Fig. 1.

DenseVNet network architecture. First, 72³ feature maps are computed using a strided convolution. Second, a cascade of dense feature stacks and strided convolutions generate activation maps at three resolutions. Third, a convolution unit is applied at each resolution reducing the number of features. Fourth, after bilinear upsampling back to 72³, the maps are concatenated and a final convolution generates the likelihood logits. Finally, these are added to the upsampled spatial prior to generate the segmentation logit. Parameters for individual components are given in Table II.