Chen et al. propose to apply residual units [1,2] to segmentation of brain scans. As the scans represent volumetric information, a 3D convolutional neural network is used. The network is summarized in Figure 1.
Additionally, Chen et al. use the proposed architecture in an auto-context fashion. This means that one VoxResNet is trained on the training set. Based on the probability maps produced by this VoxResNets, another VoxResNet is trained taking these probability maps as additional "context"-input.