Assessment of statistical and neural networks methods in NMR spectral classification and metabolite selection

摘要: Magnetic resonance spectroscopy opens a window into the biochemistry of living tissue. However, spectra acquired from different tissue types in vivo or vitro and body fluids contain large number peaks range metabolites, whose relative intensities vary substantially complicated ways even between successive samples same category. The realization full clinical potential NMR relies, part, on our ability to interpret quantify role individual metabolites characterizing specific conditions. This paper addresses problem classification by analysing using statistical neural network methods. It assesses performance models methods compares them with artificial models. also consistency selecting, directly spectra, subsets most relevant for differentiating types. analysis techniques are examined eight classes normal tumours obtained rats. We show that, given data set, linear non-linear is comparable, possibly due small sample size per class. that subset selected discriminant further networks improves accuracy, reduces necessary correct classification.