A new method for breast micro-calcification detection and characterization using digital temporal subtraction of mammogram pairs

Abstract

Breast cancer is one of the most deadly malignancies worldwide and the second leading cause of death in women. Mammography, i.e. the screening for breast cancer with x-ray imaging, has significantly improved the prognosis of patients diagnosed with the disease. The evaluation of mammograms requires a panel of radiologists, but even well trained experts can err in their assessment. For this reason, Computer-Aided Detection (CAD) systems are becoming more prevalent. In this paper, we introduce a novel approach for breast Micro-Calcification (MC) diagnosis using temporal sequences of digital mammograms. The goal is to increase the MC detection accuracy by subtracting prior images. A new dataset, with precise marking of MC locations, was created specifically for this study. The proposed approach began with temporal subtraction of mammograms, after demon-based registration, which effectively removed unchanged regions and MCs (17.3% reduction in the number of MCs). The second step was the classification of the MCs as benign or suspicious using the subtracted images. A set of diverse features were selected for the classification. Four different classifiers were tested with leave-one-patient-out cross-validation. For comparison, the MC classification was also performed, using single mammograms, without temporal subtraction. The average accuracy of the classification of the MCs as benign or suspicious was 91.3% without and 99.2% with temporal subtraction using Support Vector Machines (statistically significant $\mathbf{p}=0.026$ ). These results show that temporal subtraction could be a valuable addition to CAD systems to assist radiologists in effectively detecting breast MCs.