External validation of a deep-learning algorithm for perimetric glaucoma detection on color fundus photographs in the Rotterdam Study.

Purpose

To externally validate a deep-learning (DL) algorithm detecting perimetric glaucoma on color fundus photographs (CFPs) in a population-based cohort.

Design

External validation in a prospective population-based cohort.

Participants

Participants aged 50-80 years from three Rotterdam Study cohorts.

Methods

We applied the winning Artificial Intelligence for Robust Glaucoma Screening (AIROGS) challenge algorithm, trained on 101,442 macula-centered CFPs to detect referable presumed perimetric glaucoma. Performance was assessed at eye and participant level, with participant-level cluster bootstrap confidence intervals. Misclassified cases underwent qualitative review. Longitudinal detection was evaluated by Kaplan-Meier analysis. Additionally, paired analyses compared macula- with disc-centered CFPs. A post-hoc analysis examined the effect of image quality on positive predictive value (PPV). MAIN OUTCOME MEASURES Sensitivity at 95% specificity, area under the receiver operating characteristic curve (ROC-AUC), partial AUC (pAUC) for 90-100% specificity, PPV, and precision-recall AUC (PR-AUC).

Results

Among 19,088 eyes (9,837 participants), 201 eyes (1.1%) had perimetric glaucoma. At eye level, sensitivity at 95% specificity was 70.6% (95% CI: 63.8-76.8), with an ROC-AUC of 0.911 (95% CI: 0.887-0.934) and a PR-AUC of 0.343 (95% CI: 0.274-0.427). At participant level, sensitivity was 68.8% (95% CI: 61.8-75.9) and ROC-AUC 0.909 (95% CI: 0.882-0.934). Overall, this resulted in 5.1 referrals per detected glaucoma case. True positives showed more advanced visual field damage than false negatives (median mean deviation -10.8 vs -5.8 dB; P < 0.001). Among true positives with prior imaging, 33 of 61 (54%) were flagged before conversion, typically 4-7 years before diagnosis. Discrimination in the 90-100% specificity range was higher for disc- than macula-centered CFPs (pAUC 0.860 vs 0.802; P = 0.003). PPV improved substantially after excluding CFPs with incomplete optic nerve head (ONH) visibility.

Conclusions

The DL-algorithm achieved meaningful discrimination for perimetric glaucoma in an external population-based dataset. Correctly identified cases were frequently flagged at an earlier examination, years before the perimetric diagnosis. Despite having been trained exclusively on macula-centered CFPs, the DL-algorithm showed no loss of discriminative performance on disc-centered CFPs. Incomplete ONH visibility was an important source of false positives. These findings support CFP-based deep learning as a scalable triage tool for glaucoma screening and highlight image acquisition standards for implementation.