ASRS 2026: Large language and vision-language models could ease the burden on retina specialists

At the 2026 annual meeting of the American Society of Retina Specialists (ASRS), held July 15-18 in Montreal, Dhruv Patel, a senior at the University of Chicago and undergraduate research assistant in the Department of Ophthalmology and Visual Science, presented a poster on how large language models (LLMs) and vision-language models (VLMs) can support retina specialists. Patel developed the study with co-author Rohin Sagar, a researcher affiliated with the University of Chicago and Northwestern Medicine, along with Rahil Bhatia, Howard Fine, MD, MHSc, Uma Patel, and Jay Chhablani.

Clinical applications

The poster frames LLMs and VLMs as tools that augment retina specialists rather than replace them. According to Patel and Sagar, VLMs can analyze imaging such as optical coherence tomography (OCT) to flag findings like increased subretinal fluid, while LLMs draw on a patient's electronic health record, including prior history, laboratory results, and medications, to help build a more tailored treatment approach. The models can also flag inclusion and exclusion criteria for clinical trials, helping identify patients who may qualify for a study but are unaware that they do.

Patel and Sagar pointed to the growing burden of electronic health records and data on clinicians as a driver of the research, noting that this burden reduces time available for direct patient interaction. They said LLM and VLM tools are meant to consolidate that information and streamline workflow—not make final clinical decisions.

Balancing innovation with oversight

The study also outlines limitations that must be addressed before wider adoption, including AI hallucinations, health equity, and preservation of existing patient biases within medicine. Patel and Sagar described instances in which a VLM misread an OCT scan or an LLM cited a source that did not exist. Because of these risks, they said the clinician must remain the final decision-maker, using AI output as a supporting reference rather than a directive. Currently, AI in ophthalmology is used mainly on the administrative side, assisting with forms and paperwork rather than direct patient care.

Looking ahead

Patel and Sagar anticipate that within 5 to 10 years, multimodal AI models trained alongside physicians could help identify ocular biomarkers and the physiology underlying specific conditions from retinal images. They also expect future systems to connect patient health data directly to ongoing clinical trials, automating the process of matching patients to studies for which they qualify. Patel and Sagar concluded that LLMs and VLMs represent an essential tool to enhance—not replace—the clinician's workflow, allowing more time for diagnosis, treatment, and a more personalized care plan.

Reference:

1. Patel D, Bhatia R, Sagar R, Fine H, Patel U, Chhablani J. Clinical decision support in retina using large language models and vision-language models. Presented at: American Society of Retina Specialists (ASRS) 44th Annual Scientific Meeting; July 15-18, 2026; Montreal, Quebec, Canada.