Fig. 1.

The proposed 3D-CNN U-net: A patch-wise approach with a patch size of 160 × 160 × 160 cubic voxels of 1-mm edge length was used. A fully convolutional encoder-decoder architecture with 3D convolutions, residual-block-connections, and four reductions of the feature map size was employed. The network was trained for simultaneous segmentation of the left and right thalamus, remaining deep gray matter structures, and remaining total intracranial volume