Differentiating Glaucomatous Optic Neuropathy from Non-Glaucomatous Optic Neuropathies Using Deep Learning Algorithms

Abstract

Purpose : A deep learning framework to differentiate glaucomatous optic disc changes (GON) from non-glaucomatous optic neuropathy-related disc changes (NGON). Design : Cross-sectional study. Method : A deep-learning system was trained, validated, and externally tested to classify optic discs as normal, GON, or NGON using 2,183 digital color fundus photographs. A Single-Center data set of 1,822 images–660 images of NGON, 676 images of GON, and 486 images of normal optic discs–was used for training and validation, whereas 361 photographs from four different data sets were used for external testing. Our algorithm removed the redundant information from the images using an optic disc segmentation (OD-SEG) network, following which we performed transfer learning with various pre-trained networks. Finally, we calculated sensitivity, specificity, F1-score, and precision to show the performance of the discrimination network in the validation and independent external data set. Results : For classification, the algorithm with the best performance for the Single-Center data set was DenseNet121, with a sensitivity of 95.36%, precision of 95.35%, specificity of 92.19%, and F1 score of 95.40%. For the external validation data, the sensitivity and specificity of our network for differentiating GON from NGON were 85.53% and 89.02%, respectively. The glaucoma specialist who diagnosed those cases in masked fashion, had a sensitivity of 71.05% and a specificity of 82.21%. Conclusions : The proposed algorithm for the differentiation of GON from NGON yields results that have a higher sensitivity than those of a glaucoma specialist, and its application for unseen data thus is extremely promising.