Fig. 2.

Diagram of CNN topology. CNNs were constructed from multiple blocks in a U-net topology. Convolutional, maxpool, and upsampling layers in each block are depicted in yellow, red, and blue, respectively. Batch normalization and dropout layers were also included in each block, and a softmax layer was included in the segmentation output, but these are omitted from this figure for simplicity. Direct and skip connections between blocks are depicted as arrows. Dimensions were either halved or doubled between each block and are listed below each convolutional layer. For the 2d CNN, the initial dimensions were 256 × 256 voxels by 8 channels, and the U-net had 14 blocks. For the 3d CNN, the initial dimensions were 128 × 128 × 64 voxels by 16 channels, and the U-net had 10 blocks, to maintain a similar number of parameters. The input of the CNN was the CT scan, and the outputs were a segmentation and a diameter map