Breast masses identification through pixel-based texture classification

Abstract

Mammographic image analysis plays an important role in computer-aided breast cancer diagnosis. To improve the existing knowledge, this paper proposes a new efficient pixel-based methodology for tumor vs non-tumor classification. The proposed method firstly computes a Gabor feature pool from the mammogram. This feature set is calculated through multi-sized evaluation windows applied to the probabilistic distribution moments, in order to improve the accuracy of the whole system. To deal with a high dimensional data space and a large amount of features, we apply both a linear and non-linear pixel classification stage by using Support Vector Machines (SVMs). The randomness is encoded when training each SVM using randomly sample sets and, in consequence, randomly selected features from the whole feature bank obtained in the first stage. The proposed method has been validated using real mammographic images from well-known databases and its effectiveness is demonstrated in the experimental section. © 2014 Springer International Publishing.