Figure 3.

AUC for the proposed model on the subset of the Stanford test set that was graded by a glaucoma fellowship-trained ophthalmologist, with standard deviations computed over five runs of the model to plot the shaded areas. To assign a ground-truth label, human grader had access to other screening data, including fundus images, OCT RNFL and GCIPL printouts, IOP values, and visual field parameters, and also had access to patient history and physical examination data, while the model only had access to the OCT scan cube.