Feature based analysis of axial-shear strain imaging for breast mass classification

Normal and shear strain elastography have demonstrated the potential for differentiating benign from malignant breast masses in-vivo. In this paper we present classification results using multiple features that are exacted from axial-strain images, including the size ratio (SR), and stiffness contrast (CN) and the normalized axial shear strain area (NASSA) feature from axial-shear strain images. We report on the analysis of images obtained from 109 radiofrequency data sets acquired from three different hospitals with a benign/malignant (B/M) ratio of 58/51 established via biopsy results. Radiofrequency data were acquired during a free-hand palpation study using Siemens Antares or Elegra clinical ultrasound systems. Axial displacement and strains were estimated using a multi-level pyramid based two-dimensional cross-correlation algorithm. Since the mass boundaries on 19 of the B-mode images were isoechoic, size ratio analysis could be performed only on 90 data sets. Receiver operating characteristic (ROC) analysis using leave-one-out approach shows that the area under the curve (AUC) for the NASSA feature alone was 0.90 and the stiffness contrast was 0.61 for 109 patients. The AUC for the size ratio feature was 0.84 based on 90 cases. A linear support vector machine using a leave-one-out cross-validation approach was used to study the performance improvement obtained using a combination of these features. Since the size ratio feature was only available for 90 cases, the ROC analysis for the combined features was done on 90 patients. The best classification performance was obtained with the utilization of all the features with an AUC of 0.94. ROC analysis demonstrates the potential of using the above combination of strain features for in-vivo breast mass differentiation.