Extracting and Evaluating Morphological Features from Microcalcifications in Breasts Mammograms

Abstract

Breast cancer is the one that most affects women worldwide. Mammographic examinations are considered the best way of detecting non-palpable breast lesions, although there are several factors that make difficult the diagnostic accuracy in breast cancer screening. Computer-aided diagnosis (CADx) systems have been developed aiming to help breast cancer diagnosis. They are based on microcalcifications (MCs) characteristics and can be treated as a three-step system: (i) MCs segmentation; (ii) extraction and selection of features from the segmented MCs; (iii) lesions classification. In this work, a set of 25 morphological features were extracted from lesions containing MCs presented on 190 mammographic images. The MCs of these images were segmented and binarized in a previous work, using Mathematical Morphology techniques. After extracting the morphological features, a selection method based on mutual information ranked the features, and a linear discriminant analysis used a forward procedure to search, among the ranked features, the best subset to classify the lesions. The best classification performance was achieved with 4 morphological features: Fourier Factor, Zero crossing, Long-axis to short axis ratio and Mean value of the Normalized Radial Length.