UCL Discovery
UCL home » Library Services » Electronic resources » UCL Discovery

Artificial neural networks in pattern recognition: Classification and analysis of proton NMR spectra of human brain tumours

El-Deredy, Wael; (1998) Artificial neural networks in pattern recognition: Classification and analysis of proton NMR spectra of human brain tumours. Doctoral thesis (Ph.D), UCL (University College London). Green open access

[thumbnail of Artificial_neural_networks_in_.pdf] Text

Download (12MB)


Nuclear magnetic resonance (NMR) spectroscopy is emerging as one of the most powerful tools for the study of living tissue biochemistry. Proton NMR spectroscopy provides access to a large range of metabolites, thus enabling the investigation of a variety of disorders. However, the spectra are complex and only the high concentration metabolites can be easily selected for analysis. This analysis ignores a vast amount of information, largely embedded in noise, with potential clinical and biochemical value. Recently, statistical pattern recognition and neuro-computing techniques have been proposed to classify these spectra automatically. However, a major problem with designing automated systems for proton spectra is the large dimensionality of the spectral vectors compared to the available number of spectra. With too many spectral components the system will have a large number of undetermined parameters, overfitting the data and performing poorly. In this thesis we are concerned with the development of reliable and robust neural networks for the classification and analysis of proton NMR spectra from human brain tumours in vitro. This involves both theoretical studies in pattern recognition and the practical implementation of algorithms to analyse the spectra. We first demonstrates that neural networks outperform classical non-parametric pattern recognition techniques in classifying the spectra. Changes to standard neural network algorithms are then proposed to speed up their learning procedures and improve their generalisation performance to suit applications in clinical and biomedical NMR spectroscopy. We also propose new feature selection techniques, integrated in the network training procedure, to reduce the dimensionality of the spectral vectors automatically. They operate by assessing the relevance of each spectral component to the classification or prediction process during training. The implementation of these feature selection techniques should help in understanding the biochemical processes revealed by the spectra and allow the development of networks that make finer distinction between these spectra. The use of the above algorithms is demonstrated in two applications. The first classifies high and low grade gliomas from their high resolution proton spectra obtained from short term cell culture extracts of human tumours. The second predicts the in vitro chemosensitivity of human malignant glioma cell cultures towards the drug CCNU, prior to their actual treatment with CCNU. We conclude, first, that the realisation of the full clinical potential of NMR spectroscopy relies on understanding the role played by the constituent metabolites in the characterisation of and the differentiation between tissue types, as well as on the ability to relate metabolism to other tissue properties such as the response to drugs. Second, techniques of pattern recognition, particularly those of neuro-computing, could play an important role towards achieving these goals by helping in the interpretation of the spectra and the understanding of tissue biochemistry.

Type: Thesis (Doctoral)
Qualification: Ph.D
Title: Artificial neural networks in pattern recognition: Classification and analysis of proton NMR spectra of human brain tumours
Open access status: An open access version is available from UCL Discovery
Language: English
Additional information: Thesis digitised by ProQuest.
Keywords: Biological sciences; Health and environmental sciences; Nuclear magnetic resonance spectroscopy
URI: https://discovery.ucl.ac.uk/id/eprint/10102108
Downloads since deposit
Download activity - last month
Download activity - last 12 months
Downloads by country - last 12 months

Archive Staff Only

View Item View Item