AI Reads the Hidden Shape of the Optic Nerve to Spot Glaucoma in Highly Myopic Eyes

Diagnosing glaucoma in people with severe nearsightedness is one of ophthalmology’s stubborn problems. High myopia physically reshapes the optic nerve head, the point where the optic nerve exits the eye, and those myopia-driven distortions mimic the damage that glaucoma itself causes. Conventional imaging markers, and even standard deep learning models trained on them, can therefore confuse a stretched, tilted nerve with a diseased one. A new development study published in Journal of Translational Medicine now reports a different strategy: instead of measuring how thick tissues are, it analyzes their fundamental shape using a branch of mathematics called persistent homology, paired with deep learning segmentation of the optic nerve head.

The research team, led by Yingxiang Han and Xiaofei Wang with colleagues including Jost B. Jonas and Ya Xing Wang, built their framework on spectral-domain optical coherence tomography. Each eye contributed six radial B-scans, cross-sectional images centered on the optic nerve head. A deep learning system segmented six distinct tissues within these scans, from the retinal nerve fiber layer and lamina cribrosa to the peripapillary sclera. Segmentation quality, quantified by Dice coefficients, ranged from 0.747 for the lamina cribrosa, the mesh-like structure through which nerve fibers pass, up to 0.935 for the retinal layers, indicating substantial but imperfect agreement with reference contours.

The innovation lies in what happened next. Rather than reducing the segmented tissues to a handful of thickness numbers, the researchers assembled the tissue contours into eye-level three-dimensional point clouds and converted them into H1 persistence images, a compact visual encoding of topological features that persist across spatial scales. A multilayer perceptron, a simple type of neural network, then classified these topological fingerprints as glaucomatous or not. The development cohort comprised 200 eyes from 154 participants: 100 highly myopic eyes without glaucoma and 100 highly myopic eyes with glaucoma, with all comparisons made using participant-grouped splits repeated across five rounds of stratified five-fold cross-validation.

The results were strong. In the development cohort, a RETFound deep learning classifier fed the segmentation maps achieved an area under the receiver operating characteristic curve of 0.936 ± 0.009 (95% confidence interval, 0.898–0.967). The topological route performed comparably: a multilayer perceptron using persistence features derived from the retinal nerve fiber layer alone reached an AUC of 0.931 ± 0.008 (95% CI, 0.891–0.962), while features combining the nerve fiber layer, peripapillary sclera and lamina cribrosa reached 0.942 ± 0.007 (95% CI, 0.908–0.976), a difference from the nerve fiber layer features alone that was not statistically significant after adjustment (P > 0.999).

Perhaps the most telling comparison came from stripping the approach down to its conventional counterpart. The same neural network architecture fed 12 standard retinal nerve fiber layer thickness values managed only an AUC of 0.737 ± 0.012, and logistic regression on those values reached 0.839 ± 0.007. The topological features significantly outperformed both (adjusted P < 0.001 and P = 0.001, respectively). In other words, the shape information captured by persistent homology carried diagnostic signal that traditional thickness measurements discarded in this cohort.

Crucially, the team then tested whether the pipeline, trained once on the development cohort, could survive contact with new data. Applied without any retraining to a temporally separated 2023 cohort of 79 eyes from 79 participants imaged on the same platform, the nerve fiber layer-based classifier achieved an AUC of 0.923 (95% CI, 0.849–0.983). Testing on temporally separated data is a meaningful stress test for any diagnostic model, since scanner drift, patient mix and image quality can shift over time.

The authors are careful about what these numbers do and do not mean. They frame retinal nerve fiber layer-derived topological features as an investigational adjunct for characterizing optic nerve head structure, not a replacement for clinical assessment. The study was a development study with participant-grouped cross-validation rather than a prospective clinical trial, and the cohort was limited to highly myopic eyes imaged on a single OCT platform. The researchers state explicitly that prospective multicenter and cross-device validation remains necessary.

Within those limits, the study demonstrates that topological features extracted from OCT contours can discriminate glaucoma in highly myopic eyes even when myopia distorts the usual anatomical landmarks, and that a classifier trained on one cohort retained its discrimination on a later cohort from the same platform. Whether that performance extends across devices, centers and broader populations is exactly what the remaining validation work will have to establish.

Subject of Research: Deep learning–assisted topological analysis of optic nerve head structures for glaucoma discrimination in highly myopic eyes

Article Title: Deep learning–assisted topological analysis of optic nerve head structures for glaucoma discrimination in highly myopic eyes: a development study

Article References: Han, Y., Wang, H., Dai, J., Zhang, C., Jonas, J. B., Lv, H., Wang, Y. X., & Wang, X. (2026). Deep learning–assisted topological analysis of optic nerve head structures for glaucoma discrimination in highly myopic eyes: a development study. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-09094-x

Image Credits: AI Generated

DOI: 10.1186/s12967-026-09094-x

Keywords: glaucoma, high myopia, optic nerve head, deep learning, persistent homology, optical coherence tomography, retinal nerve fiber layer, lamina cribrosa, diagnostic AI, topological data analysis, RETFound, OCT imaging