Date of Award

August 2018

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Engineering

First Advisor

Roshan M D'Souza

Second Advisor

Vitaliy L Rayz

Committee Members

David Saloner, Krishna M Pillai, Zeyun Yu

Keywords

4D flow MRI, Algorithm, brain aneurysm, computational fluid dynamics (CFD), patient-specific modeling

Abstract

Time-resolved three-dimensional spatial encoding combined with three-directional velocity-encoded phase contrast magnetic resonance imaging (termed as 4D flow MRI), can provide valuable information for diagnosis, treatment, and monitoring of vascular diseases. The accuracy of this technique, however, is limited by errors in flow estimation due to acquisition noise as well as systematic errors. Furthermore, available spatial resolution is limited to 1.5mm - 3mm and temporal resolution is limited to 30-40ms. This is often grossly inadequate to resolve flow details in small arteries, such as those in cerebral circulation. Recently, there have been efforts to address the limitations of the spatial and temporal resolution of MR flow imaging through the use of computational fluid dynamics (CFD). While CFD is capable of providing essentially unlimited spatial and temporal resolution, numerical results are very sensitive to errors in estimation of the flow boundary conditions. In this work, we present three novel techniques that combine CFD with 4D flow MRI measurements in order to address the resolution and noise issues. The first technique is a variant of the Kalman Filter state estimator called the Ensemble Kalman Filter (EnKF). In this technique, an ensemble of patient-specific CFD solutions are used to compute filter gains. These gains are then used in a predictor-corrector scheme to not only denoise the data but also increase its temporal and spatial resolution. The second technique is based on proper orthogonal decomposition and ridge regression (POD-rr). The POD method is typically used to generate reduced order models (ROMs) in closed control applications of large degree of freedom systems that result from discretization of governing partial differential equations (PDE). The POD-rr process results in a set of basis functions (vectors), that capture the local space of solutions of the PDE in question. In our application, the basis functions are generated from an ensemble of patient-specific CFD solutions whose boundary conditions are estimated from 4D flow MRI data. The CFD solution that should be most closely representing the actual flow is generated by projecting 4D flow MRI data onto the basis vectors followed by reconstruction in both MRI and CFD resolution. The rr algorithm was used for between resolution mapping. Despite the accuracy of using rr as the mapping step, due to manual adjustment of a coefficient in the algorithm we developed the third algorithm. In this step, the rr algorithm was substituded with a dynamic mode decomposition algorithm to preserve the robustness. These algorithms have been implemented and tested using a numerical model of the flow in a cerebral aneurysm. Solutions at time intervals corresponding to the 4D flow MRI temporal resolution were collected and downsampled to the spatial resolution of the imaging data. A simulated acquisition noise was then added in k-space. Finally, the simulated data affected by noise were used as an input to the merging algorithms. Rigorous comparison to state-of-the-art techniques were conducted to assess the accuracy and performance of the proposed method. The results provided denoised flow fields with less than 1\% overall error for different signal-to-noise ratios. At the end, a small cohort of three patients were corrected and the data were reconstructed using different methods, the wall shear stress (WSS) was calculated using different reconstructed data and the results were compared. As it has been shown in chapter 5, the calculated WSS using different methods results in mutual high and low shear stress regions, however, the exact value and patterns are significantly different.

Download

Share

COinS