ESTRO38 Congress Report

corrected CBCT projections or images

raw CBCT projections or images

Convert block

+

512 × 512 × 1

512 × 512 × 1

Mix block Convolution 1x1

512 × 512 × 1

Res block

Mix block Up block

512 × 512 × 8

Down block

Mix block Up block

256 × 256 × 16

Down block

Mix block Up block

128 × 128 × 32

Down block

Mix block Up block

64 × 64 × 64

Down block

32 × 32 × 128

Down block Up block

16 × 16 × 256

Figure 1 U-net architecture consisting of contracting and expansive paths. The dimensions at each depth are indicated on the figure, with the third dimension representing channels. Input and outputs are two dimensional CBCT projections or image slices.

(a)

(b)

(c)

(d) (f) Figure 2 Opposed pencil-beam-scanning, single-field-uniform-dose proton dose distribution on the (a) reference image and (b) the image corrected by the U-net. (c) shows the dose difference. (d-f) The same for volumetric arc therapy photon doses. Unet3 stands for training on reconstructed images. (e)

PHYSICS | Congress report

20