Episode Details
Back to Episodes
AI Detected Signs of Glaucoma Years Before Diagnosis Using Ordinary Eye Photos — A New 19,000-Eye Study
Description
This audio article is from VisualFieldTest.com.
Read the full article here: https://visualfieldtest.com/en/ai-detected-signs-of-glaucoma-years-before-diagnosis-using-ordinary-eye-photos-a-new-19-000-eye-study
Test your visual field online: https://visualfieldtest.com
Support the show so new episodes keep coming: https://www.buzzsprout.com/2563091/support
Excerpt:
AI Detected Signs of Glaucoma Years Before Diagnosis Using Ordinary Eye Photos — A New 19,000-Eye Study Important clarification: the study found that an artificial-intelligence algorithm sometimes flagged eye photographs taken years before a person first met the study’s visual-field-based definition of glaucoma. It did not show that the algorithm predicted when glaucoma would begin, or prove that those people already had clinically confirmed glaucoma at the earlier visits. Introduction Glaucoma can damage the optic nerve—the connection between the eye and the brain—without causing obvious symptoms at first. A photograph of the back of the eye can show some structural changes associated with glaucoma, but photographs alone cannot establish the full diagnosis. A study published online in Ophthalmology on November 2, 2026, tested a glaucoma-screening algorithm on archived photographs from the population-based Rotterdam Study. It included 19,088 eyes from 9,837 people. At a threshold chosen to produce 95% specificity, the algorithm detected about seven in ten eyes that met the study’s definition of perimetric glaucoma—glaucoma accompanied by visual-field loss. Its area under the receiver operating characteristic curve was about 0.91. Source 1 () The result that has attracted the most attention is this: among 61 correctly flagged glaucoma cases with earlier photographs available, 33 had also been flagged on an earlier photograph, typically four to seven years before the study recorded their visual-field-based diagnosis. That is intriguing—but “flagged before diagnosis” is not the same as “predicted glaucoma before it began.” The distinction matters for understanding what the study proves, what it does not prove, and whether this approach is ready for mass screening. What the Rotterdam Study tested A photo-based triage tool, not a complete eye examination The researchers applied the winning algorithm from the Artificial Intelligence for Robust Glaucoma Screening challenge to color photographs taken with retinal cameras. The algorithm had been trained on 101,442 photographs from 54,274 people, collected through 486 diabetic-retinopathy screening centers. Its task was to identify photographs likely to show referable, presumed perimetric glaucoma. Source 1 () The photographs were color images of the retina, not ordinary snapshots taken with a household camera. The algorithm analyzed the images; it did not measure eye pressure, examine the eye’s drainage angle, or test how well the person could see at different points in their visual field. In other words, it was designed as a possible screening and referral tool, not as a substitute for an eye examination. The Rotterdam Study is a prospective cohort, meaning participants were followed over time. But the algorithm’s analysis of their stored photographs was retrospective: the researchers applied the model to images from previous visits and compared its output with the study’s glaucoma classification. Complete performance data reported in the study The researchers reported performance at two levels: per eye and per person. Those figures answer different questions. For the full table, please open this article on visualfieldtest.com. At the eye level, the algorithm correctly flagged 142 of 201 glaucoma eyes. It class