Machine learning for recovering spectral signatures of disease

In vivo single voxel metabolic magnetic resonance spectra (MRS) and its multi-voxel imaging counterpart, magnetic resonance spectroscopic imaging (MRSI), have emerged in recent years as powerful noninvasive techniques in medical imaging. The high-dimensionality and complexity of the MRSI challenges most traditional techniques in nuclear magnetic resonance (NMR) community, and significantly limit its wide clinical availability.; This thesis designs, develops, rigorously characterizes and demonstrates a machine learning based computer aided diagnosis (CAD) system, which recovers biochemically meaningful and physically interpretable spectral signatures of a variety of neurological disorders. The system holds promise for improving the sensitivity and specificity of the diagnosis of neurological disorders by integrating metabolic information from MRSI with structural information from MRI, thus ultimately increase therapeutic opportunities by providing an early and accurate diagnosis.; Considering that each MRSI dataset consists of a mixture of spatial (concentration distribution) and spectral information (metabolite resonances and tissue type), this thesis proposes a fast unsupervised multivariate analysis approach that simultaneously considers the spatial and spectral patterns of the entire high-dimensional dataset, taking advantage of the intrinsic relationships among these patterns to guide and improve the quality of analysis by minimizing the partial volume averaging problem. By adopting physically realistic constraints such as positivity, the approach recovers biochemically meaningful and physically interpretable spectral patterns which are significant for biomarker identification and further analysis. The computational efficiency of the approach makes it well-suited for use in a real-time diagnostic work-up.; Both unsupervised and supervised classification methods are integrated into the system to summarize the recovered patterns into disease specific images which can be presented to clinicians in a way that allows efficient and reliable assessment of the diagnostic information. Thus the system would greatly contribute to enhancing the practical utility and widespread clinical use of MRSI.; This thesis also demonstrates that the proposed novel approaches are not only useful in the analysis of human brain MRSI, but also have general biomedical application for identification of biomarkers from spectral data, such as in NMR-based metabolomics and assessment of age-related macular degeneration (AMD) given ophthalmologic images.