In this thesis, we explore methods for incorporating invariance into a support vector machine (SVM) classifier in detection of clustered... Show moreIn this thesis, we explore methods for incorporating invariance into a support vector machine (SVM) classifier in detection of clustered microcalcifications (MC) in mammogram images. Unlike standard SVM, both virtual SVM and tangent vector SVM can include prior information into a trained model. We formulate MC detection as a two-class classification problem and apply these three types of SVM classifiers to this problem. The issue of dimensional reduction is considered in the tangent vector SVM, which has influence on the computational cost and complexity of the algorithm. We test and compare their performance on a set of 200 clinical mammogram images which contain a total of 5,115 MCs. In the experiments these classifiers are optimized with a training procedure for model selection. We evaluate the detection performance using both receiver operating characteristic (ROC) curves and free-response operating characteristic (FROC) curves. The results show that both virtual SVM and tangent vector SVM can outperform the standard SVM. The use of dimensional reduction in tangent vector SVM can effectively reduce the computational cost. M.S. in Electrical Engineering, December 2011 Show less
