a valuable extension of the medical industry, from diagnostics to medication discovery. A high-tech screening tool developed by Google and an international team of researchers for diagnosing diabetic retinopathy, a diabetes condition in the eye, has been added to its extensive list of applications. In this article, we’ll look at how Artificial Intelligence can help with eye testing.

What role does artificial intelligence play in eye testing?

Let’s have a look at how AI can assist with eye exams!

A study of over 3,000 diabetic patients done at two eye care centers in India — Aravind Eye Hospital in Madurai and Sankara Nethralaya in Chennai — found that the AI outperformed the traditional manual grading approach for detecting diabetic retinopathy. The AI’s specificity and sensitivity were both about 90%. The findings were reported in the journal JAMA Ophthalmology.

What is diabetic retinopathy, and how does it affect you?

Diabetic retinopathy is an eye disease that can result in vision loss and blindness in diabetics. It affects the blood vessels in the retina (the light-sensitive layer of tissue in the back of your eye).

You should have a full dilated eye exam at least once a year if you have diabetes.

Diabetic retinopathy may not have any symptoms at first, but catching it early might help you keep your eyesight.

Maintaining a healthy lifestyle, which includes staying physically active, eating well-balanced food, and taking your medications, can help you avoid or delay vision loss.

That is why routine eye examinations are necessary. According to Professor Mingguang He, Principal Investigator in Ophthalmic Epidemiology at CERA, early detection of diabetic retinopathy might be considerably improved if screening was available in more GP and endocrinology clinics.

“In any case, the test is that many general practitioners lack the specialist knowledge to examine the retina – the back of the eye,” Professor Mingguang explains. He declares. AI may be able to help with this. He’s created an AI screening tool that’s both imaginative and useful.
This was created to provide GPs and endocrinologists with a simple way to recognize signs of eye disease.

Diabetic Retinopathy

They dilate the pupil to allow more light to enter the eye and brighten the rear of the eye with this technology.

The pictures of the eye were taken with a specialized retinal fundus camera.

In most cases, this procedure necessitates a visual examination of the patient’s retina.

Advantages

• It is quite simple to operate, and the camera’s price has just dropped dramatically.
• The photos are then uploaded into the computer, where they are screened for diabetic retinopathy using an AI technique.
• The AI’s specificity and sensitivity were both about 90%.
• The patient can get their eye test report, as well as the other regular diabetic test reports, in around two minutes.

Eye Care Options With AI

1. IDx-DR

IDx-DR (IDx Technologies) is an FDA-approved autonomous AI system for the identification of DR. The operator uses a fundus camera to take two images per eye, which are then sent to the IDx-DR Client.

In less than a minute, IDx-DR analyses photos for indicators of DR and offers results. Patients who test negative for anything more than mild DR are encouraged to retest in 12 months, while those who test positive for more than mild DR are referred to an eye doctor. IDx-DR showed 87 percent sensitivity and 90 percent specificity in detecting more than mild DR in fundus pictures in a clinical investigation of 900 diabetics. 2

IDx is working on algorithms that can detect disease across a variety of imaging modalities, with a focus on fundus photography and OCT, according to the company. There are prototypes for detecting AMD and glaucoma.

2. EyeArt AI Eye Screening System

To identify DR, the EyeArt AI Eye Screening System (Eyenuk) uses AI algorithms to evaluate images of the retina collected with a fundus camera. The method uses fundus images of the eye to provide automated screening in a single office visit, which comprises retinal imaging, DR grading based on worldwide standards, and report production. The DR screening findings are accessible to read and export to a PDF report in less than 60 seconds after the patient’s fundus photos have been collected and submitted to the EyeArt AI System.

A person with minimum experience can take a digital photograph of the fundus, upload it, and screen patients for DR using the EyeArt system.

3. IRIS

According to the company, IRIS (Intelligent Retinal Imaging Systems) is an FDA class 2 retinal diagnostic tool that integrates into clinical primary care workflows. IRIS was intended to detect early diabetes in individuals, improve outcomes, and lower healthcare expenditures in the long run. IRIS and Remidio Innovative Solutions established cooperation in October to combine the IRIS software and services program with Remidio’s handheld camera to allow service providers to implement a telemedicine program to check, identify, and diagnose diabetic patients retinal illness. The Remidio camera features AI-ready features such as quick image grade ability.

4. DeepMind

DeepMind has developed a prototype that scans a patient’s retina in real-time to diagnose potential abnormalities. The device first scans the retina, following which DeepMind’s algorithms evaluate the images and deliver a diagnostic and an “urgency score.” DeepMind claims that their prototype system, which has been in development at Moorfields Eye Hospital in London over the past three years, can detect a variety of disorders such as DR, glaucoma, and AMD.

30-Year Vision

Around 30 years ago, Abràmoff began looking into automating the detection of eye problems. Ophthalmologists usually identify such disorders by examining a color snapshot of the back of the eye or a cross-section of the retina taken using an imaging technique known as optical coherence tomography (OCT). But, at least at first, Abràmoff doubted that a computer program could replace a highly qualified specialist.

Since the 1950s, machine learning has shown promise in image analysis, using data and custom-built algorithms to train robots to execute jobs. Even when Abràmoff began his research 40 years later, the hardware wasn’t powerful enough to make machine learning viable for analyzing real-world medical images.

Abràmoff, on the other hand, meticulously created mathematical equations to explain specific lesions in the retina, and then wrote algorithms to identify them. He had written multiple studies on the subject by the early 2000s, and as the decade proceeded, he gained pertinent patents in the hopes of licensing them to a pharmaceutical or biotechnology business. However, the concept did not catch on. He claims that “nothing happened.”

The video game industry gave the usage of AI systems in medical imaging a tremendous boost in the late 2000s. The demand for more realistic images prompted the development of increasingly powerful graphics cards, which were excellent for the parallel processing required by AI systems.

Artificial neural networks, which are inspired by the way neurons interconnect in the brain, were made easier to create with these graphics cards. These networks are made up of layers of connected nodes that process various aspects of an image. Each attribute is assigned a weight, which the system then adds together to produce an output, such as a determination of whether or not an eye has diabetic retinopathy.

Researchers were able to develop deep-learning networks that can undertake sophisticated tasks far beyond touches of conventionally programmed software, such as beating some of the world’s best players of the ancient board game Go, by combining artificial neural networks with significant processing power and massive image data sets. “There’s been this tremendous jump ahead,” says Aaron Lee, an ophthalmologist at the University of Washington in Seattle. “All these things that used to be pie in the sky are now theoretically realistic to execute.”

Carry On As Humans Do

AI systems will have to perform more than just diagnose a particular eye illness in the future. “When a physician examines someone’s eye, they pick up on a lot of typical problems,” Wong explains. “You can’t just say, ‘I’m solely interested in seeing if you have diabetic retinopathy,'” she says. That is why Wong and others, including Abràmoff, are working on AI systems that can identify many eye illnesses at once.

Rather than teaching AI algorithms particular disease traits to look for (like Abràmoff did with IDx-DR), some researchers have them sift through a large number of photos taken from healthy and diseased eyes. The AI systems must then figure out how to distinguish between them on their own.

In 2017, Wong and his colleagues trained an AI system using retinal scans from multiple studies, including the Singapore National Diabetic Retinopathy Screening Program2. They put it to the test in 11 multi-ethnic diabetes cohorts, demonstrating that their AI engine could detect not just diabetic retinopathy but also glaucoma and AMD using variations in retinal pictures. For diabetic retinopathy, the system’s screening ability matched that of a human specialist around 90% of the time.

DeepMind and Moorfields Eye Hospital researchers have gone even further. They created an artificial intelligence algorithm that taught itself to make referral decisions for 50 common eye diseases3. In an OCT retinal scan, the system detects indicators of eye disease and determines the urgency with which a person should see a specialist.

DeepMind’s AI system might significantly reduce ophthalmologists’ patient burden. “People aren’t aware of the vast number of instances we handle,” adds Keane. Last year, the National Health Service in England planned 8.25 million outpatient ophthalmology appointments.

An AI algorithm is often trained using a massive quantity of data and is only capable of performing restricted tasks; for example, an algorithm educated to play Go by telling it to play itself 30 million times would be useless at chess. However, a technique known as transfer learning could help AI computers learn to execute similar jobs faster by requiring less task-specific data.

An ophthalmologist at the University of California, San Diego, in La Jolla, led a team that applied an AI algorithm that had been pre-trained on tens of millions of photos of daily items from the public repository ImageNet to a set of roughly 100,000 OCT retinal images4.

Despite using a small number of retina-specific images to train the system, the team’s AI program was able to accurately diagnose two common causes of vision loss — diabetic macular edema and choroidal neovascularization (often a result of advanced AMD) — and determine who needed an urgent referral to a specialist.

The algorithm’s error rate was doubled when the number of OCT retinal pictures utilized in the training was reduced to roughly 4,000, but its performance was still equivalent to that of human experts.

In the next two years, Zhang, Keane, and Wong plan to undertake clinical studies to see if their AI systems are as effective at diagnosing as ophthalmologists – a vital step before getting regulatory permission. However, the further effort will be needed to create a commercial product that can be used in a number of scenarios. “Scientists must make it as user-friendly as an iPhone,” Wong argues.

AI Applications

Artificial intelligence (AI) can be utilized as a sophisticated kind of triage. It can assist evaluate which patients with urgent illnesses should be seen by a specialist first, vs those who can wait for a normal appointment. Patients who have no symptoms or early indicators of pathology can have their screenings repeated the following year.

Another advantage of AI screening tools is that they will likely boost early detection of DR by putting the screening process in the hands of a larger spectrum of health care practitioners. Early detection leads to early intervention, which reduces the need for costly and intrusive procedures as well as follow-up visits. It can also lower the percentage of people who go undiagnosed with DR until it’s too late, resulting in DR-related blindness.

Screening may be made easier using AI technologies without sacrificing speed or accuracy. With the ability to be used by technicians or trained laypeople, it has the potential to aid patients who, for whatever reason, are unable to get to an eye care professional’s office for baseline screening. It also offers the ability to provide on-the-spot fundus imaging screening for patients who are unable to see an eye doctor due to financial constraints. This includes people who are at high risks, such as those who are HIV positive.

Increasing The Availability Of Eye Screening

Professor, He has been energized by this study to design a fully computerized, administrator-free version of the screening gadget as the stimulating next stage.

“In the future, eye disease screening could be as simple as snapping a snapshot corner,” he says.

This innovation is expected to increase access to diabetic retinopathy screening for all Australians. It has the potential to make a huge impact in territorial and rural areas where eye health administrations are lacking.

“Man-made logic may be able to close the gap in eye-care administrations. It has the potential to improve the early detection of the four most common blinding eye diseases. It will help reduce the burden of eyesight loss in the Australian population group that most needs it.”

Take, for example, Google’s DeepMind AI, which is quite active in the healthcare industry. It was prepared with the use of 15,000 OCT scans from roughly 7,500 patients.

These individuals were treated at locations run by Moorfields, Europe’s and North America’s largest eye medical clinic. Their outputs were used to maintain the framework, which was supplemented by human experts’ conclusions.

It was able to find out how to discern the many anatomical components of the eye as a result of this (a procedure known as division).

Following then, clinical activity was suggested based on the many symptoms of illnesses found in the sweeps.

Not Just A Question Of Technology

These AI systems’ abilities may, in certain situations, surpass those of humans. Bernhard Weber, a geneticist at the University of Regensburg in Germany, and his colleagues, for example, have developed a deep-learning algorithm for categorizing the progression of AMD5, a leading cause of visual loss in adults over 50. Although late-stage AMD is easily detectable, Weber discovered that his team’s AI engine could also detect the disease in its early stages. “That’s difficult stuff,” he admits, even for an ophthalmologist.

Although the accuracy of such AI systems aids in regulatory approval, it is possible that this green light will not be enough to win clinicians’ and patients’ trust. “As a culture, are we ready to put these things in place?” Lee wonders.

The secretive nature of many AI systems, which operate as black boxes, makes it difficult to acquire users’ trust. It’s not always evident how such algorithms arrive at a judgment. “With a black-box algorithm, you have no understanding why the algorithm made that diagnosis,” Lee explains.

Conclusion

Artificial intelligence (AI) will assist humans in a variety of vital tasks. The time has come for AI to establish itself as a foundation for ongoing technological advancement. Disease screening with AI offers eye care practitioners a huge potential to enhance outcomes.

One of these earliest building stones is the use of AI in eye exams. There’s still a long way to go. Finally, we have a solution to the issue, “How might Artificial Intelligence help in Eye Testing?”