Fig. 2.

The fully convolutional neural network architecture (a) comprises a number of building blocks, referred to as bottlenecks (b). An input 256 × 256 image undergoes a series of convolutions (Conv), nonlinear rectifications (ReLU), and batch normalizations (Norm). Down-sampling (↓) and up-sampling (↑) of the processed images are applied in the contracting and expansion paths, respectively. The l, k, m, and n values in (b) are determined by the image size and number of channels at the input and output of each bottleneck as shown in (a)