Framework Detects Glaucoma Across Hospitals

A neuro-symbolic AI system combines image analysis with clinical rules to identify glaucoma from fundus photographs.

Updated on Nov. 11, 2026 in Stroke

Bold flat-color editorial illustration of a stylized optic disc with concentric rings and branching vessel lines on a deep teal field, evoking retinal imaging for glaucoma detection.
Researchers presented a neuro-symbolic framework that detects glaucoma from retinal images and holds up across hospitals and imaging devices, including a switch from Canon to Nidek cameras. AI Illustration. Upload story photo >

Researchers presented a neuro-symbolic framework that detects glaucoma from fundus images, combining deep learning with clinical fuzzy logic rules. The system reached a test AUC of 0.946 and sensitivity of 0.900 on its primary dataset.

Why it matters

An estimated 79.6 million people were affected by glaucoma in 2020, a figure projected to reach 111.8 million by 2040. Most automatic detection systems do not generalize well across hospitals and imaging devices, which limits their clinical use.

The full model reached a test AUC of 0.946 and sensitivity of 0.900 on 9,100 training images drawn from a 12,619-image corpus across eight datasets. Cross-hospital performance was lower, with the best variant reaching an AUC of 0.6997 on a 485-image held-out set, and the authors note that the cup-to-disc proxy's agreement with expert measurements remains weak and inconsistent.

The details

The framework combines an EfficientNet-B4 backbone with dual spatial-channel attention, a six-node anatomical graph attention network, and a differentiable symbolic constraint module that applies four clinical fuzzy logic rules involving cup-to-disc ratio, rim thinning, RNFL defects and vessel density. It was tested in a cross-hospital setting moving from Canon CR-2 to Nidek AFC-210 imaging devices without using segmentation masks.

Timeline

  1. 2020: An estimated 79.6 million people were affected by glaucoma.

  2. 2026-11-11

    The article was published with early access.

  3. 2040: Projected glaucoma cases are expected to rise to 111.8 million.

Health Landscape

The framework extends the neuro-symbolic approach to glaucoma screening by pairing EfficientNet-B4 image features with four encoded clinical fuzzy logic rules. Its cross-hospital results point to the device-shift problem that has limited earlier automated glaucoma detection systems.

For patients, automated screening that works across different clinics and cameras could make glaucoma checks more consistent when a patient moves between providers. The framework's cross-hospital AUC of 0.6997 shows that device-to-device performance is still a limitation before such tools reach routine care.

The takeaway

Automated glaucoma detection remains strong on its own data but drops sharply when moved to new hospitals and devices. Any clinical tool will need testing across imaging equipment before it can be trusted in everyday screening.

Further reading

Explore more on Stroke to see how automated detection research fits into the wider field of neurological and eye-related disease.

Source note: This article includes information reported by Nature.