Fig. 2.

Architecture of our generative neural network. The generative neural network has a structure of layers similar to 3D U-Net [41]. It comprises about 1.7M parameters distributed in the five encoder blocks, five decoder blocks, five skip blocks, and one output block. The number of channels produced by the convolution in each block is shown on the top of that block. Each skip block yields additional four channels that are concatenated with the output channels of the convolution in the respective decoder block. The filter is of size 3x3x3 voxels in each convolutional layer from the encoder and decoder blocks, and 1x1x1 voxels from the skip blocks. Reflection padding strategy is applied for all convolutions. The downsampling layers perform decimation through a stride of 2x2x2 voxels, while the upsampling layers employ trilinear interpolations with a factor of 2. The output block consists of a convolutional layer with a filter of size 1x1x1 voxels, and a sigmoid layer that normalizes the network output.