Robust methodology for the discrimination of brain tumours from in vivo magnetic resonance spectra

Magnetic resonance spectroscopy (MRS) is a non-invasive technique representing a biochemical fingerprint of tissue composition. However, signal variation and noise cause considerable mixing between tissue categories, making class assignments unreliable. Noise filtering is investigated by benchmarking independent component analysis against univariate T² tests as pre-filters for variable selection, followed by an evaluation of the predictive power of the resulting models. We examine different discrimination strategies of variable selection and classifier validation and propose a robust methodology for the discrimination of 5 types and grades of brain tumours and cysts from 98 in vivo PROBE spectra. We show that use of the bootstrap technique for the selection of subsets of predictor spectral frequencies, and for estimating bias-corrected misclassification errors, gives more reliable and robust results. Bayesian estimates of the class conditional probabilities are obtained with linear discriminants and neural network models, and the optimal classifier structure is found to consist of coupled pairwise models. The bias-corrected overall classification rate achieved in this study is 73%