%0 Thesis
%9 Doctoral
%A Zhu, Y.
%B Department of Statistical Science
%D 2009
%F discovery:16359
%I UCL (University College London)
%P 219
%T Statistical aspects of elastic scattering spectroscopy with applications to cancer diagnosis
%U https://discovery.ucl.ac.uk/id/eprint/16359/
%X Elastic scattering spectroscopy (ESS), is a non-invasive and real-time in vivo  optical diagnosis technique sensitive to changes in the physical properties of human  tissue, and thus able to detect early cancer and precancerous changes. This thesis  focuses on the statistical issue on how to eliminate irrelevant variations in the high-dimensional  ESS spectra and extract the most useful information to enable the  classification of tissue as normal or abnormal.  Multivariate statistical methods have been used to tackle the problems, among  which principal component discriminant analysis and partial least squares  discriminant analysis are the most explored throughout the thesis as general tools for  supervised dimension reduction and classification. Customized multivariate methods  are proposed in the specific context of ESS.  When ESS spectra are measured in vivo by a hand-held optical probe, differences  in the angle and pressure of the probe are a major source of variability between the  spectra from replicate measurements. A customized spectral pre-treatment called error  removal by orthogonal subtraction (EROS) is designed to ameliorate the effect of this  variability. This pre-treatment reduces the complexity and increases both the accuracy  and interpretability of the subsequent classification models when applied to early  detection of cancer risk in Barrett’s oesophagus.  For the application of ESS to diagnosis of sentinel lymph node metastases in  breast cancer, an automated ESS scanner was developed to take measurements from a  larger area of tissue to produce ESS images for cancer diagnosis. Problems arise due  to the existence of background area in the image with considerable between-node  variation and no training data available. A partially supervised Bayesian multivariate  finite mixture classification model with a Markov random field spatial prior in a  reduced dimensional space is proposed to recognise the background area  automatically at the same time as distinguishing normal from metastatic tissue.