A Novelty Detection Approach to Classification of Breast Tissue Containing Microcalcifications

Abstract

Appearance of microcalcifications in mammograms is one of the early signs of breast cancer. In this work, one-class support vector machines (SVM), a novelty detection method, is utilized for detection of the mammogram samples containing microcalcifications. These samples are small regions of the mammograms with the size of 25x25 pixels. Each of the samples are represented by 25 features that are already proven to be accurate identifiers of the microcalcifications. Since the obtained classification performance of one-class SVM with all these 25 features is very low (accuracy = 0.5575, sensitivity = 0.2107, specificity = 0.9042), number of these features is reduced by using principal component analysis (PCA). Training a classifier only with the PCA features achieves an improved performance (accuracy = 0.9464, sensitivity = 1.0000, specificity = 0.8927) where the number of false negative samples is reduced from 206 to 0.