Deep learning for computed tomography reconstruction - learned methods, deep image prior and uncertainty estimation

Abstract

X-ray computed tomography (CT) is a highly relevant imaging technique with clinical and industrial applications. At its core, CT involves an image reconstruction task from detector measurements that are acquired from multiple projection angles. Improving CT reconstruction using deep learning, which is being explored and utilized in various fields, is a subject of recent and current research. This thesis comprises six papers, whose contributions can be summarized as two-fold. First, several deep learning approaches are compared quantitatively and qualitatively, involving the creation of a benchmark dataset as well as the realization and evaluation of challenges for learned low-dose and sparse-view CT reconstruction. Second, several extensions of the deep image prior (DIP)—an unsupervised deep learning image reconstruction framework—are investigated. This includes its application to CT using total-variation regularization, pretraining on synthetically generated data, and uncertainty estimation via a probabilistic model. These extensions benefit DIP-based CT reconstruction in several ways, such as an improved reconstruction quality, an accelerated reconstruction process, and the identification of potential errors in the reconstruction. Additionally, a Bayesian experimental design approach utilizing the uncertainty estimation is studied for the selection of scanning angles based on a pilot scan. Complementing the papers, which are included without any modifications in the second part of this thesis, the first part introduces relevant foundations, as well as a large overview of literature on deep learning for CT reconstruction.