Search

Purpose: 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. Design: Evaluation of a diagnostic technology. Subjects, Participants, and Controls: Overall 66 715 photographs from 1636 OHTS participants and an additional 5 external datasets of 16 137 photographs of healthy and glaucoma eyes. Methods: 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. Main Outcome Measures: 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. Results: 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. Conclusions: 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 assess whether the predictive accuracy of machine learning algorithms using Kalman filtering for forecasting future values of global indices on perimetry can be enhanced by adding global retinal nerve fiber layer (RNFL) data and whether model performance is influenced by the racial composition of the training and testing sets. Design: Retrospective, longitudinal cohort study. Participants: Patients with open-angle glaucoma (OAG) or glaucoma suspects enrolled in the African Descent and Glaucoma Evaluation Study or Diagnostic Innovation in Glaucoma Study. Methods: We developed a Kalman filter (KF) with tonometry and perimetry data (KF-TP) and another KF with tonometry, perimetry, and global RNFL data (KF-TPO), comparing these models with one another and with 2 linear regression (LR) models for predicting mean deviation (MD) and pattern standard deviation values 36 months into the future for patients with OAG and glaucoma suspects. We also compared KF model performance when trained on individuals of European and African descent and tested on patients of the same versus the other race. Main Outcome Measures: Predictive accuracy (percentage of MD values forecasted within the 95% repeatability interval) differences among the models. Results: Among 362 eligible patients, the mean ± standard deviation age at baseline was 71.3 ± 10.4 years; 196 patients (54.1%) were women; 202 patients (55.8%) were of European descent, and 139 (38.4%) were of African descent. Among patients with OAG (n = 296), the predictive accuracy for 36 months in the future was higher for the KF models (73.5% for KF-TP, 71.2% for KF-TPO) than for the LR models (57.5%, 58.0%). Predictive accuracy did not differ significantly between KF-TP and KF-TPO (P = 0.20). If the races of the training and testing set patients were aligned (versus nonaligned), the mean absolute prediction error of future MD improved 0.39 dB for KF-TP and 0.48 dB for KF-TPO. Conclusions: Adding global RNFL data to existing KFs minimally improved their predictive accuracy. Although KFs attained better predictive accuracy when the races of the training and testing sets were aligned, these improvements were modest. These findings will help to guide implementation of KFs in clinical practice.