From Machine to Machine: An OCT-trained Deep Learning Algorithm for Objective Quantification of Glaucomatous Damage in Fundus Photographs

PURPOSE Previous approaches using deep learning (DL) algorithms to classify glaucomatous damage on fundus photographs have been limited by the requirement for human labeling of a reference training set. We propose a new approach using quantitative spectral-domain (SD) OCT data to train a DL algorithm to quantify glaucomatous structural damage on optic disc photographs. DESIGN Cross-sectional study. PARTICIPANTS A total of 32 820 pairs of optic disc photographs and SD OCT retinal nerve fiber layer (RNFL) scans from 2312 eyes of 1198 participants. METHODS The sample was divided randomly into validation plus training (80%) and test (20%) sets, with randomization performed at the patient level. A DL convolutional neural network was trained to assess optic disc photographs and predict SD OCT average RNFL thickness. MAIN OUTCOME MEASURES The DL algorithm performance was evaluated in the test sample by evaluating correlation and agreement between the predictions and actual SD OCT measurements. We also assessed the ability to discriminate eyes with glaucomatous visual field loss from healthy eyes with area under the receiver operating characteristic (ROC) curves. RESULTS The mean prediction of average RNFL thickness from all 6292 optic disc photographs in the test set was 83.3±14.5 μm, whereas the mean average RNFL thickness from all corresponding SD OCT scans was 82.5±16.8 μm (P = 0.164). There was a very strong correlation between predicted and observed RNFL thickness values (Pearson r = 0.832; R2 = 69.3%; P < 0.001), with mean absolute error of the predictions of 7.39 μm. The area under the ROC curves for discriminating glaucomatous from healthy eyes with the DL predictions and actual SD OCT average RNFL thickness measurements were 0.944 (95% confidence interval [CI], 0.912-0.966) and 0.940 (95% CI, 0.902-0.966), respectively (P = 0.724). CONCLUSIONS We introduced a novel DL approach to assess fundus photographs and provide quantitative information about the amount of neural damage that can be used to diagnose and stage glaucoma. In addition, training neural networks to predict SD OCT data objectively represents a new approach that overcomes limitations of human labeling and could be useful in other areas of ophthalmology.