Graph Modeling for Identifying Breast Tumor Located in Dense Background of a Mammogram

Abstract

Identifying breast tumor in a mammogram is a challenging task even for experienced radiologists if the tumor is located in a dense tissue. In this study, a novel superpixel based graph modeling technique is proposed to extract texture features from the computer identified suspicious regions of mammograms. Graph models are constructed from specific structured superpixel patterns and used to generate feature vectors used for classifications of regions of mammograms. Two mammographic datasets were used to evaluate the effectiveness of the proposed approach: the publicly available Digital Database for Screening Mammography (DDSM), and a local database of mammograms (BSSA). Using Linear Discriminant Analysis (LDA) classifier, an AUC score of 0.910 was achieved for DDSM and 0.893 for BSSA. The results indicate that graph models can capture texture features capable of identifying masses located in dense tissues, and help improve computer-aided detection systems.