The ability of deep learning architectures to identify glaucomatous optic neuropathy (GON) in fundus photographs was evaluated. A large database of fundus photographs (n = 14,822) from a racially and ethnically diverse group of individuals (over 33% of African descent) was evaluated by expert reviewers and classified as GON or healthy. Several deep learning architectures and the impact of transfer learning were evaluated. The best performing model achieved an overall area under receiver operating characteristic (AUC) of 0.91 in distinguishing GON eyes from healthy eyes. It also achieved an AUC of 0.97 for identifying GON eyes with moderate-to-severe functional loss and 0.89 for GON eyes with mild functional loss. A sensitivity of 88% at a set 95% specificity was achieved in detecting moderate-to-severe GON. In all cases, transfer improved performance and reduced training time. Model visualizations indicate that these deep learning models relied on, in part, anatomical features in the inferior and superior regions of the optic disc, areas commonly used by clinicians to diagnose GON. The results suggest that deep learning-based assessment of fundus images could be useful in clinical decision support systems and in the automation of large-scale glaucoma detection and screening programs.

To compare the diagnostic accuracy and explainability of a Vision Transformer deep learning technique, Data-efficient image Transformer (DeiT), and ResNet-50, trained on fundus photographs from the Ocular Hypertension Treatment Study (OHTS) to detect primary open-angle glaucoma (POAG) and identify the salient areas of the photographs most important for each model’s decision-making process. Evaluation of a diagnostic technology. Overall 66 715 photographs from 1636 OHTS participants and an additional 5 external datasets of 16 137 photographs of healthy and glaucoma eyes. Data-efficient image Transformer models were trained to detect 5 ground-truth OHTS POAG classifications: OHTS end point committee POAG determinations because of disc changes (model 1), visual field (VF) changes (model 2), or either disc or VF changes (model 3) and Reading Center determinations based on disc (model 4) and VFs (model 5). The best-performing DeiT models were compared with ResNet-50 models on OHTS and 5 external datasets. Diagnostic performance was compared using areas under the receiver operating characteristic curve (AUROC) and sensitivities at fixed specificities. The explainability of the DeiT and ResNet-50 models was compared by evaluating the attention maps derived directly from DeiT to 3 gradient-weighted class activation map strategies. Compared with our best-performing ResNet-50 models, the DeiT models demonstrated similar performance on the OHTS test sets for all 5 ground-truth POAG labels; AUROC ranged from 0.82 (model 5) to 0.91 (model 1). Data-efficient image Transformer AUROC was consistently higher than ResNet-50 on the 5 external datasets. For example, AUROC for the main OHTS end point (model 3) was between 0.08 and 0.20 higher in the DeiT than ResNet-50 models. The saliency maps from the DeiT highlight localized areas of the neuroretinal rim, suggesting important rim features for classification. The same maps in the ResNet-50 models show a more diffuse, generalized distribution around the optic disc. Vision Transformers have the potential to improve generalizability and explainability in deep learning models, detecting eye disease and possibly other medical conditions that rely on imaging for clinical diagnosis and management.

Purpose: To apply computational techniques to wide-angle swept-source optical coherence tomography (SS-OCT) images to identify novel, glaucoma-related structural features and improve detection of glaucoma and prediction of future glaucomatous progression. Methods: Wide-angle SS-OCT, OCT circumpapillary retinal nerve fiber layer (cpRNFL) circle scans spectral-domain (SD)-OCT, standard automated perimetry (SAP), and frequency doubling technology (FDT) visual field tests were completed every 3 months for 2 years from a cohort of 28 healthy participants (56 eyes) and 93 glaucoma participants (179 eyes). RNFL thickness maps were extracted from segmented SS-OCT images and an unsupervised machine learning approach based on principal component analysis (PCA) was used to identify novel structural features. Area under the receiver operating characteristic curve (AUC) was used to assess diagnostic accuracy of RNFL PCA for detecting glaucoma and progression compared to SAP, FDT, and cpRNFL measures. Results: The RNFL PCA features were significantly associated with mean deviation (MD) in both SAP (R² = 0.49, P < 0.0001) and FDT visual field testing (R² = 0.48, P < 0.0001), and with mean circumpapillary RNFL thickness (cpRNFLt) from SD-OCT (R² = 0.58, P < 0.0001). The identified features outperformed each of these measures in detecting glaucoma with an AUC of 0.95 for RNFL PCA compared to an 0.90 for mean cpRNFLt (P = 0.09), 0.86 for SAP MD (P = 0.034), and 0.83 for FDT MD (P = 0.021). Accuracy in predicting progression was also significantly higher for RNFL PCA compared to SAP MD, FDT MD, and mean cpRNFLt (P = 0.046, P = 0.007, and P = 0.044, respectively). Conclusions: A computational approach can identify structural features that improve glaucoma detection and progression prediction.