Fig. 2.

Schematic representation of the training procedure and overview of both U-Nets in the developed cascade architecture. Transposed convolution kernel and stride sizes in the decoder are equal to the stride size in the identical encoder resolution depth, highlighted in red in the architecture diagram. Feature map depths at each resolution are displayed above the convolution blocks and are capped at 320 regardless of the number of encoder-decoder stages. The output of the first 3D U-Net is upsampled and combined with the full resolution PET and CT images as an input to the second 3D U-Net