Preview

PDF, English Download (24MB) | Terms of use

Abstract

In radiotherapy, ionizing radiation is used to accurately treat tumors. To spare healthy tissue the treatment plan is optimized on a computed tomography (CT) image. On-board cone-beam CT (CBCT) images cannot be used for a daily-updated plan because of their insufficient image quality. The aim of this thesis is to reduce two artifacts responsible for image quality loss: scatter and beam hardening. Novel deep learning-based methods are adapted and developed for the correction of projections, and are compared to correction methods of state-of-the art CBCT scanners. Here, a deterministic solver of the linear Boltzmann equation is used to generate training data for the deep scatter estimation (DSE). The proposed deep beam hardening corrections were designed to incorporate the contributions from bones as well as soft tissue, bringing an advantage to the commonly applied water precorrection method, which only considers a single material. DSE reduces the mean absolute error in test scans by approximately 96 %, outperforming the projection-based reference method in image quality, and is over 29 times faster than the reference in image domain. The proposed beam hardening correction significantly reduces the error in bone, with the remaining error reduced by 40 % in comparison to the water precorrection.