Magnetic Resonance Spectroscopy (MRS) is a technique which has evolved rapidly over the past 15 years. It has been used specifically in the context of brain tumours and has shown very encouraging correlations between brain tumour type and spectral pattern. In vivo MRS enables the quantification of metabolite concentrations non-invasively, thereby avoiding serious risks to brain damage.
While Magnetic Resonance Imaging (MRI) is commonly used for identifying the location and size of brain tumours, MRS complements it with the potential to provide detailed chemical information about metabolites present in the brain tissue and enable an early detection of abnormality.
However, the introduction of MRS in clinical medicine has been difficult due to problems associated with the acquisition of in vivo MRS signals from living tissues at low magnetic fields acceptable for patients. The low signal-to-noise ratio makes accurate analysis of the metabolites represented by the spectra very difficult. Furthermore, for clinicians, it is time-consuming to analyze and interpret the MR spectra. This task requires considerable experience from an expert spectroscopist. Automated processing, analysis and interpretation of the MRS spectra is therefore highly valuable.
The high dimensionality of the spectra, the presence of noise and artefacts, and the low amount of data of specific pathologies (e.g. specific brain tumour types) available complicate the classification. In this thesis, advanced classification techniques based on Least Squares Support Vector Machines (LS-SVM) are developed and applied to brain tumour classification. LS-SVM classifiers using linear as well as nonlinear Radial Basis Function (RBF) kernels are compared with the classical techniques.
Given the problem of classifying four major types of brain tumours: glioblas-tomas, meningiomas, metastases and astrocytomas, LS-SVM classifiers perform very well regarding discrimination and prediction of these tumour types, except for the discrimination of glioblastomas vs. metastases, due to the high similarities among the available training spectra.
Dimensionality reduction, by selecting frequency regions or peak integration could possibly reduce the influence of disturbing noise and artefacts in the spectra. Using some prior knowledge for selecting resonance peaks allows to reach a similar performance as using complete spectra. Analogously, a rough quantification method using peak integration yields a relatively high classification performance.
Without the need of prior knowledge and/or dimensionality reduction, LS-SVM classifiers with linear or nonlinear kernels can be applied directly, and automatically give a reliable result. A good generalization performance could be expected from the maximal margin separation between the classes. |