Fig. 1. Deep learning segments clinical features on par with human experts from retinal OCT images.

a F1 scores for 11 clinical features segmented on a test set from 37 to the algorithm previously unseen patients. b Example of OCT images selected to illustrate various segmentation classes with ground truth and predicted segmentation maps. c F1 scores for subretinal hyperreflective material, fibrovascular PED, drusenoid PED as well as fibrosis on 30 challenging test patients containing these features. d Example OCT images with consensus ground truth, annotations from three different annotators as well as predicted segmentation maps displaying segmentation of multiple features. Yellowish color was not evaluated and either stands for a crop of the image or vitreous cavity.