AI in Ophthalmology: What Every Resident Should Know

Last updated: October 2026

Ophthalmology was the first specialty to have an autonomous AI diagnostic system cleared by the FDA, and imaging-heavy eye care remains at the front of the AI wave. Residents do not need to code, but they do need to understand what these systems can and cannot do, how to read an AI paper, and where the pitfalls lie.

Terms you must know

• Machine learning / deep learning: algorithms that learn patterns from labelled data. Convolutional neural networks (CNNs) dominate image analysis.
• Assistive vs autonomous AI: assistive AI supports a clinician's decision. Autonomous AI gives a result with no clinician reading the image.
• Sensitivity, specificity and AUC: always ask at which operating threshold, against what reference standard, and in what population.
• Gradability: the proportion of images the system can actually read. Ungradable images must be referred, which changes real-world performance.

FDA-cleared autonomous diabetic retinopathy screening

Elsewhere, Scotland has used automated triage in its national programme since 2011, and Singapore deploys the SELENA+ system nationally. In India, smartphone-based fundus cameras with offline AI have been validated in community screening, which suits camps without reliable internet.

Beyond diabetic retinopathy

• OCT triage: deep learning on OCT can match retinal experts' referral decisions across many macular diseases (DeepMind and Moorfields, Nature Medicine 2018).
• Glaucoma: disc photograph and OCT analysis, and visual field progression forecasting.
• ROP: plus-disease grading from retinal images, valuable where trained screeners are scarce.
• Oculomics: using retinal images to predict systemic markers such as age, cardiovascular risk and kidney function. This is still a research area.
• Large language models (LLMs): helpful for drafting letters, summarising literature and studying, but they can produce confident errors ("hallucinations"). Verify every fact, and never paste identifiable patient data into public tools.

Real-world pitfalls

• Image quality: in a multisite Thai deployment, 8–39% of images were rejected for quality. Lighting, pupil size and connectivity all matter.
• False positives vary widely: a 2025 English evaluation of eight CE-marked systems found very high sensitivity for proliferative disease (95.8–99.5%), but false-positive rates in no-DR encounters ranged from 4.3% to 61.4%.
• Dataset shift and bias: a system trained on one camera, ethnicity or prevalence may underperform elsewhere. Look for external validation in a population like yours.
• Spectrum: primary-care trials include few advanced cases, and AI usually misses non-DR pathology unless it was trained to find it.
• Regulatory clearance does not guarantee readiness for real-world use.

How to read an AI paper: a quick checklist

1. What was the reference standard (expert graders, reading centre, adjudication)?
2. Was validation external and prospective, or only internal and retrospective?
3. What were the disease prevalence and image gradability?
4. Are sensitivity and specificity given at a fixed, pre-specified threshold?
5. Does the reporting follow STARD-AI, CONSORT-AI or SPIRIT-AI?
6. Who is accountable for a missed diagnosis?

Viva pearls

• IDx-DR (2018) was the first FDA-authorised autonomous AI diagnostic in any specialty.
• Autonomous means no human reads the image; ungradable results must still be referred.
• Explainability tools such as heat maps and saliency maps show where a network "looked", but not why.
• AI and thesis topics combine well: validation of a screening tool in your own population is a feasible MS/DNB project.

Related: diabetic retinopathy for residents, how to read an OCT and thesis topic ideas.

Key sources

• Retinal Physician: AI screening for diabetic retinopathy (2026)
• Ophthalmology Times: AEYE Health portable autonomous AI clearance
• Abràmoff MD et al. Pivotal trial of an autonomous AI-based diagnostic system. npj Digital Medicine 2018;1:39
• De Fauw J et al. Clinically applicable deep learning for diagnosis and referral in retinal disease. Nature Medicine 2018;24:1342–50