Computer-Aided Diagnosis for 3-D Power Doppler Breast Ultrasound

The breast cancer is always a major cause of death for women among all kind of cancers. In recent years, the computer-aided diagnosis systems have been developed rapidly and they can not only detect the tumors but also distinguish malignant tumors from benign ones. Therefore, the need of the breast biopsy of the detected tumors might be further decreased. Since the blood flow is successfully detected with the conventional US image by the Doppler ultrasound (US), the studies of tumor vascularity, tumor morphology, and texture of tumor have played important roles to diagnose diseases of breast recently. In general, malignant tumors need more complex blood vessels to obtain sufficient nutrients for growing, and the volume and surface of malignant tumors are larger and more irregular than those of benign tumors. In the past, researches about Doppler US only used the characteristics based on the tumor or vascular information for diagnosing tumor. In this paper, we demonstrate a computer-aided diagnosis (CAD) system for three-dimensional (3-D) power Doppler breast US images that can quantify the characteristics of both vascularity and tumor. In order to obtain the features of tumor and vascularity, the level-set method is applied to segment the tumor margin and thinning approach is used to skeletonize the vascularity. Then, the features including the texture information based on co-occurrence matrix, shape information, and ellipsoid fitting information are extracted based on the segmented 3-D tumor contour, and the vascular morphology are quantified from the skeletonized vessels. The features are used to classify the benign and malignant tumors by the binary logistic regression model. In the experiment, 82 biopsy-proved lesions including 41 benign tumors and 41 malignant tumors are used to test the diagnosis performance of the proposed CAD system. From the experimental results, it is found that the features of tumor combined with vascular features has better performance than using single type of features. Moreover, the proposed method could achieve a high performance with the accuracy, sensitivity, specificity and Az value being 85.37% (70/82), 85.37% (35/41), 85.37% (35/41), and 0.9104, respectively.