Figure 3.

General scheme of the processing pipeline: each 2D slice from a set of 3D scans is processed by the network to obtain four prediction masks (one for each segmentation class), from which 3 area borders are calculated: posterior cortical vitreous (PCV), inner limiting membrane (ILM), and retina pigment epithelium (RPE). Predicted segmentations $\widehat{I}$ are compared with the ground truth data I, and the difference between them (in the form of loss) is used to update the network weights. During the test phase, error metrics of Dice score, Mean Absolute Error (MAE), and Topology Incorrectness Index (TII) are computed.