Anne Rosemary Tate
Magnetic resonance spectroscopy (MRS) provides a unique non-invasive method for obtaining information on the biochemistry of living tissue in situ, and therefore has great potential as a clinical tool. However, presently in vivo MRS is used mainly for research, rather than for clinical applications. There are a number of reasons for this. The information may be difficult to extract from the spectrum due to low signal-to-noise ratio and other problems associated with obtaining a signal from living tissue. Interpretation may be difficult due to the large number of metabolites represented by the spectra. Another problem is that most current methods for analysing MRS data are targeted at providing information on specific metabolites, rather than the more general information appropriate for clinical applications, such as the disease stage or state of the tissue being examined. This thesis shows how pattern recognition techniques may be used to help overcome these problems and to provide methods for classifying in vivo spectra according to their tissue type. A prototype system for classifying spectra is developed using features that are extracted automatically, using the whole spectrum, rather than selected peaks. These features were selected purely on the basis of their power to discriminate between different types of spectra, using no prior knowledge of biochemistry. Among the techniques used were wavelets, principal component analysis and linear discriminant function analysis. These techniques were tested on two sets of in vivo data: 75 \carbon\ spectra obtained from healthy human volunteers from three different dietary groups of adipose tissue in the leg and 55 \phos\ spectra obtained from tumorous and normal tissue in rats. For both datasets most of the spectra were assigned to their correct groups (94\% of the \carbon\ and 86 -- 100\% of the \phos\ spectra) without the need for explicit identification or measurement of peaks.
Download compressed postscript file