2. The architecture of 3D U-Net (Çiçek et al., 2016), a type of fully convolutional network. It applies an end-to-end architecture using only valid convolutions (Conv) with no paddingand kernel sizes of 3 × 3 × 3. Rectified Linear units (ReLU) are used as activation functions. This results in a smaller output size than input size and requires cropping of when mappinglower level feature maps of the analysis path to the synthesis path of the network via concatenation (Concat). Max-pooling (Max pool) is used to reduce the resolution of feature maps,while up-convolutions (Up-conv) are used for up-sampling the feature maps back to higher resolutions. The number of extracted feature maps is noted above each layer. We show the inputand output size of feature maps at each level of the network. These parameters are kept constant for all experiments performed in this study. Batch normalization (BatchNorm) is usedthroughout the network for improved convergence (Ioffe and Szegedy, 2015).