Comparison of Black Box Implementations of Two Algorithms of Processing of NMR Spectra, Gaussian Mixture Model and Singular Value Decomposition

摘要

Analysis of NMR spectra is a multi-stage computational process performed with the use of appropriately chosen sequence of algorithms. Initial stages of this process, called pre-processing, including filtering, baseline correction, phase correction and removal of unwanted components, are aimed at improving the quality of NMR spectral signal by rejection of noise, removing unnecessary spectral components and irregularities. After pre-processing the basic operations on NMR spectra are aimed at estimation of levels of certain metabolites by analysis of appropriate structural properties of NMR spectral signals. In this paper authors present design and implementation of two signals modelling methods. The first one is based on singular value decomposition of the induction decay signal. The second is done with use of mixture model constructed for frequency spectrum. Authors present all assumption that need to be satisfied and processing steps that must be performed before final analysis. The methods studied in the paper are implemented under the black - box assumption; i.e., prior knowledge of parameters of metabolites in the spectra is not used. As a second part of the project authors present a comparison of obtained result with popular modelling techniques and software LCmodel and Tarquin, based on experimental phantom dataset. Comparisons between different methods are based on the commonly used quality indexes, mean squared errors corresponding to levels of detected metabolites and specificities and sensitivities of the process of detection of metabolites. Using the presented comparisons we authors are able to characterize advantages and drawbacks of the studied approaches.