Figure 1.

Our proposed model architecture. The Initialization block receives the input data and creates a feature space. This is handed over to the Refinement block which is built from multiple dilation blocks, to integrate global context into the feature space. The Self-Attention block scales features in our feature space with access to the global context and returns 64 output channels. The Layer Head transforms 2D feature maps into a 1D representation where in our case every 1D output channel represents the height of an OCT layer. Convolutions are indicated as C(kernel-size, filters) and dilation blocks as DB(filters, dilation-1, dilation-2).