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

Images are an essential part in radiotherapy treatments, from the planning stage to adaptive treatment delivery. Cone Beam Computed Tomography (CBCT) plays a key role in Image Guided Radiation Therapy (IGRT). However, there is always a compromise between image quality and the additional dose delivered to patients during the imaging procedures. Over the last years several correction methods that account for most of the detrimental effects associated with scattered radiation have been developed, but they are normally difficult to implement and computationally expensive. In this work we used a validated EGSnrc Monte Carlo (MC) model of the kV and MV imaging systems of the Varian TrueBeam STx Linac as a ground truth to train a deep learning model, based on a U-Net Convolutional Neural Network (CNN). We generated projections of full tomographic acquisitions, separating the scatter and primary contributions in a single scan from a modified version of the egs_cbct code. Two training strategies have been analyzed: using total (primary+scatter) images to predict either primary or scatter. So far, the second method has produced better results. The CNN has shown successful performance on test data with speed-up factors of 3 to 4 orders of magnitude than equivalent MC based algorithms. On a Dell Precision T7920 workstation it takes 11.7 m/s projection to apply scatter corrections with the trained model, which amounts to 4.2 s for a complete 360 projections study. Comparison of profiles of the CNN predicted projections with the ground truth has shown a mean error of the order of approximately 0.5%. Further tests must still be carried out to verify the applicability of the CNN scatter model to experimental data. (author)