| dc.description.abstract | Magnetic resonance spectroscopic imaging (MRSI), combining both magnetic resonance spectroscopy (MRS) and magnetic resonance imaging (MRI) techniques, has proven to be a powerful approach to reveal information about metabolite distributions and discriminate multiple resonant frequencies in the spectrum. Due to its nondestructive nature and sensitivity to the molecular environment of individual atoms, MRSI is widely applied in the clinical community. The typical applications include mapping abnormal tissues in the brain, in the prostate, in the breast, as well as pathologic analysis after resection operations. All of these clinical studies require satisfactory resolution in both spatial and spectral dimensions, which in conventional MRSI demands a great deal of acquisition time. Unfortunately, patient discomfort, motion artifacts and cost will significantly increase when the acquisition time lengthens. To overcome these problems, we propose to implement an imaging protocol that only acquires an optimal subset of data to accelerate the collection process without sacrificing spatial and spectral accuracy in reconstruction.
In most MRSI applications, the spectral domain has great sparsity, which raises the possibility of reconstructing spectral information with limited time series data. Therefore, an efficient sequential backward selection (SBS) technique is proposed to select a limited set of but the most informative echo-time values, which are then applied to echo-planar imaging (EPI) acquisition. By exploiting multi-echo EPI, multiple k-space frames can be acquired within one excitation to further reduce the acquisition time. To achieve this purpose, we modify the selection method to a more efficient approach. Instead of selecting echo-time value one by one, the modified algorithm selects multiple echo-time values simultaneously, which would then be used in one excitation acquisition.
For the EPI technique, a k-space frame cannot be collected instantly. In other words, every k-space sample will have a different time delay even in the same k-space frame. Consequently, selecting one echo-time value for a whole k-space frame might not be accurate enough. We then extend the data selection method to both k-space and time domains. In addition, an advanced EPI strategy is introduced. On the other hand, if SBS algorithm is the only restriction for k-t data selection, the acquisition efficiency might be reduced, which leads to longer observation procedure. Considering this, sequential k-t selection with constraint will be studied for a better selection result in a more efficient way.
Due to the selection method and the EPI acquisition technique, the collected data are normally nonuniform and time varying. Therefore, fast Fourier Transforms (FFTs) are not capable reconstructing the spatial and spectral information directly. Besides, an FFT cannot separate spatial information of different metabolite resonances. On the other side, conventional optimization methods, such as conjugate gradients (CG), require very high computational effort and large memory storage to find the matching parameters of the images. A fast reconstruction method combining polynomial approximation with FFTs is investigated, which can greatly accelerate the reconstruction process without sacrificing reconstruction quality.
During the reconstruction procedure, the finite data set raises a practical problem: frequency local minima, which means convergence to the global minimum is not guaranteed. In order to overcome this problem, we study the origin of the local minima, and propose an easily implemented method: varying estimated decays during the optimization. However, this method has its own limitations, especially in reconstruction efficiency and accuracy. Therefore, another advanced technique, applying weighted scalars to the cost function model, is presented. The results show that the second method can efficiently attenuate frequency local minima effect, while avoiding the reconstruction speed and accuracy problems. | en_US |
