Fusing Low-Level Visual Features and High-Level Semantic Features for Breast Cancer Diagnosis in Digital Mammograms

Mammography is still the most effective procedure for early diagnosis of breast cancer. Computer-aided Diagnosis (CAD) systems can be very helpful in this direction and an assistance tool for radiologists to recognize abnormal and normal regions of interest in digital mammograms faster than traditional screening program. In this paper, we propose a new method for breast cancer recognition of all types of lesions in digital mammograms using multimodal features in a quaternionic representation. The proposed method consists of two steps: First, a feature extraction module utilizes two dimensional discrete transforms based on ART, Shapelets and Gabor filters to extract low-level visual descriptors of ROS. In addition, semantic mammogram information provided by radiologists, is encoded into a 16-bit length word used as high-level feature vector. All multimodal features are fused in a compact quaternionic representation prior to their presentation to the classification module in the second step. For the classification task, ROSs are diagnosed using trained Support-Vector-Machines (SVMs). The proposed method is evaluated on regions taken from the DDSM database. The proposed feature set achieves a mean accuracy of 87.22% (±2.678) and AUC = 0.98, that is 34% higher than using only visual descriptors, proving the significant information that semantic descriptors contain.