AI Eye Screening: Diabetic Retinopathy, Glaucoma, Macular Degeneration

A quick photo of the back of your eye, checked by approved AI software, can now spot diabetic eye damage in a family doctor's office. Here is what these tools can and cannot do, and what is still research.

The problem today: eye disease

Three common eye diseases can steal sight in adults: diabetic retinopathy (damage to the tiny blood vessels at the back of the eye caused by diabetes), glaucoma (slow damage to the optic nerve, the cable that carries signals from the eye to the brain) and age-related macular degeneration, or AMD (wear and tear of the macula, the small central part of the retina that you use for reading and recognising faces).

All three have something cruel in common. In the early stages they usually cause no pain and no change in vision that you would notice. By the time someone says "my eyesight is getting worse", some of the damage may already be permanent. That is why doctors rely on screening: taking pictures of the retina (the light-sensitive layer at the back of the eye) in people who feel fine, and looking for early warning signs.

The trouble is that screening needs trained people to look at a very large number of images. People with diabetes are advised to have regular eye checks, but many miss them. Reasons include waits, travel, cost and not knowing the check matters. In many parts of the world there are too few eye specialists to read every picture.

This is where artificial intelligence (AI) comes in. AI here means computer software that has learned to recognise patterns in medical images by studying very large collections of pictures that experts had already labelled. Some of these programs are now approved medical devices. Others are still research tools. A newer field called oculomics goes one step further: it uses eye scans to estimate the risk of diseases elsewhere in the body, such as heart disease, kidney disease and dementia.

How AI eye screening works

The retina is the only place in the body where doctors can look directly at blood vessels and nerve tissue without cutting the skin. That makes it a useful window. Two main types of picture are used.

• Fundus photography: a colour photograph of the back of the eye, taken through the pupil with a special camera. It shows the retina, its blood vessels and the optic nerve head. It takes seconds and the camera does not touch the eye.
• Optical coherence tomography (OCT): a scan that uses harmless near-infrared light to build a cross-section of the retina, a bit like slicing a cake to see its layers. It shows swelling, fluid and thinning that a flat photo can miss.

Diabetic retinopathy

In diabetes, high blood sugar damages small blood vessels. On a photo this shows up as tiny bulges in vessel walls, small bleeds, fatty deposits and pale patches where tissue lacks oxygen. AI software is trained to spot these signs. Most approved systems give a simple answer for each patient: either "refer to an eye specialist" (because disease that needs a closer look is present) or "rescreen in the usual interval".

Glaucoma

Glaucoma damages the optic nerve. On a photo, the central "cup" of the nerve head gets larger compared with the whole disc, and the rim of nerve tissue gets thinner. On OCT, the layer of nerve fibres around the optic nerve becomes thinner. AI tools measure these features and compare them with what is normal for healthy eyes. Glaucoma is harder for AI than diabetic eye disease, because a large cup can be normal in some people and because a full diagnosis also needs eye pressure and a visual field test (a test of your side vision).

Macular degeneration

In the early "dry" form of AMD, yellow deposits called drusen build up under the retina. In the late dry form, called geographic atrophy, patches of retinal cells die and leave blind spots in the centre of vision. In the "wet" form, abnormal blood vessels grow and leak fluid under the macula. AI tools can count and measure drusen on photos and OCT, and can detect fluid on OCT scans, which helps doctors decide when an injection is needed.

Oculomics: the eye as a window to the body

The small blood vessels in the retina share features with vessels in the heart, brain and kidneys. High blood pressure, diabetes and hardening of the arteries can change how retinal vessels look: their width, how twisted they are and how they branch. The retina also develops from the same early tissue as the brain. Researchers have therefore trained AI to look for patterns linked with heart attack, stroke, kidney disease and conditions such as Parkinson's disease.

A big step came in 2023, when researchers at Moorfields Eye Hospital and University College London (UCL) published RETFound in the journal Nature. RETFound is a foundation model: a general-purpose AI that first taught itself what normal and abnormal retinas look like from about 1.6 million unlabelled images (904,170 colour photos and 736,442 OCT scans). It can then be adapted to many tasks with relatively few labelled examples. In 2025, UCL announced a global version trained on more than 100 million fundus photos from over 65 countries, aimed at working better for people of every background.

The main tools and where they stand

AI eye tools fall into two groups. Autonomous tools give a screening result without a doctor looking at the image first. Assistive tools help a clinician by highlighting problems or estimating risk, but a person makes the decision.

LumineticsCore (formerly IDx-DR), Digital Diagnostics

In April 2018 the US Food and Drug Administration (FDA) authorised IDx-DR as the first medical device that uses AI to detect diabetic eye disease without a specialist interpreting the image. It is used in primary care with a desktop retinal camera and checks for more-than-mild diabetic retinopathy in adults with diabetes. The FDA said it should not be used in people who have already had eye laser treatment, eye surgery or injections, those with persistent vision loss or certain known eye conditions, or pregnant women with diabetes. It is listed among the CE-certified diabetic retinopathy AI systems in the EU.

EyeArt, Eyenuk

EyeArt is an autonomous system for diabetic retinopathy, cleared by the FDA and launched in the US in 2020. In June 2023 an FDA clearance extended it to a third camera model. In January 2023 Eyenuk announced certification under the EU Medical Device Regulation as a Class IIb device covering diabetic retinopathy, AMD and glaucoma-related optic nerve damage. In the UK, NICE (the body that advises the NHS on value for money) has reviewed evidence on EyeArt and similar grading software.

AEYE-DS, AEYE Health

AEYE-DS is an autonomous diabetic retinopathy system. It was first cleared in the US with a desktop camera. In April 2024 the company announced FDA clearance for use with the Optomed Aurora handheld camera, which it described as the first fully autonomous AI clearance for portable diabetic retinopathy screening.

Other CE-certified diabetic retinopathy systems

A 2026 review in the journal Ophthalmology and Therapy counted 13 AI systems for diabetic retinopathy within its scope of CE-certified products in the EU. They include IDx-DR, EyeArt, RetCAD, SELENA+, Retmarker DR, OphthAI and Remidio's Medios AI, which works with a smartphone-based camera. The review found the quality and type of evidence differs a lot between products.

CLAiR, Toku (heart risk from the eye)

CLAiR is an oculomics tool. It analyses a standard retinal photo to estimate a person's risk of cardiovascular disease (heart attack and stroke) in adults with no known heart disease. In February 2024 Toku announced CE and UKCA marks. In the US it has FDA Breakthrough Device designation, which means faster review, but it is not FDA cleared. A US pivotal study (NCT06808334) started in February 2025 and is recruiting.

New treatments that early detection can lead to

Finding disease early only helps if there is something to do about it. For wet AMD and diabetic macular swelling, the standard treatment is an injection into the eye of a drug that blocks a growth signal called VEGF. A higher-dose form of aflibercept (8 mg, sold as Eylea HD in the US) was approved by the FDA in August 2023 for wet AMD, diabetic macular edema and diabetic retinopathy, and by the European Commission in January 2024 for wet AMD and diabetic macular edema. After three monthly starting doses, some patients can space injections out to every 8 to 16 weeks, which means fewer clinic visits.

For geographic atrophy (late dry AMD), the picture is split by region. The FDA approved pegcetacoplan (Syfovre) in February 2023 and avacincaptad pegol (Izervay) in August 2023. Both slow the growth of the damaged areas. However, the European Medicines Agency (EMA) refused Syfovre, confirming its decision in September 2024, because slowing the lesions did not clearly help people's everyday sight while the repeated eye injections carried risks such as inflammation and conversion to wet AMD. Astellas withdrew its EU application for Izervay in 2024. The UK regulator, the MHRA, also rejected Syfovre in January 2025. Japan gave Izervay a conditional approval in September 2025.

What the evidence shows for AI eye screening

Two measures come up again and again. Sensitivity is how many people who really have the disease the test catches. Specificity is how many people without the disease the test correctly clears.

A few points stand out. For diabetic eye disease, the best-studied systems are accurate enough to be approved for screening, and the newest EyeArt review pooled real-world data from many clinics, not just one trial. But high sensitivity often comes with lower specificity. That means some people without serious disease will be sent for a specialist check they did not need.

The 2026 EU review warned that information about where training images came from, which cameras were used and how unreadable images were handled is often incomplete. It also found that evidence for detecting the most sight-threatening stages is reported less often and varies more.

For glaucoma the evidence is weaker. A small 2025 study in Frontiers in Ophthalmology found an AI model was more accurate overall than first-year eye doctors in training at spotting likely glaucoma, but it missed many cases on its own. Combining the AI with the doctor's own judgement worked better than either alone.

In oculomics, the RETFound results for heart failure (a score of 0.79 on a scale where 0.5 is a coin toss and 1.0 is perfect), stroke (0.75) and Parkinson's (0.67) show real but limited signal. A 2023 study from Singapore in the Journal of the American Medical Informatics Association found that retinal photos could help detect diabetic kidney disease, and worked better when combined with blood pressure and blood sugar data. These are associations that help sort people by risk. They do not prove that a particular person has, or will get, the disease.

Who can benefit from AI eye screening

Approved AI screening is mainly for:

• Adults with type 1 or type 2 diabetes who need routine eye screening and have no known diabetic eye disease.
• People who find it hard to reach an eye clinic, because the test can be done in a family doctor's practice, pharmacy-style setting or community clinic.

It is usually NOT for:

• People who already have diagnosed diabetic eye disease, macular swelling, or who have had eye laser treatment, surgery or injections. They need specialist follow-up instead.
• Pregnant women with diabetes, because eye disease can progress quickly in pregnancy and they need closer monitoring.
• Anyone with new symptoms such as sudden blurred vision, flashes of light, a shower of floaters or a dark curtain over the vision. These need urgent in-person care, not a screening photo.
• Anyone wanting a full eye health check. Most approved tools look for one disease only.

Heart-risk tools such as CLAiR are aimed at adults without known heart disease. They are an extra source of information, not a replacement for blood pressure, cholesterol and other standard checks.

What to expect as a patient

1. Invitation or referral. You may be invited by a national screening programme, or your family doctor, diabetes nurse or optician may offer the test during a routine visit.
2. Before the photo. Bring your glasses and a list of medicines. Tell the staff about any eye surgery, injections or recent vision changes. If you are pregnant, say so.
3. The photo. You sit in front of a camera, rest your chin and forehead on a support and look at a target light. There is a bright flash or a scan. Nothing touches your eye. It usually takes a few minutes for both eyes.
4. Eye drops, if needed. Many AI systems are designed to work without drops. If your pupils are small or you have a cataract (a cloudy lens), the picture may not be clear enough. You may then get drops to widen your pupils. These can blur near vision and make you sensitive to light for some hours, so do not drive until your vision is back to normal.
5. The result. With autonomous systems, the result can be ready within minutes during the same visit. It will say "refer" or "no referable disease found", or that the image could not be graded.
6. If you are referred. An optometrist or ophthalmologist (eye doctor) will examine you more closely, often with OCT, to confirm the result and decide whether you need watching, laser or injections.
7. Follow-up. Even with a clear result, keep having regular screening. In England, people whose last two screens showed no diabetic retinopathy have been invited every two years since October 2023; others are seen yearly or more often.

AI eye screening: access and cost in the US, UK and EU

United States

Autonomous AI screening for diabetic eye disease has its own billing code, CPT 92229. Medicare set a national payment for it from 2022. A 2025 analysis reported a payment of about $40 per test and found that the code was used in only 2.2% of more than 154,000 diabetic eye imaging visits from 2021 to 2023, partly because of the cost of cameras and IT set-up. Your share of the cost depends on your insurance plan and deductible. There is no official data on a typical out-of-pocket price, so ask your clinic and insurer before the test. The new geographic atrophy drugs Syfovre and Izervay are available in the US.

United Kingdom

The NHS runs diabetic eye screening programmes that are free at the point of use. Most screening is still graded by trained people. Scotland has used automated grading software (iGradingM) to clear clear-cut normal images in its national programme, and in 2024 researchers reported how a newer deep learning system performed on Scottish data. In England, NICE has reviewed AI grading tools including EyeArt, noting that the evidence comes from observational studies. Heart-risk eye scans such as CLAiR are offered privately and are not part of NHS care. Syfovre was rejected by the MHRA in January 2025.

European Union

Many AI screening products carry a CE mark, which means they meet EU safety and performance rules. A CE mark does not mean your public insurer pays for it. Each country decides on funding, so availability varies by country and region. Ask your diabetes clinic or eye doctor whether AI screening is used locally. The geographic atrophy drugs Syfovre and Izervay are not authorised in the EU, while higher-dose aflibercept is.

Elsewhere and for international patients

Japan granted Izervay a conditional approval in 2025. AI eye screening is also used in other countries, but the details differ and we have not listed them where we could not confirm official data. If you are thinking of travelling for an AI eye test or a new treatment, check these points:

• Approval: Is the device or drug approved by the regulator in the country where you will be treated? Ask for the product name and approval details.
• Experience: How many patients has the centre screened or treated with this tool? Is an eye specialist on site to review positive results?
• Records to bring: Previous eye reports, OCT scans and photos, your diabetes records (such as recent HbA1c, a measure of average blood sugar), blood pressure readings and a list of medicines.
• Follow-up at home: Injection treatments for wet AMD or geographic atrophy need repeat doses over months or years. Make sure a doctor at home can continue care and that the drug is available there.
• Insurance: Check what your insurer covers abroad and whether complications or follow-up visits are included.

Risks, limits and realistic expectations

• False alarms. When a disease is uncommon in the group being screened, even a good test produces many false positives. For example, imagine 1,000 people of whom 20 have referable disease, screened with a test that catches 95% and correctly clears 81%. About 19 true cases would be flagged, but so would about 186 people without the disease. Most "refer" results would then turn out to be false alarms. This can cause worry and longer waits at eye clinics.
• Missed disease. No test catches everyone. Keep going for screening and report new symptoms straight away.
• Unreadable images. Small pupils, cataract, dry eyes and camera problems can stop the AI from giving a result. You may need eye drops or a specialist visit.
• Only one disease checked. A diabetic retinopathy AI does not look for retinal tears, eye tumours or other conditions, and it does not measure eye pressure for glaucoma. A clear result is not a clean bill of health for the whole eye.
• Fairness across populations. AI can perform less well in groups that were under-represented in its training images. Retinal colour varies with pigmentation, and the EU review found that information on the diversity of training data is often incomplete. Ask whether the tool has been tested in people like you.
• "Black box" results. Oculomics models do not explain why they give a risk score. Heart, kidney and brain specialists may want standard tests before acting on them.
• Over-reliance. Doctors and patients may trust a computer result too much. AI is a tool that supports clinical judgement, not a replacement for it.

Clinical trials and what is next for eye disease

• CLAiR pivotal study (NCT06808334), Toku Eyes: tests whether a retinal photo can identify adults at raised cardiovascular risk compared with standard risk calculators. Started February 2025, recruiting, planned size about 942 people.
• Digital Diagnostics pivotal trial (NCT07844317): a new trial of an automated AI system for detecting diabetic retinopathy. Started August 2026, recruiting, planned size about 800 people.
• Retinal hyperspectral imaging in neurodegenerative diseases (NCT07545473), Centre for Eye Research Australia: uses a camera that captures many colours of light to look for early signs of Alzheimer's, Parkinson's and other dementias. Recruiting, planned size about 930 people.
• GlaukomAI (NCT07668193), Fondazione G.B. Bietti in Italy: a clinical validation of an AI system for early glaucoma screening, planned for about 1,200 people, not yet recruiting as of October 2026.
• Global RETFound: the UCL-led consortium training a foundation model on more than 100 million images from over 65 countries, aiming to reduce bias and widen the range of diseases that can be detected.

The key open question for oculomics is whether eye-based risk scores lead to better heart, kidney or brain outcomes. Until larger studies answer that, treat it as promising research.

Questions to ask your doctor about AI eye screening

Take this checklist to your next diabetes, eye or family doctor appointment:

• How often should I have my eyes screened, and when is my next check due?
• Is the AI tool used here approved by the regulator in this country, and for which disease?
• Does the tool check only for diabetic eye damage, or also for glaucoma or macular degeneration?
• Has it been tested in people of my age and ethnic background?
• What happens if the result is "refer" or "image could not be graded", and how long will I wait to see a specialist?
• Will an eye doctor look at my images as well, or only the software?
• Will I need eye drops, and can I drive home afterwards?
• Is the test covered by my health system or insurance, and what will I pay?
• If I have wet AMD or diabetic macular swelling, am I a candidate for longer-acting injections?
• If I have geographic atrophy, which treatments are available where I live, and are there trials I could join?

Questions patients ask