DETECTION OF DISEASE IN CT SCANS OF THE LUNG USING TEXTURE ANALYSIS AND MACHINE LEARNING TECHNIQUES

ii Here we aim to evaluate a range of classifiers for their use in the detection of disease in Computed Tomography scans. The classifiers evaluated range from basic algorithms such as the k nearest neighbour to Neural Network approaches and Support Vector Machines. Each classifier then is trained and tested on two different types of data sets compiled by a radiologist, consisting of 44 textural features obtained from image blocks found in the CT scans. The first set contains highly representative images of each of the patterns taken from 79 patients, and the second of areas deemed broadly representative of each of the patterns taken from 91 patients. The classification accuracy for training and testing on these sets is presented for each classifier over the factors that affect their performance. Each of the classifiers examined in this report are required to be able to provide indications to how confidant they are of a classification and when they are operating outside their training region. This is essential for in this application for both safety i.e. recognition of novel points, and post processing of the classified image. In finding the maximum accuracy for each classifier, forward feature selection was used. This allow