In Vivo Classification of Breast Masses Using Features Derived From Axial-Strain and Axial-Shear Images

Breast cancer is currently the second leading cause of cancer deaths in women. Early detection and accurate classification of suspicious masses as benign or malignant is important for arriving at an appropriate treatment plan. In this article, we present classification results for features extracted from ultrasound-based, axial-strain and axial-shear images of breast masses. The breast-mass stiffness contrast, size ratio, and a normalized axial-shear strain area feature are evaluated for the classification of in vivo breast masses using a leave-one-out classifier. Radiofrequency echo data from 123 patients were acquired using Siemens Antares or Elegra clinical ultrasound systems during freehand palpation. Data from four different institutions were analyzed. Axial displacements and strains were estimated using a multilevel, pyramid-based two-dimensional cross-correlation algorithm, with final processing block dimensions of 0.385 mm × 0.507 mm (three A-lines). Since mass boundaries on B-mode images for 21 patients could not be delineated (isoechoic), the combined feature analysis was only performed for 102 patients. Results from receiver operating characteristic (ROC) demonstrate that the area under the curve was 0.90, 0.84, and 0.52 for the normalized axial-shear strain, size ratio, and stiffness contrast, respectively. When these three features were combined using a leave-one-out classifier and support vector machine approach, the overall area under the curve improved to 0.93.