OCT Signal Enhancement with Deep Learning.

PURPOSE: To establish whether deep learning methods are able to improve the signal-to-noise ratio of time-domain (TD) OCT images to approach that of spectral-domain (SD) OCT images. DESIGN: Method agreement study and progression detection in a randomized, double-masked, placebo-controlled, multicenter trial for open-angle glaucoma (OAG), the United Kingdom Glaucoma Treatment Study (UKGTS). PARTICIPANTS: The training and validation cohort comprised 77 stable OAG participants with TD OCT and SD OCT imaging at up to 11 visits within 3 months. The testing cohort comprised 284 newly diagnosed OAG patients with TD OCT images from a cohort of 516 recruited at 10 United Kingdom centers between 2007 and 2010. METHODS: An ensemble of generative adversarial networks (GANs) was trained on TD OCT and SD OCT image pairs from the training dataset and applied to TD OCT images from the testing dataset. Time-domain OCT images were converted to synthesized SD OCT images and segmented via Bayesian fusion on the output of the GANs. MAIN OUTCOME MEASURES: Bland-Altman analysis assessed agreement between TD OCT and synthesized SD OCT average retinal nerve fiber layer thickness (RNFLT) measurements and the SD OCT RNFLT. Analysis of the distribution of the rates of RNFLT change in TD OCT and synthesized SD OCT in the two treatment arms of the UKGTS was compared. A Cox model for predictors of time-to-incident visual field (VF) progression was computed with the TD OCT and the synthesized SD OCT images. RESULTS: The 95% limits of agreement were between TD OCT and SD OCT were 26.64 to -22.95; between synthesized SD OCT and SD OCT were 8.11 to -6.73; and between SD OCT and SD OCT were 4.16 to -4.04. The mean difference in the rate of RNFLT change between UKGTS treatment and placebo arms with TD OCT was 0.24 (P = 0.11) and with synthesized SD OCT was 0.43 (P = 0.0017). The hazard ratio for RNFLT slope in Cox regression modeling for time to incident VF progression was 1.09 (95% confidence interval