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

Diagnosing glaucoma in highly myopic (HM) eyes remains challenging due to myopia-induced structural alterations of the optic nerve head (ONH) that confound conventional imaging biomarkers and deep learning models. This study developed and evaluated a shape-topology-based diagnostic framework that integrates deep learning–based ONH tissue segmentation with persistent homology (PH) analysis for glaucoma discrimination in HM eyes. The development cohort comprised 200 eyes from 154 participants: 100 HM eyes from 84 participants and 100 HM eyes with glaucoma from 70 participants, with six radial spectral-domain optical coherence tomography (OCT) B-scans per eye. Six ONH tissues were segmented; tissue contours were assembled into eye-level three-dimensional point clouds, transformed into H 1 persistence images, and classified with a multilayer perceptron (MLP). Image-based deep learning classifiers and PH feature–based MLP classifiers used identical participant-grouped splits with five repeats of stratified five-fold cross-validation. The PH pipeline developed in the development cohort was subsequently evaluated in a 2023 temporally separated cohort of 79 eyes from 79 participants. Segmentation Dice coefficients ranged from 0.747 for the lamina cribrosa to 0.935 for the retinal layers. In the development cohort, the RETFound classifier using segmentation maps as input achieved an area under the receiver operating characteristic curve (AUC) of 0.936 ± 0.009 (95% confidence interval [CI], 0.898–0.967). The MLP classifier using RNFL-derived PH features achieved an AUC of 0.931 ± 0.008 (95% CI, 0.891–0.962), whereas the MLP classifier using PH features derived from the RNFL+peripapillary sclera+lamina cribrosa combination achieved 0.942 ± 0.007 (95% CI, 0.908–0.976), without a significant adjusted difference from RNFL alone ( P > 0.999). The MLP classifier using RNFL-derived PH features outperformed the same-architecture MLP using 12 RNFL thickness values (AUC, 0.737 ± 0.012; adjusted P < 0.001) and logistic regression using those values (AUC, 0.839 ± 0.007; adjusted P = 0.001). In the temporally separated cohort, the development-trained MLP classifier using RNFL-derived PH features achieved an AUC of 0.923 (95% CI, 0.849–0.983). The MLP classifier using PH features derived from RNFL contours showed strong participant-grouped discrimination and retained performance when applied to a later-period same-platform cohort without retraining. These findings support RNFL-derived PH features as an investigational adjunct for characterizing ONH structure rather than a replacement for clinical assessment. Prospective multicenter and cross-device validation remains necessary.