Superresolution Reconstruction for Magnetic Resonance Spectroscopic Imaging Exploiting Low-Rank Spatio-Spectral Structure

Magnetic resonance spectroscopic imaging (MRSI) is a rapidly developing medical imaging modality, capable of conferring both spatial and spectral information content, and has become a powerful clinical tool. The ability to non-invasively observe spatial maps of metabolite concentrations, for instance, in the human brain, can offer functional, as well as pathological insights, perhaps even before structural aberrations or behavioral symptoms are evinced. Despite its lofty clinical prospects, MRSI has traditionally remained encumbered by a number of practical limitations. Of primary concern are the vastly reduced concentrations of tissue metabolites when compared to that of water, which forms the basis for conventional MR imaging. Moreover, the protracted exam durations required by MRSI routinely approach the limits for patient compliance. Taken in conjunction, the above considerations effectively circumscribe the data collection process, ultimately translating to coarse image resolutions that are of diminished clinical utility. Such shortcomings are compounded by spectral contamination artifacts due to the system pointspread function, which arise as a natural consequence when reconstructing non-band-limited data by the inverse Fourier transform. These artifacts are especially pronounced near regions characterized by substantial discrepancies in signal intensity, for example, the interface between normal brain and adipose tissue, whereby the metabolite signals are inundated by the dominant lipid resonances. In recent years, concerted efforts have been made to develop alternative, non-Fourier MRSI reconstruction strategies that aim to surmount the aforementioned limitations. In this dissertation, we build upon the burgeoning medley of innovative and promising techniques, proffering a novel superresolution reconstruction framework predicated on the recent interest in low-rank signal modeling, along with state-of-the-art regularization methods. The proposed framework is founded upon a number of key tenets. Firstly, we proclaim that the underlying spatio-spectral distribution of the investigated object admits a bilinear representation, whereby spatial and spectral signal components can be effectively segregated. We further maintain that the dimensionality of the subspace spanned by the components is, in principle, bounded by a modest number of observable metabolites. Secondly, we assume that local susceptibility effects represent the primary sources of signal corruption that tend to disallow such representations. Finally, we assert that the spatial components belong to a class of real-valued, non-negative, and piecewise linear functions, compelled in part through the use of a total variation regularization penalty. After demonstrating superior spatial and spectral localization properties in both numerical and physical phantom data when compared against standard Fourier methods, we proceed to evaluate reconstruction performance in typical in vivo settin