Development and evaluation of pattern recognition techniques for fluorescence diagnosis of atherosclerosis

摘要

A field where fluorescence spectroscopy might be of great interest for diagnosis, is coronary atherosclerosis and therefore spectroscopic characterization of cardiovascular tissues has been extensively studied. Nevertheless there are several limitations in the precise interpretation of the spectroscopic differences, between normal and atherosclerotic arteries since the tissue is a complex and multilayer structure. Therefore the spectra of individual chromophores could overlap and re-absorption phenomena could occur, too. Another major difficulty arises from the necessity of convenient classification algorithms and the assessment of their feasibility to use fluorescence information, for accurate diagnosis. As a result in order to assess the feasibility of utilizing spectral information to discriminate arterial tissue type several classification algorithms were developed and evaluated. In this work the following pattern recognition techniques have been tested and evaluated: (1) Distance measure (or norm, or metric) based pattern recognition techniques. Methodologically speaking, based on the histopathological diagnosis, a training set of spectra has been classified into four different categories (healthy, fibrous, calcified, heavy calcified) and in each of these four training groups a representative spectrum has been recorded. (2) A pattern recognition method based on statistical considerations. Discrimination between either the four aforementioned classes (categories) or pairs of them is achieved since peak intensities in appropriate wavelengths appear to correlate efficiently with tissue type. The difference of each training set member from the corresponding representative has been defined by using various appropriate distance measures and the sample statistical properties for each category of the training group has been found. Appropriate statistical analysis has been performed in order to deduce the distribution of the distance measures and of the coefficients of the whole population for each one of the four categories, with at least 99% confidence interval. A validation set of samples has been used in order to test and compare the aforementioned pattern recognition algorithms. A performance comparison of the aforementioned algorithms has been undertaken.