AI in Ophthalmology: Can Artificial Intelligence Detect Diabetic Eye Disease?

AI in Ophthalmology: Can Artificial Intelligence Detect Diabetic Eye Disease? | Helathora

Last updated: October 21, 2026

Editorial note: This article is based on peer-reviewed clinical studies, FDA records and published research on AI screening for diabetic retinopathy. Helathora has not tested any of the systems described and is not affiliated with any manufacturer. Accuracy figures come from specific studies with specific cameras and patient groups; they are estimates, not guarantees of how a system will perform in every clinic.

2018Year the FDA authorized the first autonomous AI system to detect diabetic retinopathy

87.2%Sensitivity of LumineticsCore (then IDx-DR) in its 900-patient pivotal trial

95.5%Sensitivity of EyeArt in its peer-reviewed pivotal study, without pupil dilation

>20%Share of images rejected as low quality in Google’s 2020 field study in Thai clinics

Figures come from four different studies and are not directly comparable. Sources are listed at the end of the article.

Yes: artificial intelligence can detect diabetic eye disease from photographs of the retina, and in the United States several systems are cleared to do it without a doctor reviewing the images first. That makes diabetic retinopathy one of the few areas of medicine where AI is already used as an autonomous diagnostic test, not just a second opinion.

The more useful question is how well it works, for whom and where. This article explains what these systems actually detect, what the trial numbers mean, why real-world results can differ, and what a patient should do with a positive, negative or ungradable result.

Quick Answer

AI screening tools can reliably flag diabetic retinopathy that needs a specialist’s attention. In their pivotal studies they detected roughly 87% to 96% of cases. They are screening tests, not a full eye exam: they look for one disease, they work only with approved cameras and patient groups, and a negative result does not rule out other eye problems. Their biggest value is reaching people with diabetes who would otherwise skip their regular eye check.

Why Diabetic Eye Disease Is a Natural Target for AI

Diabetic retinopathy is damage to the small blood vessels of the retina, the light-sensitive layer at the back of the eye, caused by long-term high blood sugar. It can also cause diabetic macular edema, a swelling of the central retina. In the early stages there are usually no symptoms, and by the time vision changes the disease may already be advanced. That is why people with diabetes are advised to have regular retinal screening.

The problem is not a lack of effective treatment; it is that many people never get screened. AEYE Health, one of the companies in this field, cites estimates that only 15% to 50% of people with diabetes in the U.S. follow screening guidelines. Needing a separate appointment with an eye specialist is one of the main barriers.

Retinal screening also suits AI unusually well. The test is based on a photograph, the signs of disease (tiny bleeds, fluid leaks, abnormal vessels) are visible patterns, and large collections of graded images exist to train and test algorithms. Placing an AI system in a family doctor’s office removes the extra appointment.

How AI Screening Works in Practice

The process is short and, in most systems, takes place during a routine visit:

Retinal photographs: a trained staff member, not necessarily an eye specialist, takes one or more photos of each eye with a fundus camera. Some systems work without dilating the pupils.
Image quality check: the software first decides whether the images are good enough to analyze. If not, it asks for new photos or returns an “insufficient quality” result.
Analysis: the algorithm looks for signs of disease and classifies the result, typically as positive or negative for “more than mild” diabetic retinopathy.
Result and referral: a positive result means the patient should see an eye care professional. A negative result usually means rescreening at the next recommended interval.

Key term: “More than mild diabetic retinopathy” (mtmDR) is the threshold most systems are designed to detect. It is the point at which referral to an eye specialist is recommended. Very early, mild changes are not the main target.

The FDA-Cleared Systems and What Their Studies Showed

Three systems are frequently cited as cleared by the FDA for autonomous diabetic retinopathy screening. Where a peer-reviewed publication exists, the figures below come from it rather than from company announcements.

System	FDA authorization	Performance reported in pivotal trials	Points to note
LumineticsCore (formerly IDx-DR), Digital Diagnostics	April 2018	Sensitivity 87.2% (95% CI 81.8–91.2%), specificity 90.7% (95% CI 88.3–92.7%) and imageability 96.1% for more than mild DR, in 900 patients at 10 U.S. primary care sites	The first autonomous AI diagnostic system authorized by the FDA in any field of medicine. Renamed LumineticsCore in 2023; the indications for use did not change.
EyeArt, Eyenuk	August 2020	More than mild DR without dilation: sensitivity 95.5% (95% CI 92.4–98.5%), specificity 85.0% (95% CI 82.6–87.4%). Vision-threatening DR: sensitivity 95.1%, specificity 89.0%. 893 patients analyzed	Reports vision-threatening disease separately. Specificity rose to 87.8% after statistical correction for the study’s enriched sample.
AEYE-DS, AEYE Health	November 2022 (desktop camera); April 2024 (handheld camera added)	Sensitivity 93% and specificity 91.4% with a desktop camera; 91.9% and 93.6% with a handheld camera, according to the company’s phase 3 results	Uses one image per eye. Confidence intervals were not available in the sources reviewed for this article.

Sensitivity vs specificity: sensitivity is the share of people with the disease that the test correctly flags. Specificity is the share of people without it that the test correctly clears. A screening test favors high sensitivity, because a missed case is more harmful than an unnecessary referral.

About the numbers: each figure is an estimate from a single study. Its precision depends on sample size and study design, which is what the confidence interval (CI) shows: the range within which the true value is likely to lie. None of these results is a guaranteed performance level in routine practice.

Reported sensitivity for more than mild diabetic retinopathy (pivotal trials)

LumineticsCore (2018)

87.2%

AEYE-DS, desktop

93%

EyeArt (2021 paper)

95.5%

Bars show each study’s reported sensitivity on a 0–100% scale. Sources: Abràmoff et al., npj Digital Medicine 2018; Ipp et al., JAMA Network Open 2021; AEYE Health phase 3 results.

Do not read this as a ranking: these studies used different patient populations, cameras, reference standards and protocols (for example, with or without pupil dilation). A higher number in one study does not prove one system is better than another. Only head-to-head studies on the same patients can answer that, and few exist.

What Happens Outside Clinical Trials

Pivotal trials take place under controlled conditions. Three well-known studies show what can change once AI meets everyday clinics. They looked at different questions, so their results should not be mixed.

Seven algorithms on the same real-world images

Researchers at the University of Washington tested seven AI screening algorithms from five companies on more than 311,000 retinal images from nearly 24,000 veterans screened at VA centers in Seattle and Atlanta between 2006 and 2018. The results, published in Diabetes Care in 2021, showed significant differences between algorithms. Several performed worse than the human graders already working in the VA’s screening program, and performance also varied between the two sites.

The authors concluded that every algorithm should be tested rigorously on real-world data before clinical use. The study population is also worth noting: almost all patients were men with type 2 diabetes, which limits how far the findings can be generalized.

Google’s workflow study in Thai clinics

In a study presented at the CHI conference in 2020, Google Health researchers observed how a deep learning system worked in daily practice across 11 clinics in Thailand. The system had shown high accuracy on research datasets, but in the clinics more than a fifth of images were rejected as too low quality to grade, often because of lighting in the screening rooms. Slow internet connections also delayed uploads, and nurses had to manage patients whose images were rejected.

A large prospective study in Thailand’s screening program

A separate prospective study, published in The Lancet Digital Health in 2022, evaluated a deep learning system within Thailand’s national screening program in 7,651 patients. For vision-threatening diabetic retinopathy it reported a sensitivity of 91.3% and a specificity of 96.3%. This is a different evaluation from the workflow study above: one measured accuracy in a large patient group, the other described how the tool fitted into clinic routines.

Together, these studies illustrate a lesson that applies across medical AI: an accurate algorithm is only one part of a working screening service. Cameras, lighting, staff training, connectivity and referral pathways matter just as much.

What Your Result Means

Result	What it means	What to do
Positive	Signs of diabetic retinopathy above the referral threshold were detected. It is not a diagnosis of severity.	Book the referral with an eye care professional promptly. The benefit of screening depends on this follow-up.
Negative	No referable diabetic retinopathy was detected in these images.	Keep to your recommended screening interval. It does not rule out other eye conditions.
Ungradable / insufficient quality	The images could not be analyzed. This is not a negative result.	Ask how and when you will be examined instead, for example with new photos, dilation or an eye clinic appointment.

What AI Screening Does Not Do

It screens for one disease. These systems are designed to detect diabetic retinopathy. A negative result does not rule out glaucoma, cataract, age-related macular degeneration or other conditions a full eye exam can find.
It has a defined intended use. FDA clearances typically cover adults with diabetes who have not already been diagnosed with more than mild diabetic retinopathy. People already under follow-up for retinopathy should continue with their eye specialist.
It depends on the approved equipment. Each clearance is tied to specific cameras. Results with other devices are not covered by the same evidence.
It is not a treatment decision. A positive result means “see a specialist”, not a decision about what treatment is needed.
It does not replace symptoms. Sudden vision changes, new floaters, flashes or loss of vision need prompt medical attention whatever a recent screening showed.

Is AI Diabetic Eye Screening Available in Europe?

FDA clearance in the United States does not by itself establish that a system is authorized or used in the same way in Europe. Availability depends on each device’s European regulatory status and intended use, and on how national and regional health services organize screening. EyeArt, for example, reports CE marking as a class IIa medical device in the European Union.

For patients in Spain and other European countries, AI should be seen as a possible way to support diabetic retinopathy screening, not as a replacement for the eye examinations and follow-up recommended by their healthcare team. The practical step is to ask your diabetes care team or local ophthalmology service which screening pathway is used where you live.

Lesson for Patients: Using AI Screening Well

If you are offered AI retinal screening, these checks help you get real value from it:

Ask what the test looks for: diabetic retinopathy only, or other conditions too.
If the result is positive, make sure the referral appointment is booked.
If the images were of insufficient quality, ask what happens next rather than assuming all is well.
If the result is negative, ask when you should be screened again.
Tell the clinician if you already have a retinopathy diagnosis, are pregnant, or have other eye conditions, since these may change whether AI screening is appropriate.
Report any change in vision immediately, even soon after a normal result.

Related reading

Limitations of This Article

Different evidence types: this article combines pivotal trials, real-world studies and company-reported results. AEYE-DS figures come from company announcements; the LumineticsCore and EyeArt figures come from peer-reviewed publications.
Separate study designs: sensitivity and specificity from different studies cannot be compared directly.
Focus on U.S. regulation: availability, approvals and screening programs differ by country.
A moving field: software versions, cleared cameras and indications change over time. Check current information with the manufacturer or regulator.

The Bottom Line

AI can detect diabetic eye disease, and it is one of the best-validated uses of autonomous AI in medicine. Its real promise is access: bringing a reliable retinal check to the family doctor’s office, where many people with diabetes already go. Its limits are just as real. It screens for one condition, depends on good images and approved equipment, and only helps if a positive or ungradable result leads to a proper eye examination.

Frequently Asked Questions

How does AI screening compare with an eye exam by an ophthalmologist?

For one specific task, detecting more than mild diabetic retinopathy in retinal photographs, FDA-cleared systems showed high sensitivity and specificity in their pivotal studies. An ophthalmologist’s examination covers much more: other diseases, the severity of retinopathy and treatment decisions. Real-world studies also show that performance varies between algorithms and settings.

Do I need eye drops for an AI screening?

Not always. Several systems are designed to work without dilating the pupils, although dilation may be needed if the images are not clear enough. In the EyeArt pivotal study, for example, dilating eyes with ungradable images raised the share of gradable eyes from 87.4% to 97.4%.

Can AI screening detect glaucoma or macular degeneration?

The FDA-cleared autonomous systems described here are indicated for diabetic retinopathy. AI research is exploring other eye diseases, but a negative diabetic retinopathy screening should not be taken as a check for other conditions.

What should I do if my images could not be analyzed?

Treat it as an incomplete test, not as a normal result. Ask the clinic whether new photos, pupil dilation or a referral to an eye service is the next step.

Sources

Abràmoff MD et al. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. npj Digital Medicine, 2018. PubMed Central
Ipp E et al. Pivotal evaluation of an artificial intelligence system for autonomous detection of referrable and vision-threatening diabetic retinopathy. JAMA Network Open, 2021. PubMed
FDA 510(k) database: AEYE-DS, K221183 (November 2022). accessdata.fda.gov
FDA 510(k) summary: AEYE-DS, K240058 (2024, Topcon NW400 and Optomed Aurora cameras). accessdata.fda.gov
AEYE Health clinical results as reported at clearance. medtechdive.com
Digital Diagnostics: rebrand of IDx-DR to LumineticsCore, March 2023. digitaldiagnostics.com
Lee AY et al. Multicenter, head-to-head, real-world validation study of seven automated artificial intelligence diabetic retinopathy screening systems. Diabetes Care, 2021. PubMed Central
Beede E et al. A human-centered evaluation of a deep learning system deployed in clinics for the detection of diabetic retinopathy. CHI 2020. Coverage: MIT Technology Review
Prospective deep learning screening study in Thailand’s national program, The Lancet Digital Health, 2022, as summarized in a review of prospective validation studies. PubMed Central

Medical disclaimer: This article is for informational and educational purposes only and is not medical advice. It does not replace an examination by a qualified eye care professional. If you have diabetes, follow the screening schedule recommended by your healthcare team, and seek prompt care for any change in your vision. Helathora is not affiliated with any company or product mentioned. Last reviewed: October 21, 2026.