Spectrum separation for brain tumor magnetic resonance spectroscopy imaging

Altered metabolic activity in cancerous tissues leads to abnormal metabolite concentrations, which are reflected in abnormal spectral profiles in magnetic resonance spectroscopic imaging (MRSI). Currently, MRSI is utilized in conjunction with MRI to diagnose brain tumors. However, spectral profiles often show increased variability which limits the diagnostic potential of MRSI.;This thesis addresses this problem by quantifying the abundance (volume fraction) within a voxel for each tissue type instead of the conventional estimation of metabolite concentrations from spectral resonance peaks. This "spectrum separation" method simultaneously decomposes the observed spectra of multiple voxels into abundance distributions and constituent spectra.;Our results show that the proposed method can successfully extract spectral profiles consistent with different tissue types. We confirmed the physiological and clinical relevance of the extracted spectra by correlating the analysis results with pathologically proven tumors. The results demonstrate reduced cross-subject variability, which leads to improved discrimination between high and low-grade gliomas.;The accuracy of the estimated abundances was validated on phantom data. The results indicate that the method correctly decomposes the observed data into constituent spectra corresponding to the different solutions with their specific metabolite concentrations. It validated the interpretation of abundance estimates as partial volume fraction and established bias and confidence intervals for its estimates.;Furthermore, we explored the potential of NMF in characterizing tissue heterogeneity. Our results showed that if sufficient diversity is apparent, our method can decompose MRSI data into multiple constituent spectra to capture tissue heterogeneity by modifying the number of the underlying sources.;We also tested semi-nonnegative matrix factorization (semi-NMF), a generalized factorization algorithm that allows mixed data matrix, for spectrum separation. Our results indicate that once the data has sufficient quality, semi-NMF is better suited for MRSI data.;Taken together, this thesis validated the physiological and clinical relevance of components extracted from brain tumor MRSI indicating that spectral separation of MRSI via NMF algorithm is not simply a mathematical decomposition but is clinically and physiologically meaningful. These additional physiologically relevant components obtained from the data can provide doctors with more clinical information and thus improve the diagnostic effectiveness of MRSI.