Flow-chart of the proposed two-level deep learning pipeline. In each step, an auto-encoder learns to encode the input data in a lower-dimensional embedding. First, the local encoder transforms each A-Scan into a 20-dimensional local representation, resulting in 20 2D feature maps. This local representation forms the input of the second stage, the global encoder. The global features provide a compact representation of an entire three-dimensional dataset in only 20 numbers.