Abstract
PURPOSE: To develop and evaluate methods to improve the generalizability of convolutional neural networks (CNNs) trained to detect glaucoma from optical coherence tomography retinal nerve fiber layer probability maps, as well as optical coherence tomography circumpapillary disc (circle) b-scans, and to explore impact of reference standard (RS) on CNN accuracy. METHODS: CNNs previously optimized for glaucoma detection from retinal nerve fiber layer probability maps, and newly developed CNNs adapted for glaucoma detection from optical coherence tomography b-scans, were evaluated on an unseen dataset (i.e., data collected at a different site). Multiple techniques were used to enhance CNN generalizability, including augmenting the training dataset, using multimodal input, and training with confidently rated images. Model performance was evaluated with different RS. RESULTS: Training with data augmentation and training on confident images enhanced the accuracy of the CNNs for glaucoma detection on a new dataset by 5% to 9%. CNN performance was optimal when a similar RS was used to establish labels both for the training and the testing sets. However, interestingly, the CNNs described here were robust to variation in the RS. CONCLUSIONS: CNN generalizability can be improved with data augmentation, multiple input image modalities, and training on images with confident ratings. CNNs trained and tested with the same RS achieved best accuracy, suggesting that choosing a thorough and consistent RS for training and testing improves generalization to new datasets. TRANSLATIONAL RELEVANCE: Strategies for enhancing CNN generalizability and for choosing optimal RS should be standard practice for CNNs before their deployment for glaucoma detection.