Author: Claude, from an owner-approved source list · Reviewed by: Dr Mina Abdelmseih · Reviewed: 2026-11-15 · Review due: 2027-11-15

Educational content only. Not personalised medical advice. Clinical decisions remain the responsibility of the treating clinician.

Artificial Intelligence in Retina and OCT

Deep-learning systems can grade retinal photographs and OCT scans for diabetic retinopathy (DR), diabetic macular oedema (DMO) and AMD with high accuracy in studies [S1] [S10] [S12]. Diabetic retinopathy screening is the most mature use, with prospective and real-world evidence and regulator-cleared autonomous systems [S1] [S3] [S4]. Most other uses are earlier-stage [S13].

Resource tiers: this is a specialist research topic outside the Tier A / Tier B / Tier C low-resource focus of this library, so no resource tiers are defined here. AI-supported DR screening is the part most relevant to low-resource settings [S6] [S7].

Red flags

• An algorithm validated elsewhere may fail locally. Most DR algorithms were developed and validated in high-income countries, none was developed or implemented in Africa, and local validation is needed [S7].
• Internal results overstate performance. Models performed better on internal than external validation [S10].
• Ungradable images. Accuracy figures apply to gradable images [S2]; AI gradability was at least 95% in only 60% of cohorts [S1].
• Specificity varies by product. Pooled specificity was 81% for EyeArt [S4] and 91% for IDx-DR [S3], so referral volumes differ.
• "Oculomics" claims. 30 of 31 studies predicting systemic disease from fundus photographs had a high risk of bias [S13].

Recognise it

Where AI is used:

Confirm it

Questions to ask before using a system:

• Has it been validated prospectively, on your type of camera and your population [S1] [S7]? Camera type and prevalence explained differences in accuracy [S1].
• Does it have regulatory approval? Only 12 of more than 400 DR algorithms had regulatory approval [S7].
• What happens to ungradable images and uncertain results [S1] [S11]?
• Who checks performance over time? Quality assurance and ongoing monitoring are needed [S1].

Manage it

Before acting on this section, check the national and local guidelines, product information and regulatory approvals that apply where you practise. See the medical disclaimer.

Specialist and programme use

1. Autonomous DR screening. Pooled sensitivity of the FDA-authorised systems was 95% for IDx-DR (13 studies) [S3] and 95% for EyeArt (17 studies, 162,695 examinations) [S4].
2. Improving attendance. In the ACCESS RCT (2024) in youth with diabetes, point-of-care autonomous AI raised eye exam completion to 100% vs 22% with referral [S8]. A meta-analysis (2025) found higher uptake of follow-up appointments after AI assessment (OR 1.89) [S9].
3. OCT as a second step to cut false referrals. In a 2026 RCT in Hong Kong, adding AI-OCT to photograph-based screening reduced false-positive DMO referrals from 69.1% to 24.1%, with no DMO missed among non-referred participants [S11].
4. Low-resource deployment. Portable fundus cameras and trained general physicians showed good agreement with reference standards in developing countries [S6]. Success depends on infrastructure, local validation and regulatory support [S7].
5. Patient data. Do not upload identifiable patient images or data to unapproved tools; follow your local data-protection rules.

Refer when

Pitfalls

• Applying published accuracy to a different camera or population [S1] [S7].
• Ignoring ungradable images [S1] [S2].
• Trusting internally validated models [S10].
• Treating oculomics predictions as diagnostic [S13].
• Vendor involvement: EyeArt specificity varied with vendor involvement [S4].

Evidence notes

• Diagnostic-accuracy meta-analyses show high heterogeneity [S12].
• Not covered by these sources: generative AI and large language models, cost-effectiveness and medicolegal responsibility.

Related articles

• OCT interpretation: macula and retinal nerve fibre layer
• Diabetic retinopathy screening programmes
• Smartphone and portable fundus imaging

References

See the numbered source list below. Each marker [S#] links to its PubMed and DOI record.