General Time-resolved Computed Tomography from First Principles

Abstract

Conventional computed tomography (CT) methods assume the sample being scanned is static. They are insufficient for dynamic systems where the sample is evolving during scanning, and typically lead to artifacts in CT reconstruction. Dynamic CT methods aim to address this issue by capturing the dynamics to reduce artifacts during reconstruction and provide time-resolved imaging. However, current dynamic CT methods are often limited to specific applications or single classes of dynamics. Moreover, these methods are commonly tested on two or fewer experiments, which poses a risk of overfitting the developed methods to these limited examples. In this thesis, I aim to address these issues in dynamic CT. First, I develop and release pt4 (PhanTom-4d), an open-source software tool used to generate time-evolving phantoms (simulated test samples) for testing dynamic CT methods across a wide range of dynamics. Second, I devise a general dynamic CT method. To achieve this, I categorise general dynamics into either kinematic or non-kinematic dynamics, allowing for distinct assumptions to be placed on each. Kinematic dynamics are assumed to exhibit slow and slowly varying motion, while non-kinematic dynamics are assumed to be sparse. I employ a deformation vector field (DVF) to parameterise the kinematic dynamics and sparse arrays of patches to parameterise the non-kinematic dynamics. Subsequently, I develop a methodology to estimate my dynamics model from a time series of groundtruth volumes. My dynamics model is validated on phantoms generated using pt4 and real-world tomographic data. I demonstrate that my dynamics model faithfully captures dynamics only requiring 0.2 % to 1 % of the data for an equivalent timeseries of volumes. Finally, I derive a CT reconstruction algorithm tailored to my dynamics model and assess the reconstruction performance given an estimated dynamics model. My findings reveal that my techniques yields up to a 33 % improvement in root-mean-square error (RMSE) over static techniques. To end, I present a proof-of-concept demonstrating the feasibility of estimating the dynamics model from measured projection data from a simulated CT scan, thereby illustrating the potential applicability of my method to real-world datasets.