Chen et al. extend the idea of residual networks to volumetric convolutional neural networks (i.e. 3D CNNs) for brain segmentation. Their main contribution is twofold: first, they present a volumetric residual module allowing to design deeper networks; second, they propose an auto-context network for brain segmentation. In the following, I will focus on the volumetric residual module.
The proposed volumetric residual unit is illustrated in Figure 1b). The network architecture illustrated in Figure 1a) is called VoxResNet and consists of multiple volumetric residual modules overall heavily increasing the depth compared to other 3D convolutional neural networks. As can be seen in Figure 2, this setup is very similar to the original residual unit as presented in . However, Chen et al. Additional introduce batch normalization between the two weight layers.