An Investigation Of The Potential Value Of Neural Network Computing In Diagnosis And Assessment Of Breast Cancer: Analysis Of Blood Plasma By 1H Nuclear Magnetic Resonance Spectroscopy

Abstract

Introduction 'H Nuclear Magnetic Resonance (WR) spectra were obtained from plasma samples from patients with breast cancer, with benign breast disease and from healthy volunteers. Preprocessing of the data included the use of principal component (PC) analysis and resulted in reduction of the data dimensions to 6-8 PC scores. A backpropogation neural network with two hidden layers was used to learn how to convert these PC scores (used as inputs) into outputs indicative of class membership. Examination of the similarity between the PC scores for these samples was performed using cluster analysis and from the weights obtainedfrom a single layer network From these results it could be predicted that, contrary to our prior expectations, it should be relatively easy to distinguish benign breast disease from the other two clwses but that there would be considerable overlap between cancer and normal subjects. fiis was also the conclusion of an assessment of the reliability of the network when c1assifLing samples that had not been included in the learning stage. Ihe best resultsfrom this study were that 85% of samples in the normal versus benign diseace datmet could be correctly assigned after having been omitted from the learning stage. i%e other 15% were designated ar unclassified since the output scores were ambiguous. Moderately good results from the most clinically interesting pair of classes, benign disease versus cancer, were improved on by incorporating information about type of treatment and secondary diseases into additional output scores. l%is approach appears to be helpful in reducing the problems associated with heterogeneity within one or more of the classes.