Transactions on Mass-Data Analysis of Images and Signals (ISSN:1868-6451)
Volume 2- Number 1 - September 2010 - Pages 96-111
Classification of LIBS Protein Spectra Using Multilayer Perceptrons
T. Vance1, D. Pokrajac1, A. Lazarevic2, A. Marcano1, Y. Markushin1, S. McDaniel1 and N. Melikechi 1
1State University, 1200 N. Dupont Highway, Dover, DE USA 19901
2United Technologies Research Center, 411 Silver Lane, East Hartford, CT USA 06108
The spectroscopy community has recently begun to explore various artificial intelligence techniques for analysis of specimen based on spectroscopy data. In this paper, we investigate a variety of neural network learning algorithms for classification of laser-induced breakdown spectroscopic data of four proteins: Bovine Serum Albumin (the most abundant protein in blood plasma), and three other proteins—Osteopontin, Leptin and Insulin-like Growth Factor II—that have been identified as potential biomarkers of ovarian cancer. For the purpose of dimensionality reduction, we utilize a version of the principal component analysis algorithm suitable for high dimensional data, while for classification we use artificial neural networks (multilayer perceptrons) algorithms with one and two hidden layers and selected optimal number of principal components. We also perform analysis of variance (ANOVA) experiments to determine the neural network parameters and their influence on classification accuracy. Our experimental results show that the best classification accuracy, 93.41%, is achieved by using the Resilient Backpropagation learning algorithm with 22 principal components, 19 hidden neurons, and one hidden layer.