Fig. 1.

Convolutional Neural Network Structure (modified U‐Net, adapted from Ronneberger et al. [15]). The input CT slice is down‐sampled due to GPU memory limitations. The downward path is the VGG16 model from keras trained on ImageNet with locked weights. The upward path mirrors the VGG16 path with some modifications to enable faster convergence. Activation functions not shown for clarity. The output of then neural network is then up‐sampled for the full resolution prediction. Up‐sampling and down‐sampling is done with 2:1 nearest neighbor resizing (both for pre/post‐processing and within the GPU neural network flow).