Leveraging prior knowledge for learning-based approaches to computed tomography

Abstract

Computed Tomography (CT) is a widely used imaging technique in medical, industrial, and scientific applications. CT reconstruction is a classical inverse problem in which the internal structure of an object is recovered from X-ray measurements. In practical settings, these measurements are often sparse, noisy, or otherwise incomplete, making accurate reconstruction challenging. Prior knowledge about the underlying object can therefore play an essential role in reducing ambiguity and guiding the reconstruction towards plausible solutions.

This thesis explores different ways of acquiring and incorporating prior knowledge into learning-based approaches for CT reconstruction. First, we investigate supervised learning, where prior knowledge is learned explicitly from paired training data. We introduce a low-dose industrial CT dataset and study the use of depth information for CT denoising. We further investigate how supervised deep learning can be integrated at different stages of the CT processing pipeline to efficiently reduce artifacts. Second, we explore self-supervised learning, where prior knowledge is extracted directly from the CT data itself. By exploiting similarities between different orientations of CT images, we develop methods for improving anisotropic image resolution without requiring external high-resolution reference data. Third, we investigate joint reconstruction of multiple similar objects using Implicit Neural Representations (INRs), combining their implicit regularization with statistical information shared across objects to improve individual reconstructions. Finally, we study generative models as expressive learned priors and systematically benchmark diffusion-based methods for CT reconstruction under both simulated and real-world experimental conditions.

Together, these contributions demonstrate how different forms of prior knowledge can complement incomplete or noisy measurements in CT reconstruction. They also highlight the trade-offs between the strength of learned priors, data consistency, computational efficiency, and the availability of training data. By examining prior knowledge from multiple perspectives, this thesis provides insights into the opportunities and practical limitations of learning-based approaches for solving challenging CT reconstruction problems.