Design of Machine Learning Algorithms with Applications to Breast Cancer Detection

• Main Author: Mu, Tingting
• Published: University of Liverpool 2008
• Subjects: 616.99
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.490621

Machine learning is concerned with the design and development of algorithms and techniques that allow computers to 'learn' from experience with respect to some class of tasks and performance measure. One application of machine learning is to improve the accuracy and efficiency of computer-aided diagnosis systems to assist physician, radiologists, cardiologists, neuroscientists, and health-care technologists. This thesis focuses on machine learning and the applications to breast cancer detection. Emphasis is laid on preprocessing of features, pattern classification, and model selection. Before the classification task, feature selection and feature transformation may be performed to reduce the dimensionality of the features and to improve the classification performance. Genetic algorithm (GA) can be employed for feature selection based on different measures of data separability or the estimated risk of a chosen classifier. A separate nonlinear transformation can be performed by applying kernel principal component analysis and kernel partial least squares. Different classifiers are proposed in this work: The SOM-RBF network combines self-organizing maps (SOMs) and radial basis function (RBF) networks, with the RBF centers set as the weight vectors of neurons from the competitive layer of a trained SaM. The pairwise Rayleigh quotient (PRQ) classifier seeks one discriminating boundary by maximizing an unconstrained optimization objective, named as the PRQ criterion, formed with a set of pairwise const~aints instead of individual training samples. The strict 2-surface proximal (S2SP) classifier seeks two proximal planes that are not necessary parallel to fit the distribution of the samples in the original feature space or a kernel-defined feature space, by ma-ximizing two strict optimization objectives with a 'square of sum' optimization factor. Two variations of the support vector data description (SVDD) with negative samples (NSVDD) are proposed by involving different forms of slack vectors, which learn a closed spherically shaped boundary, named as the supervised compact hypersphere (SCH), around a set of samples in the target class. \Ve extend the NSVDDs to solve the multi-class classification problems based on distances between the samples and the centers of the learned SCHs in a kernel-defined feature space, using a combination of linear discriminant analysis and the nearest-neighbor rule. The problem of model selection is studied to pick the best values of the hyperparameters for a parametric classifier. To choose the optimal kernel or regularization parameters of a classifier, we investigate different criteria, such as the validation error estimate and the leave-out-out bound, as well as different optimization methods, such as grid search, gradient descent, and GA. By viewing the tuning problem of the multiple parameters of an 2-norm support vector machine (SVM) as an identification problem of a nonlinear dynamic system, we design a tuning system by employing the extended Kalman filter based on cross validation. Independent kernel optimization based on different measures of data separability are a~so investigated for different kernel-based classifiers. Numerous computer experiments using the benchmark datasets verify the theoretical results, make comparisons among the techniques in measures of classification accuracy or area under the receiver operating characteristics curve. Computational requirements, such as the computing time and the number of hyper-parameters, are also discussed. All of the presented methods are applied to breast cancer detection from fine-needle aspiration and in mammograms, as well as screening of knee-joint vibroarthrographic signals and automatic monitoring of roller bearings with vibration signals. Experimental results demonstrate the excellence of these methods with improved classification performance. For breast cancer detection, instead of only providing a binary diagnostic decision of 'malignant' or 'benign', we propose methods to assign a measure of confidence of malignancy to an individual mass, by calculating probabilities of being benign and malignant with a single classifier or a set of classifiers.