Speaker: Yi Wang. Advisor: Dennis L. Parker

Abstract:

This dissertation focuses on developing magnetic resonance imaging (MRI) acquisition strategies and algorithms for high-resolution magnetic resonance angiography (MRA) and quantitative image analysis. The problems of flow-dependent MRA with complete background suppression, flow-independent technique for visualizing blood vessels in the breast, and segmentation of multi-contrast breast MRI have been addressed.

An arterial spin labeling (ASL) based subtraction technique is presented in the first part of the dissertation, to achieve high-resolution brain MRA with complete background suppression. The imaging readout is further extended to a Look-Locker-like acquisition to capture the flow information encoded in the flow-sensitive signal. After the single in-plane slice-selective double inversion (IDOL) magnetization preparation, multiple turbo-FLASH (TFL) readouts are acquired with linear k-space ordering causing a signal variation that depends on through-slice flow velocity. This velocity measured in a hemodialyzer using an exponential curve fitting is shown to be consistent with the ground truth values.

In the second part of this dissertation, 3D two-point Dixon SSFP sequences with variations in the design of the readout gradients are broadly investigated for flow-independent breast MRA imaging. The theory is validated through simulations of the in-phase and out-of-phase signal for both fat and water signals. The experimental results show the potential of using this technique to visualize blood vessels in the breast. However, some of the image artifacts are still not fully understood. Future work will include reducing image artifacts, and comparing this two-point Dixon flow-independent technique to the contrast-enhanced method to see if all the vessels are detected.

Finally, a complete workflow composed of multi-parametric MRI inputs, pre-processing procedures, and the novel hierarchical support vector machine (SVM) algorithm is presented for breast MRI segmentation. The hierarchical SVM is compared to the conventional classification algorithms — conventional SVM and fuzzy C-mean (FCM). The classification outputs demonstrate that the presented methodology is consistent and outperforms other algorithms.