Figure 1

Diagram of the anisotropic U-Net (aU-Net) introduced by Chlebus et al.16. All convolutional blocks except for the last one are followed by batch normalization and ReLU activation function. Numbers in brackets indicate the number of features generated by each layer. All convolutions are unpadded, so that cropping of the feature map centers is required before concatenation. Downsampling is performed by max pooling, upsampling by transposed convolutions.