Robust estimation of diffusion tensor and diffusion kurtosis imaging parameters

Date: 23 October 2018

Venue: Campus Middelheim, G0.10 - Middelheimlaan 1 - 2020 Antwerpen (route: UAntwerpen, Campus Middelheim)

Time: 4:15 PM

Organization / co-organization: Department of Physics

PhD candidate: Quinten Collier

Principal investigator: Jan Sijbers & Jelle Veraart

Short description: PhD defence Quinten Collier - Faculty of Science, Department of Physics


Diffusion magnetic resonance imaging (dMRI) is primarily a medical imaging modality that allows the in vivo and non-invasive quantification of the diffusion of water molecules. This unique ability makes dMRI a valuable tool, both in preclinical research and in the clinical practice, as it is able to provide information about the microstructure of living tissue. The images acquired with dMRI are often combined in various diffusion models that produce parameters which can serve as biomarkers for various pathological changes. The work in this thesis focuses on estimating model parameters for two specific diffusion models, diffusion tensor imaging (DTI) and diffusion kurtosis imaging (DKI), in a robust way, without compromising the precision or accuracy of the parameter estimates.

In this context, two distinct aspects of robustness are considered. First, the iterative reweighted linear least squares (IRLLS) framework is proposed, with the goal of estimating diffusion model parameters in the presence of data outliers. These outliers are often the result of various imaging artifacts and can, when left untreated, severely influence the parameter estimates to such an extent that the data are no longer useful. IRLLS automatically identifies and removes these outliers, thereby providing robust DTI or DKI model parameter estimates. Secondly, clinical dMRI typically deals with a relatively low spatial resolution, causing voxels to often contain multiple different substances. For dMRI, these so-called partial volume effects can be especially problematic when a voxel contains both tissue and free diffusing water. Therefore, the DKI-FWE model is proposed with the goal of separating the signal contributions of free diffusing water and tissue. To deal with the ill-conditionedness of the DKI-FWE model, two distinct solutions are proposed. The first one deals with this ill-conditionedness on the parameter estimation level, resulting in the use of a Bayesian parameter estimation approach with a shrinkage prior (BSP). Alternatively, one could also acquire a richer dMRI data set that includes multiple echo times. Incorporating this additional information in the model, which is now called the T2-DKI-FWE model, leads to an increased stability of the parameter estimation problem.