Session Item

In recent years, the neural network (NN) has emerged as a promising tool to enhance Cone Beam Computed Tomography (CBCT) images. Among them, NNs based on the cycle consistent generative adversarial network (CycleGAN) architecture have been shown to suppress imaging artifacts in CBCT images. Such a NN is applied to a CBCT image, resulting in a synthetic CT (sCT), with processing times of a few seconds. These sCTs should allow accurate dose calculations and (re-)contouring, which is an important step toward online adaptive radiotherapy with conventional linacs.
A common concern with NNs is generalizability. When a trained NN is deployed, it is hard to predict how it will perform on out-of-distribution data, e.g. acquired in a different institute or with a different machine. This study quantifies the performance of a NN, trained on CBCTs acquired on machines without anti-scatter grid (ASG), when applied to CBCTs acquired on machines with ASG.

The NN was a pretrained, CycleGAN-based model, deployed in the Advanced Medical Image Registration Engine (ADMIRE, Elekta) v3.34. It was trained on CBCTs of the male pelvis region, acquired without ASG, downsampled to 1.64x1.64mm in-plane voxel size. The testing dataset contained 895 CBCTs (Elekta Versa HD) from 47 prostate cancer patients, each with a planning CT (Siemens Somatom go.Open Pro). Of these CBCTs, 336 and 559 were acquired with and without ASG respectively.
For each CBCT, a sCT was generated by the NN and two reference CTs were generated by deformably registering the planning CT to the CBCT and sCT respectively. The CBCT and sCT were compared to their reference CT with the structural similarity (SSIM) index, mutual information (MI), mean error (ME), mean absolute error (MAE) and the absolute percentage difference between their histograms (histogram error). The sharpness of the images was evaluated as the full width at half maximum (FWHM) value of the compact bone of the pelvis.
Statistical significance of the results was evaluated using the Wilcoxon rank-sum test for clustered data.

Examples of CBCTs, sCTs and a planning CT are shown in figure 1. Calculated metrics are summarized in table 1.
The sCTs outperformed their corresponding CBCTs on all metrics, except FWHM (P<.001 for all metrics). This shows that the NN improves correspondence with the reference CT, but loses some sharpness with respect to the CBCT.
The CBCTs with ASG outperformed the CBCTs without, except in FWHM (P<.05 for all metrics). The sCTs with ASG outperformed the sCTs without, except in FWHM (P<.05 for all metrics except MAE).

The NN in Admire reliably enhanced the CBCT images. We observed some loss of sharpness, which may be explained by the downsampling of the training data. The NN benefitted from the higher quality in CBCTs with ASG, even though it was trained exclusively on data without ASG. We expect, however, that the differences between sCTs with/without ASG may not be dosimetrically relevant.