Magnetic resonance (MR) spectroscopy provides a direct non-invasive measure of tissue biochemistry, but tissue heterogeneity causes considerable mixing between tissue categories. A systematic methodology for variable selection and performance estimation, applied to 98 in vivo spectra from cysts and five categories of brain tumour is proposed. The selection of predictive variables from the spectra, and the estimation of misclassification errors, are made robust by pre-filtering the irrelevant spectral components and repeatedly applying bootstrap resampling. Three alternative approaches to the methodology were investigated, with reference to pairwise discriminant models. The first approach is applied directly to the spectral intensity values, treated as independent covariates that are interpreted as metabolite indicators, proceeding to search for the smallest number of metabolites necessary for class discrimination. The two other approaches use independent component analysis (ICA) to separate the heterogeneous spectra into a small number of independent spectral sources of intrinsic tissue types. Given the six classes with strong inter-class mixing, the most accurate classifier based on linear discriminant models is obtained by first optimizing the discrimination between class pairs, then combining their outcome using a pairwise coupling method. Finally, the statistical and ICA pre-processing methods are compared in a retrospective study for the first class assignment pair, to separate low- and medium-grade from high-grade astrocytic tumours.
|Journal||IEE Proceedings: Science, Measurement and Technology|
|Publication status||Published - 1 Nov 2000|