GENERALIZATION PERFORMANCE USING BACKPROPAGATION ALGORITHMS APPLIED TO PATTERNS DERIVED FROM TUMOR H-1-NMR SPECTRA

1H nuclear magnetic resonance spectroscopy enables many metabolites in normal and tumour tissue to be examined and has considerable clinical potential. However, the spectra are complex and much of the biochemical information they contain is inaccessible to visual analysis. This paper assesses the training and generalization (classifying unknown patterns) performance of backpropagation neural networks using patterns derived from rat tumour spectra, and investigates the technique of training with noise added to the patterns. We used a basic set of 84 patterns (obtained from spectra of 8 tissue classes) each of 180 ordinates sampled over the range of 0·0 to 4·5 ppm. A network with 1 hidden layer was used with a consistent protocol involving different selections of 63 training and 21 test patterns from the basic set. Gaussian noise with zero mean and specified SD could be added to training or test patterns independently. With 15 hidden units and non-noisy patterns, a training score of 63/63 and a total squared error of <0·003 was reached typically within 150 sweeps. The test score averaged 19·2/21 with a test error of 0·06. Addition of test noise with SD of 0·20 decreased the test score to 17·0/21 and increased the test error to 0·14. However, concomitant training with training noise SD up to 0·075 increased the test score to 17·8 and reduced the error to 0·13. These results show, first, that a network containing only 15 hidden units can be trained to classify correctly a set of patterns derived from tumour spectra, with promising generalization to new patterns. Second, adding noise to the training patterns can partially offset the degradation of generalization performance produced by adding noise to the test patterns. If developed further, the technique of adding noise may be useful in helping neural networks perform more complex classification tasks using patterns of noisy spectra. © 1993 Academic Press. All rights reserved.