AI Diabetic Retinopathy Screening: How Deep Learning Grades DR from a Single Fundus Image

Diabetic retinopathy screening is the routine examination of the retina in people with diabetes to detect vascular damage before it threatens sight. The need in India is large. The ICMR-INDIAB study, published in The Lancet Diabetes and Endocrinology in 2023, estimated that 101 million Indians live with diabetes, and population studies place the prevalence of diabetic retinopathy at roughly 12 to 17 percent of that group. Guidelines recommend a retinal examination at least once a year for most people with diabetes, yet the number of ophthalmologists is far smaller than the number of eyes that need examining.

Deep learning changes the arithmetic. A trained algorithm can read a retinal photograph in seconds, assign a severity grade and flag which patients need to see an ophthalmologist. This article explains how AI diabetic retinopathy detection works from a single image, what the published evidence shows, and where the limits lie.

Why a single fundus image is enough to start

A fundus image is a colour photograph of the back of the eye that shows the optic disc, the macula and the retinal blood vessels. The lesions that define diabetic retinopathy are all visible on this photograph. They include microaneurysms, dot and blot haemorrhages, hard exudates, cotton wool spots, venous beading and new vessels. Fundus image analysis, whether by a human grader or an algorithm, is the search for these features and an assessment of their extent.

A non-mydriatic fundus camera captures a 45-degree field of the retina through an undilated pupil. The camera aligns on the eye using infrared light, which does not constrict the pupil, and then fires a brief visible flash to capture the image. Several current devices can image through a pupil of about 2 mm. Because no dilating drops are needed, there is no waiting period and no blurred vision afterwards, so screening can take place in a diabetes clinic, a primary health centre or a community camp rather than only in an eye hospital. Dilation is still used when the undilated image is too dark or too small to grade.

The ICDR severity scale and the referral line

Grading follows the International Clinical Diabetic Retinopathy (ICDR) severity scale, a consensus classification published by Wilkinson and colleagues in Ophthalmology in 2003. It simplifies the 13-level ETDRS research scale into five clinical stages:

• No apparent retinopathy. No abnormalities.
• Mild non-proliferative DR (NPDR). Microaneurysms only.
• Moderate NPDR. More than microaneurysms, but less than severe NPDR.
• Severe NPDR. Any of the following with no proliferative signs: more than 20 intraretinal haemorrhages in each of four quadrants, definite venous beading in two or more quadrants, or prominent intraretinal microvascular abnormalities in one or more quadrants. This is the 4-2-1 rule.
• Proliferative DR: neovascularisation, or vitreous or pre-retinal haemorrhage.

Diabetic macular oedema (DME) is graded separately as apparently present or apparently absent, because it can occur at any level of retinopathy.

Referable diabetic retinopathy is defined in the main validation studies as moderate NPDR or worse, referable DME, or both. This is the line that matters for screening. Patients below it can return for routine screening at the next interval. Patients at or above it need an ophthalmologist. Most AI systems are trained and validated against this threshold rather than asked to assign every one of the five grades with equal confidence.

How deep learning grades the image

The models used for AI diabetic retinopathy detection are convolutional neural networks. They are not programmed with rules about lesions. Instead, the network is trained on a large set of fundus photographs that ophthalmologists have already graded on the ICDR scale, and it learns which pixel patterns predict each grade. In the 2016 JAMA study by Gulshan and colleagues, the training set was 128,175 images and the labels came from a panel of 54 US-licensed ophthalmologists and senior trainees.

In use, the process runs in four steps.

1. Image quality check. The system first decides whether the photograph is gradable. Poor focus, uneven illumination, or an off-centre field returns an ungradable result, and the operator recaptures the image or dilates the pupil.
2. Feature extraction. Successive layers of the network detect progressively more complex patterns, from edges and colour contrasts to the small red dots of microaneurysms, the bright patches of exudates and abnormal vessel shapes. Some systems, such as the one behind IDx-DR, use explicit detectors for these lesion types. Others learn end-to-end from the whole image.
3. Classification. The final layer outputs a probability of referable disease, or a probability for each severity grade.
4. Operating point. A threshold on that probability decides the referral. It can be set for high sensitivity, which misses fewer cases but refers more people unnecessarily, or for high specificity, which does the opposite. Screening programmes usually choose the high-sensitivity setting.

The whole sequence takes seconds and produces a result on the spot, without a specialist present.

What the evidence shows

Three studies anchor the field.

Gulshan et al., JAMA, 2016. The algorithm was tested on two independent sets, EyePACS-1 (9,963 images) and Messidor-2 (1,748 images). At the high-sensitivity operating point, it reached 97.5 percent sensitivity and 93.4 percent specificity on EyePACS-1, and 96.1 percent sensitivity and 93.9 percent specificity on Messidor-2, for referable diabetic retinopathy. The reference standard was the majority decision of a panel of ophthalmologists.

Abràmoff et al, npj Digital Medicine, 2018. This was the pivotal trial behind IDx-DR, the first autonomous AI diagnostic system authorised by the US FDA, in April 2018. It enrolled 900 adults with diabetes and no prior retinopathy diagnosis at 10 primary care sites. Against a reading-centre standard, the system correctly identified 173 of 198 participants with more than mild diabetic retinopathy, a sensitivity of 87 percent. The images were taken by primary care staff, not eye specialists.

Gulshan et al., JAMA Ophthalmology, 2019. A prospective study of 3,049 patients with diabetes at Aravind Eye Hospital and Sankara Nethralaya in South India. The automated system reached 88.9 percent sensitivity and 92.2 percent specificity at Aravind, and 92.1 percent sensitivity and 95.2 percent specificity at Sankara Nethralaya. Manual grading by individual graders at the same sites ranged from 73.4 to 89.8 percent sensitivity. The authors concluded that the model generalised to an Indian population.

On the strength of this evidence, the American Diabetes Association has included FDA-authorised AI systems that detect more than mild retinopathy as an alternative to traditional screening in its Standards of Care since 2020, provided that referral pathways are in place.

Where AI screening fits and what it cannot do

AI screening addresses one step in the care pathway, which is image interpretation at the point of first contact. It moves that step into the diabetes clinic or primary care setting. It does not replace the ophthalmologist. A patient flagged as referable still needs a dilated examination, and often optical coherence tomography to confirm macular oedema, before any treatment decision is made.

Four limits are worth stating plainly.

• Ungradable images. Cataract, small pupils and media opacity produce images the algorithm cannot read. These patients need dilation or a conventional examination, and a screening programme has to plan for them.
• Validated threshold. Most systems are validated for the referable versus non-referable decision. Fine distinctions between individual grades, and the detection of DME from a colour photograph alone, are less certain.
• Population and camera. Performance figures come from specific cameras, image protocols and populations. A system should be validated on the camera and in the population where it will be deployed.
• Follow-up. Screening only prevents vision loss if referred patients actually reach an ophthalmologist. Without a working referral pathway, AI screening increases test volume without improving outcomes.

The algorithm also looks only for diabetic retinopathy unless it has been trained for other conditions. Glaucoma, age-related macular degeneration and hypertensive changes may go unreported.

Conclusion

Diabetic retinopathy screening works because the disease announces itself on a photograph long before it affects vision. A non-mydriatic fundus camera makes that photograph easy to obtain in any clinic, the ICDR severity scale gives a shared language for grading it, and deep learning now reads it to the referable threshold with sensitivity and specificity comparable to trained human graders. The evidence from India specifically, where the model performed as well as or better than manual grading at two large eye hospitals, supports using this approach to reach the patients who are currently not screened at all.

Forus Health AI builds on this foundation with FH-POISE (Precision Ocular Intelligence for Systemic and Eye Health), its retinal image analysis platform. In February 2026, Forus Health received approval from the Central Drugs Standard Control Organisation (CDSCO) for FH-POISE DRT, a deep learning tool that assesses retinal images and predicts diabetic retinopathy severity. It is designed for use at the point of care in general hospitals, diabetes clinics, community health centres and outreach programmes, and it complements the 3nethra range of non-mydriatic fundus cameras that Forus Health manufactures. The aim is the one described throughout this article, which is to embed reliable screening into routine diabetes care so that referable disease is found early and specialist time is spent where it is needed.

Sources

1. Gulshan V, Peng L, Coram M, et al. Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. JAMA. 2016;316(22):2402-2410. https://doi.org/10.1001/jama.2016.17216
2. Abràmoff MD, Lavin PT, Birch M, Shah N, Folk JC. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. npj Digital Medicine. 2018;1:39. https://doi.org/10.1038/s41746-018-0040-6
3. Gulshan V, Rajan RP, Widner K, et al. Performance of a Deep-Learning Algorithm vs Manual Grading for Detecting Diabetic Retinopathy in India. JAMA Ophthalmology. 2019;137(9):987-993. https://jamanetwork.com/journals/jamaophthalmology/fullarticle/2734990
4. Wilkinson CP, Ferris FL, Klein RE, et al. Proposed international clinical diabetic retinopathy and diabetic macular edema disease severity scales. Ophthalmology. 2003;110(9):1677-1682. https://pubmed.ncbi.nlm.nih.gov/13129861/
5. Anjana RM, Unnikrishnan R, Deepa M, et al. Metabolic non-communicable disease health report of India: the ICMR-INDIAB national cross-sectional study (ICMR-INDIAB-17). The Lancet Diabetes and Endocrinology. 2023;11(7):474-489. Summary: https://nmji.in/?p=6485
6. American Diabetes Association. Retinopathy, Neuropathy, and Foot Care: Standards of Care in Diabetes. Diabetes Care. 2025;48(Suppl 1):S252-S265. https://diabetesjournals.org/care
7. University of Iowa Carver College of Medicine. FDA’s first clearance of AI for autonomous detection of diabetic retinopathy. 2018. https://medicine.uiowa.edu/node/1025
8. BioSpectrum India. Forus Health secures CDSCO approval for AI-led Diabetic Retinopathy screening tool. 4 February 2026. https://biospectrumindia.com/news/16/27279/forus-health-secures-cdsco-approval-for-ai-led-diabetic-retinopathy-screening-tool-.html