Shearlet-based texture feature extraction for classification of breast tumor in ultrasound image

Abstract

To augment the classification accuracy of the ultrasound computer-aided diagnosis (CAD) for breast tumor detection based on texture feature, we proposed to extract texture feature descriptors by the shearlet transform. Shearlet transform provides a sparse representation of high dimensional data with especially superior directional sensitivity at various scales. Therefore, shearlet-based texture feature descriptors can characterize breast tumors well. In order to objectively evaluate the performance of shearlet-based features, curvelet, contourlet, wavelet and gray level co-occurrence matrix based texture feature descriptors are also extracted for comparison. All these features were then fed to two different classifiers, support machine vector (SVM) and AdaBoost, to evaluate the consistency. The experimental results of breast tumor classification showed that the classification accuracy, sensitivity, specificity, positive predictive value, negative predictive value and Matthew's correlation coefficient of shearlet-based method were 91.0±3.8%, 92.5±6.6%, 90.0±3.8%, 90.3±3.8%, 92.6±6.3%, 0.822±0.078 by SVM, and 90.0±2.8%, 90.0±4.0%, 90.0±2.3%, 89.9±2.4%, 90.1±3.6%, 0.803±0.056 by AdaBoost, respectively. Most of the shearlet-based results significantly outperformed those of other method based results under both the classifiers. The results suggest that the proposed method can well characterize the properties of breast tumor in ultrasound images, and has the potential to be used for breast CAD in ultrasound image.