FIGURE 1.

Network architecture of the proposed three-dimensional (3D) convolutional neural network (CNN). The network has 28 layers integrating six residual blocks. Bilinear interpolating arrows indicate upsampling operations to provide dense predictions for the segmentation task. Skip connections are used to fuse low- and high-level features in the network. Batch normalization is a linear transformation of the features to reduce covariance shift and accelerate the training process. The convolution bar represents the convolution operation that computes features. The number 64 indicates the number of channels in that layer, and 3 3 3 3 3 denotes the size of the 3D CNN kernels.