Robust and time-effcient determination of perfusion parameters using time-encoded Arterial Spin Labeling MRI

In clinical routine, arterial spin labeling (ASL) faces many challenges, such as time pressure, patient- and disease-specific artifacts, e.g., in steno-occlusive and Moya-Moya disease. In addition, individually tailored parametrization of the MR pulse-sequence is frequently required. Time-encoded ASL-techniques like Hadamard time-encoded pseudocontinuous ASL (H-pCASL) offers a time and signal efficient way to measure accurately both perfusion and arterial transit-times. However, it relies on the decoding of a series of volumes. If even a single volume is corrupted this might, via the decoding process, lead to artifacts in the entire dataset and in the worst case result in the loss of the data. In this thesis a general introduction to time encoded ASL is given and three methods are introduced to increase the robustness of time-encoded ASL against image artifacts and to detect corrupted images. The first method is Walsh-ordered time-encoded H-pCASL (WH-pCASL). It proposes the Walsh-ordering of Hadamard encoding-matrices. In contrast to conventional H-pCASL, this makes perfusion-weighted images accessible during a running experiment and even from incomplete sets of encoded images. An optional additional averaging strategy is based on a mirrored matrix and results in more perfusion-weighted images without any penalty in time. The feasibility of the method is shown using five volunteer datasets. As a second method non-decoded time-encoded ASL is introduced. This novel model-based approach to quantification avoids the decoding step altogether. It models the non-decoded time encoded signal. Therefore it uses the convolution of the tissue response function with a model of the true encoded arterial input function, which is determined by the employed encoding matrix. The model was implemented in a Bayesian model-based ASL analysis framework to fit maps for hemodynamic parameters. The feasibility of the method is demonstrated in a study with five volunteers. The last method is an algorithm for the automated detection of outliers and corrupted images, which is based on variational Bayesian inference (VB). Using the variance of the posterior normal distributions, the algorithm measures the quality of a fit directly and without the need for a separate reference dataset. Its performance and feasibility is demonstrated using volunteer data and a clinical dataset.