AI in Ophthalmology: Transforming Eye Disease Screening and Diagnosis
Artificial intelligence is making significant strides in ophthalmology, particularly in screening for retinal diseases. The most mature application is in diabetic retinopathy (DR) detection, where AI-driven systems analyze retinal photographs to catch early signs of disease. In fact, the U.S. FDA has cleared three autonomous AI devices for DR screening: Digital Diagnostics’ IDx-DR (now called LumineticsCore), Eyenuk’s EyeArt, and AEYE Health’s AEYE-DS retina-specialist.comretina-specialist.com. These algorithms examine fundus images and can diagnose referable diabetic retinopathy with high accuracy (sensitivities around 87–96% and specificities ~88–91% in trials) modernretina.comretina-specialist.com. Notably, AEYE-DS, approved in 2023, is the first fully autonomous handheld AI system for DR – its portability allows primary care clinics and remote programs to detect retinopathy without an on-site specialistglaucomatoday.comglaucomatoday.com. All three FDA-cleared platforms can analyze images from standard non-mydriatic fundus cameras and instantly indicate if the patient has more-than-mild DR that warrants referralretina-specialist.comretina-specialist.com. This capability is transformative for early detection, since timely laser or pharmacologic treatment of diabetic eye disease can prevent blindness.
Despite regulatory approvals, adoption of AI in eye clinics has been gradual. A 2024 study of U.S. clinics found that only 0.09% of diabetic patients received AI-based DR screening (CPT code 92229) from 2021–2023 modernretina.commodernretina.com. In other words, fewer than 1 in 1,000 patients with diabetes were screened using AI, highlighting that the technology’s use is still nascent. Medicare billing data mirror this low uptake – just over 15,000 AI-based DR screening claims have been filed since 2022, primarily in urban academic centers retina-specialist.com. Nonetheless, there are signs of growth. IDx-DR reports deployment in over 1,000 clinics (though independent verification is pending) retina-specialist.com. A quality improvement program in 5 U.S. health systems deployed 198 AI-equipped fundus cameras by mid-2024, enabling screening coverage for ~151,000 diabetic patients pubmed.ncbi.nlm.nih.gov. Over 20,000 DR exams have been completed in that initiative, detecting more than 3,450 cases of vision-threatening retinopathy and facilitating specialist referrals pubmed.ncbi.nlm.nih.gov. These early results underscore AI’s potential to scale screening – if workflow integration and reimbursement hurdles can be overcome. Indeed, financial factors are key: Medicare introduced a reimbursement code (92229) in 2021 that pays about $40 per AI screening, which is higher than traditional teleophthalmology reads, to incentivize adoption retina-specialist.comretina-specialist.com. This higher reimbursement is intended to spur clinics to invest in the technology and help close screening gaps.
AI for Diabetic Retinopathy, Macular Degeneration, and Glaucoma
Beyond diabetic retinopathy, researchers are applying AI to other major eye diseases. In age-related macular degeneration (AMD), deep learning models have been trained on optical coherence tomography (OCT) scans to detect early macular changes and even predict progression from dry to wet AMD. For example, Google’s DeepMind (now Google Health) partnered with Moorfields Eye Hospital to develop an AI that can analyze OCT scans and classify a range of retinal conditions – including AMD – with expert-level accuracy ois.netois.net. While no autonomous AMD diagnostic is FDA-approved yet, clinical decision-support tools are emerging. AI can triage urgent cases, identifying OCT signs of neovascular (“wet”) AMD that need prompt treatment. Additionally, in clinical trials, AI models can help quantify geographic atrophy or drusen burden from images more reproducibly than human graders insight.hdrhub.orgsciencedirect.com, aiding patient selection for new therapies. These advances hint that AI will soon support retina specialists by monitoring AMD patients for conversion and optimizing injection timing.
In glaucoma, AI offers promise in both detection and management. Glaucoma is a complex optic nerve disease often requiring multiple tests – fundus exam, intraocular pressure, visual fields, OCT of nerve fibers – making autonomous diagnosis challenging glaucomatoday.comglaucomatoday.com. There is no FDA-cleared AI for glaucoma yet, but prototypes exist. For instance, researchers have shown that deep learning of fundus photographs can flag optic nerve cupping suggestive of glaucoma, and algorithms can analyze patterns in visual field tests to detect early functional loss. More impressively, AI is proving useful for tracking glaucoma progression. Recent studies demonstrate that machine learning can detect subtle changes in serial OCT and visual field data – sometimes predicting progression risk before clinicians can glaucomatoday.comglaucomatoday.com. One convolutional neural network detected glaucomatous progression in over 14,000 OCT scans with an AUC of 0.97 glaucomatoday.com. Another model predicted which patients would eventually need glaucoma surgery with >90% accuracy by analyzing initial imaging and clinical parameters glaucomatoday.com. These tools could enable “predictive monitoring,” alerting doctors to fast progressors so they can tailor treatment (e.g. earlier surgery for high-risk patients )glaucomatoday.comglaucomatoday.com. While fully autonomous glaucoma screening is distant due to the disease’s multifactorial nature glaucomatoday.comglaucomatoday.com, AI’s role as a decision support – for example, forecasting an individual’s vision loss trajectory – may become standard in glaucoma care.
Expanding Access: AI Eye Screening in Low-Resource Settings
A major appeal of AI in ophthalmology is its ability to expand eye care to underserved populations. Many regions face shortages of ophthalmologists or optometrists, leading to preventable blindness from late diagnosis. AI screening tools can act as force multipliers in such settings. A landmark study in Zambia proved this concept: an AI algorithm for diabetic retinopathy, trained on images from Singapore, was tested on thousands of patients in rural Africa. It achieved >92% sensitivity and ~89% specificity for referable DR, on par with ophthalmologists pubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. Equally important, it showed 99% sensitivity for vision-threatening DR – catching advanced cases most likely to cause blindness pubmed.ncbi.nlm.nih.gov. This 2019 study demonstrated that AI can generalize across continents and still perform well, hinting at solutions for low-resource areas pubmed.ncbi.nlm.nih.govpubmed.ncbi.nlm.nih.gov. Since then, pilot projects have deployed AI-equipped fundus cameras in community clinics and diabetes centers in countries like India, Thailand, and Botswana. Early reports show high accuracy and patient satisfaction, though challenges like funding and internet connectivity remain.
AI is also being leveraged to screen other eye diseases in infants and rural communities. A compelling example is in retinopathy of prematurity (ROP), a leading cause of childhood blindness. In 2023, researchers developed a machine learning model that analyzes smartphone retinal videos of preemies’ eyes, to identify ROP in NICUs lacking specialists. In a study of 512 infants in Mexico and Argentina, the AI had a patient-level sensitivity of 93.3%, outperforming a panel of pediatric ophthalmologists (73.3% sensitivity) at flagging which babies had ROP needing treatment jamanetwork.comjamanetwork.com. This AI’s ability to use cheap smartphone imaging instead of expensive retinal cameras is a game-changer – it can “expand access to ROP screening and optimize the scarce pediatric ophthalmologist workforce” in low-resource settings jamanetwork.com. Similarly, in adult eye care, portable fundus cameras paired with offline AI software are being trialed for tele-ophthalmology. Health workers in remote clinics can take an image of the eye; the AI immediately indicates if the person has diabetic retinopathy, cataract, or even optic disc cupping. Patients with positive findings can then be prioritized for the visiting specialist or referred to tertiary centers, while those without disease are reassured. The triage and task-shifting enabled by AI can dramatically increase screening coverage where doctors are scarce glaucomatoday.comglaucomatoday.com. For instance, the Orbis Flying Eye Hospital program and others are evaluating AI tools to help non-doctor staff identify cataracts and corneal diseases in underserved African and Asian communities pmc.ncbi.nlm.nih.govtandfonline.com. Early outcomes show AI agreeing with ophthalmologists’ diagnoses in the vast majority of cases tandfonline.com.
Overall, AI in ophthalmology (AI eye screening) is moving from research to real-world deployment. By 2024–2025, dozens of healthcare systems across the US, EU, India, and Africa have begun integrating AI retinal screenings as part of routine diabetes care retina-specialist.comretina-specialist.com. Professional bodies like the American Academy of Ophthalmology endorse these innovations cautiously: they celebrate the improved detection rates, but also emphasize ongoing oversight (e.g. ensuring AI algorithms perform equitably across ethnic groups and don’t miss other pathologies on the images) retina-specialist.comretina-specialist.com. If implementation challenges can be addressed – including physician trust, liability questions, and workflow integration – AI has the potential to prevent blindness on a global scale by catching eye diseases early and directing patients to timely care
