Figure 2.

Simplified network architecture of FD-CNN. FD-CNN takes contour masks and beam templates as input and predicts nine field doses in an axial slice. The details of downsampling, upsampling, and convolutional blocks (rounded rectangles) are omitted to highlight the transformation process from the inputs to the output. PTV±n refers to the n th PTV slice superior or inferior to the query slice. OAR includes only the OAR contour in the query slice. Each rectangle block represents a layer with the number of channels on the top and image dimensions labeled on the bottom of each layer.