A-Eye

How artificial intelligence is revolutionizing glaucoma detection and treatment

Artificial intelligence (AI) may soon help doctors detect glaucoma earlier, enabling more precise treatments for the 4.2 million Americans living with the condition. Glaucoma, a group of eye diseases that damage the optic nerve, is one of the leading causes of irreversible blindness worldwide. In the United States, it affects about 2.5% of adults ages 40 and older. The disease impacts people at different rates: African Americans have glaucoma at six times the rate of European Americans and are more likely to experience earlier onset and more severe vision loss, according to a 2022 study published in Missouri Medicine.

“Half of the people with glaucoma in the United States are not aware of their diagnosis,” says Angelo P. Tanna, MD, professor and vice chair of ophthalmology at Northwestern University’s Feinberg School of Medicine. Limited access to routine eye care and screening further widens this gap, particularly in communities already at higher risk.

The disease often progresses without noticeable symptoms. Because it does not affect central vision early on, many patients delay seeking care. As a result, many enter treatment at later stages, when vision loss is more likely.

Teaching computers to see glaucoma

Glaucoma occurs when damage to retinal ganglion cells leads to gradual vision loss. Doctors detect the disease by observing progressive structural changes in the optic nerve, especially a feature known as cupping, Tanna says.

“That really is the hallmark finding of glaucoma,” he says, adding that people with healthy eyes also can have large optic disc cups. “This is especially common in people who are very near- sighted and people with large optic nerves. To make matters more confusing, both being near-sighted and having a large optic nerve are risk factors for developing glaucoma,” he says.

Because these similar appearances can occur in healthy eyes, even specialists may disagree in borderline cases. AI is no different.

“One of the biggest challenges with the use of AI for glaucoma detection is that there isn’t a universally accepted definition of what constitutes glaucoma,” Tanna says. “The diagnosis is, in part, subjective.”

Physicians diagnose and monitor glaucoma using imaging scans such as optical coherence tomography (OCT) and visual field tests that detect subtle losses in peripheral vision, which often deteriorate before central vision.

“So we look at everything from the center to the mid-periphery in visual field testing,” Tanna says. “And we compare our patient, one eye at a time, to a normative database. The computer does it automatically.”

AI helps clinicians interpret tests more quickly and consistently by applying comparisons to global models trained on 1.6 million retinal scans, according to a study published in Nature. By analyzing images, these models can detect early signs of nerve damage that human observers may miss, researchers report.

“There are undoubtedly features of these images that we’re not able to detect with our human eyes,” says Nicholas Volpe, MD, vice chair of ophthalmology at Northwestern. “An educated machine could theoretically say these patients are the ones that are getting worse.”

Tanna envisions a future in which patients visit pharmacies or community clinics for AI-assisted eye scans to estimate glaucoma risk.

Ongoing hurdles

Current AI systems still face technical hurdles. Tanna says some research models detect glaucoma with roughly 80% to 90% sensitivity and specificity — an encouraging but insufficient level for widespread screening. Population-level screening programs require accuracy above 90%.

Similar challenges exist across healthcare, where AI systems must interpret data that varies among hospitals, imaging devices, and patient populations. Yuchen Liu, PhD, an assistant professor in the department of computer science at North Carolina State University, says results from controlled training environments often do not fully translate to real-world reliability due to differences in clinical settings and patient demographics.

AI should remain “a decision-support tool,” Liu says, with clinicians holding final responsibility for medical decisions.

The Center for Engineering in Vision and Ophthalmology (CEVO) at Northwestern uses imaging data and computational models to improve glaucoma detection.

While some researchers focus on diagnosis, others are exploring how AI might improve treatment. Biomedical engineers and ophthalmologists at Northwestern are collaborating to develop imaging-based models that could guide glaucoma surgery, which improves fluid drainage to lower pressure inside the eye. These procedures do not restore damaged optic nerve cells but aim to slow further vision loss.

When a person’s glaucoma no longer responds to eye drops, they may need surgery. But physicians don’t always know which treatment will produce the best long-term result, making surgical planning a key target for AI-assisted modeling.

The work involves Hao F. Zhang, PhD, co-director of CEVO, whose laboratory specializes in advanced eye imaging technologies.

Zhang’s team is developing models that combine imaging data with computational models to predict how surgical implants might affect fluid drainage inside the eye. AI analyzes imaging data, identifies anatomical features, and predicts how factors such as implant placement and number may affect outcomes.

These implants drain fluid from the eye to lower pressure. Zhang says current practice often relies on physician judgment rather than data-driven planning. The approach could improve long-term outcomes by reducing guesswork.

“AI provides information,” Zhang says. “But the final decision has to come from the doctor.”

The goal, researchers say, is not to replace physicians, but to help them act more effectively in a disease that often goes unnoticed until it’s too late.