Scatter correction in cone-beam computed tomography using convolutional neural networks.

A validated Monte Carlo model of the kV imaging system of the TrueBeam STx Linac based on the egs_cbct code of EGSnrc has been used to determine the scatter contribution in cone-beam computed tomography (CBCT) images. Geometrical and anatomical digital phantoms were used in this MC framework to acquire a data set of CBCT images and its respective scatter components. This data set was used to train a deep learning algorithm, based on Convolutional Neural Networks. Specifically, we analyzed the performance of UNets and Generative Adversarial Networks. UNets demonstrate more efficient and precise results, even when trained with a few hundred images. Scatter corrections in CBCT images can be achieved with a trained deep learning-based model in 3 to 4 orders of magnitude faster than MC-based methods. [ABSTRACT FROM AUTHOR]