Classification of Hyperspectral Colon Cancer Images Using Convolutional Neural Networks

Abstract

CLASSIFICATION OF HYPERSPECTRAL COLON CANCER IMAGES USING CONVOLUTIONAL NEURAL NETWORKS by Sean J. Mobilia Hyperspectral images are 3-D images, which contain data in hundreds of spectral bands as opposed to 2-D images, which contain data in at most 3 bands (red, green, and blue). Hyperspectral imagery was initially developed for remote sensing; however, recently, researchers have started to see its potential in medical diagnosis and cancer detection. Hyperspectral images provide massive amounts of data about the objects they are studying, and this causes challenges during information processing. Machine learning tools, such as convolutional neural networks (CNNs), are known to be successful in extracting features and classifying traditional 2-D images. This thesis proposes CNN architectures for processing hyperspectral data for colon cancer detection. Using data taken from a limited number of colon tissue samples, this thesis shows that the proposed CNN architecture can classify cancerous and noncancerous tissue samples utilizing hyperspectral information. The obtained results are compared to grayscale images of the same tissue samples, looking both at grayscales of the individual hyperspectral bands and panchromatic grayscale images in which the spectral bands are merged together. The CNN using the hyperspectral data shows advantages over the grayscale data, with a 5.6% improvement in accuracy and a 0.037 improvement in F1 score over the individual band grayscale images and a 21.7% improvement in accuracy and a 0.178 improvement in F1 score over the panchromatic grayscale images. The results are also compared to a K-nearest neighbor (KNN) classifier and a logistic regression (LR) classifier using the hyperspectral data, and the CNN shows advantages over both. The CNN has a 17.9% improvement in accuracy and a 0.141 improvement in F1 score over the KNN classifier and a 5% improvement in accuracy and a 0.061 improvement in F1 score over the LR classifier. ACKNOWLEDGMENTS I would like to thank my advisor, Professor Birsen Sirkeci, for all her help and support during this project and for having tremendous patience with me. I would also like to thank my other committee members, Professor Robert Morelos-Zaragoza and Professor Chang Choo, for taking the time to read and review my work.