Detection and Classification of Lesions in Mammographies Using Neural Networks and Morphological Wavelets

Breast cancer is a worldwide public health problem, with a high rate of incidence and mortality. The most widely used to perform early on possible abnormalities in breast tissue is mammography. In this work we aim to verify and analyze the application of classifiers based on neural networks (multilayer perceptrons, MLP, and radial bassis functions, RBF), and support vector machines (SVM) with several different kernels, in order to detect the presence of breast lesions and classify them into malignant of benign. We used the IRMA database, composed by 2,796 patch images, which consist of 128 x 128 pixels region of interest of real mammography images. IRMA database is organized by BI-RADS classification (normal, benign and malignant) and tissue type (dense, extremely dense, adipose, and fibroglandular), generating 12 classes. Each image was represented by Zernike moments extracted from the components of the two levels decomposition by morphological wavelets. Multilayer perceptrons with two layers were the successful methods, reaching an accuracy rate of 96.20%, proving the possibility of building a computer-aided diagnosis system to improve accuracy of mammogram analysis, contributing to improve prognosis as well.