AI in ophthalmology: Retina specialists weigh the promise and the pitfalls

Artificial intelligence is reshaping how retina specialists design clinical trials, screen patients, and interpret imaging, according to a collection of ophthalmologists and vision researchers who discussed the technology's expanding role—and its limitations—in eye care.

Redefining clinical trial design

Hasenin Al-khersan, MD, Retina Group of Florida, said AI is changing how clinical trials are designed by enabling new endpoints. Companies such as Biogen and, previously, Apellis entered the retina space “very early on, before we even really knew what to measure,” relying on tools such as RPE loss, geographic atrophy (GA), or fundus autofluorescence that “now almost seems archaic.” Trials can now stratify and enroll patients based on photoreceptor loss instead, catching disease progression earlier, since photoreceptor loss predates the GA that develops later—by which point, Hasenin said, treatment intervention already comes too late.

Hasenin also pointed to a gap between what AI can do in research and what it does in the clinic today. It would help, he said, to use these tools to personalize treatment and show patients “This is your trajectory before you were treated, and this is what you're doing now that we started treating you”—a capability that does not yet exist and that Hasenin called “one of the biggest gaps I hear from my colleagues.” The biggest barrier to closing that gap is regulatory, Hasenin said: a device company, a pharmaceutical company, or the field itself needs to invest the time and money required to gain FDA approval for a clinic-ready tool, a shift Hasenin estimated is “probably at least a few years out.”

Arshad M. Khanani, MD, MA, director of clinical research, Sierra Eye Associates, said AI can also improve trial efficiency by helping identify candidates through imaging biomarkers and electronic health record data, which could speed recruitment and reduce screen failure rates.

Automating image analysis and screening

Several clinicians pointed to imaging as the area where AI's impact is most immediate. Joel Pearlman, MD, PhD, Retina Consultants Medial Group, said automated, consistent analysis of retinal imaging would help clinicians interpret and quantify how well treatments are working. Mina Massaro-Giordano, MD, professor of ophthalmology at NYU Langone Health, said AI can help clinicians see patterns across large sets of images that the human eye misses, such as changes in how quickly the tear film evaporates on the ocular surface, but cautioned that the results depend on feeding the software accurate, well-curated data: “If we're not putting the right information in, or the computer isn't taught well, we're not going to get that.” One researcher described a related project using AI to stitch a mosaic of the entire corneal surface, allowing clinicians to examine dry-eye corneal nerves rather than “just a piece of the puzzle.”

Paul Nderitu, MBChB, MPhil, MBCS, FRCOphth, PhD, a retinal specialist at Moorfields Eye Hospital, said AI's biggest global impact may come from lowering the cost and skill threshold needed to detect conditions such as cataract, uncorrected refractive errors, glaucoma, diabetic eye disease, and macular degeneration, using low-cost or handheld imaging devices in place of a specialist, with AI serving as “the gateway to detecting them” before referral and management.

Khanani agreed that AI-assisted screening could identify patients who would otherwise go unchecked, describing it as a way to flag which patients “should go get that checked out,” not a replacement for physician evaluation: “It's not like AI is going to completely take over the field of medicine—we're always going to need human doctors working with it.”

Powering research and drug development

Pearlman said there is “tremendous opportunity” for AI beyond the clinic, particularly in drug development, where AI now helps predict how proteins fold and assists in drug design—work recognized with a Nobel Prize. Pearlman said AI is unlikely to replace physicians, framing it instead as a new collaborator: “AI will not replace doctors; it'll only replace doctors who don't use AI.” Olawale Bankole, a graduate student in the Department of Biochemistry and Physiology at the University of Oklahoma Health Sciences Center, likewise pointed to AI's usefulness in research, particularly for consolidating large volumes of data from different sources into one place to help predict outcomes and guide researchers toward therapeutic targets for a given disease.

Andrew Pucker, OD, PhD, FAAO, FSLS, FBCLA, chief development officer at Mentra Health, described a more operational use case: at Mentra Health, a contract research organization, AI agents built with partner Tilda Research now handle tasks such as an “electronic trial master file,” or eTMF, agent that can take in a thousand documents at once, categorize them, and flag errors for review. Pucker said the company is expanding that approach with an agent to help research sites onboard faster and, soon, an electronic data capture system that would also screen incoming research data for errors.

Where clinicians see risk

Not every use case drew support. Khanani said patients using AI tools such as ChatGPT to self-diagnose or self-treat is “a really bad idea,” because the underlying algorithms draw from both reliable and unreliable sources and can produce wrong answers; Khanani said patients should still see a physician for an accurate diagnosis and treatment plan.

Nimesh Patel, of Mass Eye and Ear and Boston Children's Hospital, Boston, offered a different view, saying he encourages patients to research their conditions using AI or other tools before their next visit, arguing that patients who arrive with some baseline knowledge have easier conversations about starting treatment, particularly for conditions such as geographic atrophy or neovascular age-related macular degeneration that can otherwise feel abstract to a newly diagnosed patient.

Data privacy, equity, and staying human-centered

Khanani raised confidentiality as a central concern for AI in clinical trials and practice, including who owns patient data, where it is stored, who has access to it, and how it might be used beyond its original purpose.

Nderitu raised 2 further concerns. First, AI tools should be built to answer real clinical needs rather than generating new, unnecessary predictions: “There needs to be a strong link where the clinical need is the driver for the AI—not the AI becoming the driver for the clinical need.” Second, Nderitu warned that unequal access to AI, given its growing reliance on paid subscriptions, could widen existing disparities in care if the field does not ensure broad representation at the development stage—a dynamic Nderitu called “AI poverty,” arguing that AI should be made available “at different rungs of affordability.”

Pucker said AI has clearly streamlined research operations but stressed that a human still verifies the technology's output. “You need humans to double-check it,” Pucker said, “and you're always going to need humans there for the interaction element that we all need.” That sentiment echoed across the group: AI was widely described as a tool to support—not replace—physician judgment.

Across clinicians, a consistent throughline emerged: AI in retina and ophthalmology is best understood as a collaborator, not a replacement. Its clearest wins are already visible in imaging analysis, trial design, and research operations, where the technology helps clinicians and investigators do more, faster, and with fewer blind spots. Its harder questions—regulatory approval, data ownership, unsupervised patient self-treatment, and equitable access—are still being worked out, and several clinicians said those questions could take years to resolve.

For now, most speakers landed on similar middle ground. As Nderitu put it, "At the moment, AI is probably a supportive tool rather than a fully automated one, but as these tools become more and more capable, I think we'll head in that direction." The dividing question was never whether AI belongs in ophthalmology, but how carefully the field builds the guardrails around it.