涩里番

Five years ago, a learned he had diabetes. At first, he kept his blood sugar in check, took medications on schedule and saw an eye doctor every year. But, after a difficult divorce, he stopped his medications, relied on fast food and soon his vision began to blur.

Fundus photograph of a diabetic retina exhibiting hallmark hemorrhages and microvascular damage characteristic of diabetic retinopathy, emphasizing the critical need for early detection and intervention.

At his next eye exam, the doctor found small retinal hemorrhages, a sign of diabetic retinopathy, or DR, in both eyes. Caught early, DR can often be slowed with laser treatment. Once complications set in, patients may face invasive surgery.

Caused by long-term high blood sugar that damages retinal vessels, DR is the leading cause of blindness worldwide. Yet, up to are preventable with regular screening and early treatment, according to The Lancet.

Artificial intelligence, or AI, is giving ophthalmologists new ways to detect eye disease earlier. Before patients notice blurry vision, AI can reveal cellular changes in the retina, allowing treatment to begin sooner and vision loss to be prevented altogether.

Some AI tools are already in clinics, but research continues to push the field forward. Advances in imaging and diagnosis could extend vision care even to people without access to eye specialists.

A sharper view of disease

National Center for Microscopy and Imaging Research via NIH Flickr

Laser scanning microscope image of a healthy mouse retina in the optic fiber layer. This layer is responsible for relaying information from the retina to the brain and was fluorescently stained to reveal the distribution of glial cells (green); DNA and RNA in the cell bodies of the retinal ganglion neurons (orange); and their optic nerve fibers (red); and actin in endothelial cells surrounding a prominent branching blood vessel (blue).

The eye is a biological camera, capturing not only the world around us but also early signs of systemic disease. Now, AI is giving the camera an upgrade, making it a sharper, faster, more accessible diagnostic tool that can reveal disease before symptoms appear.

Most eye exams include fundus photography, a snapshot of the retina, optic nerve and vessels. These images form the baseline of modern vision care, and with AI, they become even more powerful. Algorithms can now analyze the photos to provide earlier, clearer answers, often with quicker, more comfortable scans.

By spotting microscopic retinal changes, AI can flag DR long before symptoms, speeding treatment and improving outcomes. Beyond diabetes, AI scans are also showing early signs of cardiovascular, hypertensive and even neurological disease.

Equity in eye care

As populations age and systemic diseases rise, early diagnosis and long-term monitoring are more critical than ever.

According to , resident ophthalmologist in the department of ophthalmology and vision sciences at the University of Toronto, fundus images reveal more than vision. They reflect systemic health.

Yet the most advanced cameras remain out of reach in rural and under-resourced areas, and travelling to specialists can be difficult. In addition, low-quality images from non-specialist clinics can delay diagnoses and worsen outcomes.

Tele-ophthalmology helps bridge the gap: clinicians send images for remote interpretation, and specialists return diagnoses and treatment plans.

“It allows us to extend our reach far beyond the physical walls of our clinics,” Balas said.

By integrating AI into everyday clinics, even older devices can generate diagnostic-quality scans, a step toward earlier, more equitable care.

Had such AI-powered screening been available at his primary-care office, the manager from the story’s opening might have been flagged months earlier, sparing him invasive surgery.

Sharpening the picture

Fluids and tissues inside the eye distort light, adding blur to optical images. To correct this, researchers use , or AO, a technology originally developed in the 1950s to sharpen telescope views of distant stars.
An AO-assisted retinal imager uses a deformable mirror to cancel distortions, producing corrected images captured by a camera. The addition of AI reduces AO scan times from hours to minutes while preserving cellular detail.

Researchers at the National Eye Institute, or NEI, at the National Institutes of Health developed an AI method that extracts more data from fewer low-quality AO images, cutting analysis time from days to hours.

“So many people think about AI as a tool that you apply after the images (are collected),” , an NEI senior investigator, said. “We’re thinking about AI as part of the whole imaging procedure.”

Tam and , a biomedical engineer at NEI’s Clinical and Translational Imaging Section, that recovers features of the retinal pigment epithelium, or RPE, from standard clinical scans. RPE cells nourish photoreceptors, which enable vision and are disrupted in many eye diseases. Seeing cell-level changes early can change the course of care.

Diagram showing various cell types that make up the retina.

“Medical imaging is a tissue-level technology,” Tam said. “The ability to go in and look at individual cells is a game-changer.”

The NEI team also developed P-GAN, an AI model that sharpens AO optical coherence tomography, which images 3D retinal structures at the cellular level. Instead of collecting hundreds of 3D images to reduce noise, P-GAN cleans them computationally, even improving results from low-cost cameras.

“It can supercharge an ordinary device and make it even more helpful to both the scientists and clinicians,” Li said. “It’s not about replacing. It’s about assisting.”

Tam hopes similar tools will one day deliver real-time insights during routine exams.

Speeding up the scan

The Tam lab used AI to collect fewer images with high-quality results. However, to see some features of the retina, many images at different angles of light refraction are necessary.

At Duke University, researchers developed an open-source AI system that separates many angles of captured light into individual images, boosting the speed and accuracy of retinal scans.

“The difference between this system and a system developed on mice is that we can put a mouse in front of an imager for five hours,” Sina Farsiu, a professor of biomedical engineering and director of the Vision and Image Processing Laboratory at Duke, said. “A patient in their 50s or 60s wouldn’t be able to sit in front of a bright light for a long time.”

To overcome that challenge, Farsiu’s team combines machine learning and medical imaging into a technique called deep compressed AO scanning light ophthalmoscopy, or . The technique combines adaptive optics with machine learning, using multiplexed light signals to capture single-cell detail in a fraction of the time, cutting acquisition by nearly 100-fold.

AO focuses light on the single-cell level, while AI converts light data into clear images, like decoding a secret message. By combining the two, DCAOSLO delivers high-resolution, comprehensive images in minutes instead of hours, with less strain on patients.

The result: faster scans, earlier detection and greater comfort for patients.

Farsiu hopes that DCAOSLO will reshape screening workflows, noting that open-source AI is already “leveling the playing field” by squeezing more performance from affordable cameras.

“You can detect diseases at a much earlier stage, so your treatment will hopefully have a higher efficacy,” Farsiu said.

AI in the clinic

Trained on thousands of retinal scans, diagnostic AIs learn patterns of health and injury and flag who needs urgent referral.

A clinician takes retinal pictures with whatever camera is available; then AI analyzes the images against large databases, spotting subtle leaks, vessel damage or cell loss that the human eye might miss.

One clear example is , the first AI software approved by the U.S. Food and Drug Administration to detect and diagnose more than mild DR. In primary-care pilots, trained staff captured two images per eye. From just four retinal images per patient, the AI correctly identified nearly all patients with and without DR, in seconds.

The payoff: even if a primary-care provider isn’t comfortable reading retinas, the AI tells them who to refer, shifting detection earlier and expanding access.

For patients like the 35-year-old store manager diagnosed with DR, such tools could mean earlier referral and treatment, before vision loss becomes irreversible.

Beyond diabetes

Sonia Samtani et al., National Eye Institute, National Institutes of Health, via NIH Flickr

Confocal microscope image of an eye with choroidal neovascularization, which is common in patients with age-related macular degeneration. Endothelial cells are shown in red, F-actin of stress fibers is shown in green and cell nuclei are shown in blue.

Ophthalmologists are uncovering deeper links between eye structure and systemic disease.

“Linking structural biomarkers with genetic phenotypes is transforming how we stratify and manage patients,” , director of clinical research at the University of Pittsburgh Medical Center and a professor of ophthalmology, said.

Chhablani studies the choroid, the eye’s vascular layer, where early signs of diabetes, hypertension and cardiovascular disease often appear in small vessels and pericytes before they show in larger vessels near the heart.

Early, accurate diagnosis of progressive conditions such as age-related macular degeneration and inherited retinal diseases can prevent vision loss. Connecting imaging biomarkers, like geographic atrophy or photoreceptor loss, to specific diseases speeds diagnosis and guides treatment.

“A key challenge is translating complex imaging data into clinically actionable insights,” Chhablani said.

By tracking AI-derived biomarkers over time, clinicians could predict disease progression or treatment responses even outside the clinic. He envisions home-based imaging and remote monitoring extending care to underserved populations.

Standardized AI tools can help by improving reproducibility and clinician confidence and supporting clinical trials that track response to gene and cell-based therapies.

“Ultimately, we aim to integrate imaging, genetics and function into a single, personalized disease model,” Chhablani said.

Keeping the patient in focus

Loss of vision reshapes quality of life. Blindness progresses on a spectrum — from seeing normally, to seeing differently, to losing sight altogether.

Halting that progression requires early detection and continual monitoring. When patients receive sight-preserving treatments, regular exams confirm they’re working.

Whether AI clarifies low-quality images for remote review or flags patients for referral, the technology keeps researchers, clinicians and patients connected. Behind it all, Balas said, “the human element is still the optical nerve that keeps everything connected.”

The eye has always been a kind of camera, reflecting both vision and health. With AI, that camera is becoming sharper, faster and more accessible. For the millions worldwide at risk of diabetic retinopathy, including store clerks and CEOs, the promise of AI is not just sharper images but clearer futures.

“In ophthalmology, the future is brilliant,” Balas said. “But the present, where pixels meet people, is what gets me out of bed.”