Fig. 5.

Schematic of CNN ensemble made up of three end-to-end deep learning models (each separately fine-tuned on RNFL maps, shown as input at top) followed by dense fine-tuned layers which predict if the input image is glaucomatous (G) or not glaucomatous (NG). Predictions were averaged to arrive at the final ensemble prediction between 0 and 1, with 0.5 serving as threshold probability for binary classification.