July 8, 2026 | Thanks to several key advances in deep learning architecture, optical coherence tomography (OCT)—one of the most obtained medical images in the world—can now be used to predict disease one patient at a time. The people who stand to benefit are not only those suffering from common retinal diseases, but also conditions such as heart attack, stroke, and kidney failure simply by analyzing microscopic features in the eye, according to Aaron Lee, M.D., head of the Hardesty Department of Ophthalmology and Visual Sciences at Washington University School of Medicine in St. Louis.
The enabling technology is a new experimental AI framework, known as OCTCube-M, which is designed to analyze the three-dimensional (3D) images produced by the ubiquitous eye scans. The framework was recently found to do a better job than older 2D models of identifying eight different retinal diseases, including age-related macular degeneration and diabetic retinopathy, as reported recently in Nature Biomedical Engineering (DOI: 10.1038/s41551-026-01662-2). It also more accurately predicted how fast a severe form of the condition, called geographic atrophy, would progress.
The 3D foundation model was trained on more than 1.62 million individual retinal slices from 26,000 OCT images. It was previously technically unfeasible to take advantage of all that dense, high-resolution medical data at once, says Lee. Since humans are only used to reading images in 2D, researchers used a trio of AI models—one each for processing one (unimodal), two (bimodal), or three (trimodal) data types—to read and interpret 3D images of the eye’s retina.
When compared to the model trained on 2D images, OCTCube-M more accurately identified six of the eight retinal diseases by about 4 to 6 percentage points. That amounts to the detection of an additional 43 to 60 cases per 1,000 individuals, and the improvement held true across multiple clinical sites, patient populations, and imaging modalities, he says.
People diagnosed with macular degeneration get OCT scans on a regular basis, meaning an immense amount of eye scan data exists for training AI models to make them more intelligent, Lee points out. Macular degeneration also remains one of the leading causes of blindness in the world, despite a breakthrough therapy for the wet variety that Genentech brought to market in 2006 (Lucentis) that can stop and sometimes reverse vision loss.
Wet macular degeneration once meant almost certain blindness, Lee says. “We used to tell patients go home and memorize the faces of their grandchildren because they wouldn’t see them again in a year ... it was a really heartbreaking disease.”
The remaining problem is that no good treatments exist for geographic atrophy, the dry form of macular degeneration that causes vision loss. “We don’t have a lot to offer these patients, and those lesions just grow and grow and grow and eventually make people go blind,” says Lee.
OCTCube-M could potentially be of use to pharma companies developing treatments to make their trials “extraordinarily more powerful” by enhancing statistical analysis with virtual data, as well as rapidly identify and enroll patients, says Lee, who has been trying to make that case with industry sponsors. “We open-sourced the architecture of the AI model as well as all the weights and parameters that go into it,” he notes.
Beyond the realm of ophthalmic diseases, OCTCube-M can be used in the new research area of “oculomics” to analyze structural eye scans (like OCT) alongside 2D images (like infrared or fundus photos) to better understand other parts of the human body for disease prediction purposes. As the recent study suggests, imaging the retina in such fine detail led to significant gains in the ability to predict things such as blood sugar levels, kidney functioning, and the risk of heart disease.
The tiny blood vessels in the retina share the same basic structure as those in other organs, Lee explains. Disease processes from those other body parts also leave signatures in the eye.
“The field of ophthalmology has known for decades that there’s a tight connection between the configuration of the blood vessels and [multiple systemic and chronic diseases] ... but those have been population association type studies,” says Lee. “We’ve never had tools that were powerful enough to predict it at an individual patient level until the advent of AI.”
Lee believes that the multi-foundational model has several particularly promising applications in the design of clinical trials, including reducing the required sample size and time it traditionally takes to run a study—be it in ophthalmology or any other therapeutic area. “If we can accelerate the field of drug development, we are going to get more therapies that actually work in the hands of people who actually need them,” he says.
This is where a collaboration with Genetech, which has access to randomized controlled trial data, played a pivotal role in the latest OCTCube-M study, he adds. As was discovered, the framework could be used to substantially lower the number of patients needed for trials to get equivalent results.
A decade ago, Genentech ran a large and expensive phase 3 trials for lampalizumab to evaluate the drug's ability to slow the progression of geographic atrophy, but ultimately it showed no meaningful difference in slowing lesion growth, says Lee. Since then, Genetech and other pharma companies have been investigating other therapeutic agents that could potentially slow down the disease, he adds.
In 2023, the FDA approved a pair of drugs for treating geographic atrophy (Syfovre, developed by Apellis Pharmaceuticals and Izervay, developed by Astellas/Iveric Bio). The first wave of real-world outcome data confirms they successfully decelerate disease progression while highlighting rare but serious risks like intraocular inflammation and endophthalmitis.
In future trials, OCTCube-M could be used to help identify patients who are highly similar to each other and thus make the overall group as homogenous as possible, Lee says. This would ensure that their baseline characteristics are uniform before randomization so that differences in health outcomes can confidently be attributed to the treatment being studied rather than pre-existing variations among participants.
Homogeneity is today primarily established using inclusion/exclusion criteria alongside clinical impressions, both of which have their limitations, says Lee. Restrictive eligibility rules shrink the pool of eligible participants, while over-reliance on clinician assessment can introduce human bias, regional practice differences, and subjective variability.
Across medicine generally, OCTCube-M could help with clinical trials by more quickly identifying eligible patients being seen in routine clinical practice, he continues. Physicians who are running an hour behind in the clinic are simply too busy to make that determination at the point of care.
Perhaps the most exciting possibility is to use the new AI system to create digital twins to “encapsulate how well a patient is going to do without [actually] getting that experimental therapy because they are under the standard of care,” says Lee. “You can train these models to predict the outcome for each individual patient that is about to go through the treatment arm and thereby create a counterfactual of what would have happened to that patient had they not gone through the therapeutic arm.”
This would enable a “paired statistical trial” matching each real participant with an AI-generated virtual model that predicts that patient’s specific disease progression without the experimental treatment. Researchers can then compare the patient's actual results against their own virtual baseline to measure the true treatment effect.
Having been engaged by the Food and Drug Administration (FDA) on the topic of digital twins, Lee has some insights on regulatory thinking about their deployment in trials. The agency’s view is likely to favor applying them first in phase 2 studies, the go/no-go decision point in clinical drug development, he shares.
“The pharma companies [and] sponsors of the clinical trials are really the ones who would be taking the risk,” Lee says. They might use a digital twin model to perform a phase 2 trial in an abbreviated period and, if the treatment effect is found to be as large as they hoped, invest in the final and largest stage of testing.
For phase 3 trials, the agency would need greater assurance that the results were as good as the sponsor claims, he continues. “But there are situations in medicine where it makes almost ethical sense to think about these technologies very seriously.” These include oncology and cardiovascular clinical trials where patients don’t want to be randomized into the placebo arm, and mortality outcomes may be measured in hours or days.
Single-arm trials also make sense when it comes to investigational treatments for rare or uncommon diseases. “By the time you finish the whole phase 1, phase 2, phase 3 in a traditional paradigm, you may have treated just about everybody on earth with those conditions,” says Lee. It is more appropriate to have them potentially benefit from the therapy being investigated.
Importantly, this reflects the thinking expressed in FDA guidance documents and implemented by the agency in real-world regulatory contexts.
‘Commercially Friendly’ Model
Since the AI model has been open-sourced in a “commercially friendly” way—meaning, companies can freely use the OCTCube-M tool framework to accelerate their AI goals, says Lee. That’s a relative rarity in the world of foundation models.
“We hope to see a new generation of AI models ... [using] this as a steppingstone to do many downstream things like the quantification and identification of different retinal pathologies,” Lee says. “We don’t have great tools for that available ... [to] the clinical world.”
The latest published paper also provides clinical trial sponsors with easy access to a foundation model to support their clinical trials. Heretofore, they have often been “tied up in license agreements that were unfavorable for them,” says Lee, who has no personal commercial ambitions with the technology.
“I think that if you want to make a difference in the world you have to be okay with the idea of letting commercial companies turn those things into medical products that will actually get used because in the world of academics, we don’t have the resources to go through the regulatory process to turn a piece of software into a ... [clinical-grade] medical device,” he says. That regulatory pathway requires large prospective studies to benchmark results.
Lee says he hopes others share his optimism about the ongoing Fourth Industrial Revolution—a phrase coined by the World Economic Forum a decade ago—where “AI is going to augment, change, and completely replace in some cases, many things that humans have been doing for a very long time. We don’t even have a framework for understanding how profound of a change that we are going to see, and it’s both a very scary time but also an exciting time to be in the technology space working at the interface of AI and medicine.”