Computer-Aided Diagnosis Of Breast Tumors With Ultrasonic Shear-Wave Elastogram

Breast cancer is one of the most common malignant tumors occurred in women, which seriously threatens women’s health. For its complex pathogenesis and multitudinous influencing factors, people haven’t yet found an ideal method to cure breast cancer. The main clinical treatment for breast cancer relies on surgery, combined with chemotherapy, radiotherapy, targeted therapy, immune therapy and other auxiliary therapies. Diagnosis and treatment of breast cancer at early phase is one of the effective ways to improve the cure rate, while radiological examination is a critical way of the early breast cancer screening and diagnosis. Ultrasonic elastography techniques provide a more accurate and effective method for early breast cancer examination. The shear-wave elastography(SWE) is one new approach of ultrasonic elastography, which utilizes the principle that the propagation speed of ultrasound waves differs in tissues with different degrees of hardness. To implement the SWE, the shear waves in tissues are generated by using the acoustic radiation force, and then the ultrafast real-time scanning is utilized to dynamically capture the propagation speed of the shear waves; therefore, a two-dimensional image is formed representing the tissue elasticity information. Ultrasonic elastography can detect the disease of breast tissue before the change of the tissue morphology, which is valuable for the early diagnosis of breast cancer.By using SWE technique, a two-dimensional quantitative elasticity image is captured with high reproducibility. Thus, the SWE has gained increased attention in academia. However, in the clinical diagnosis, it needs ultrasonographers to manually interpret the SWE and B-mode ultrasound images produced by elastography imaging, and thus it leads to high subjectivity and variability, as well as poor reproducibility of the diagnosis result. To date, quantitative analysis of shear wave elastogram is still relatively lacking, thus there is an urgent need for a computer-aided diagnosis system to quantify and classify the breast tumor images, in order to reduce the workload of ultrasonographers and to enhance the objectivity, reproducibility and accuracy of breast cancer diagnosis. In this paper, we mainly focus on the four parts of a computer-aided diagnosis system, including the filtering, segmentation, feature extraction and classifier design for ultrasound images of breast tumors, by using computerized algorithms to accomplish the identification of benign and malignant breast tumors objectively and automatically.First, the speckle noise reduction in breast ultrasound image. In this paper, we introduce a new edge detector with high robustness into anisotropic diffusion(AD) model to improve its ability of speckle noise suppression. Firstly, we combine the distance mapping function and Mc Ilhagga edge detector to form an improved Mc Ilhagga edge detector which outputs continuous values. Then we embed the improved Mc Ilhagga-based edge detector into AD model to generate the Mc Ilhagga-based anisotropy diffusion(MAD) model with high robustness and edge preservation ability. Based on the similar idea, the Gabor edge detector is embedded into the AD model, generating the Gabor-based anisotropic diffusion(GAD) model. The MAD and GAD methods were compared with several traditional filtering methods in experiments on both simulated and clinical ultrasound images. For simulated image with a low speckle noise variance(<0.1), the denoising results of MAD and GAD were comparable to those of the traditional algorithms. When the noise variance was large(= 0.2), the peak signal-to-noise ratio(PSNR) of both MAD and GAD were close to the best result of the traditional filtering methods(about 15 d B); the Pratt’s figure of merit(FOM) and structural similarity index(SSIM) of MAD were 0.38 and 0.82, respectively, which were increased by 80.77% and 35.15% compared with the best results of the traditional methods; the FOM and SSIM of GAD were 0.74 and 0.86, respectively, which were increased by 259.61% and 42.41% compared with the best results of the traditional methods. After comprehensive comparisons, we chose GAD in the subsequent segmentation algorithm for more accurate quantitative analysis of the ultrasound image.Second, the breast tumor image segmentation. Since the contours of breast tumors in the SWE images are not clear and obvious, we segment the B-mode ultrasound image to get the tumor contours and then map the contours back into the SWE image. Thus, the segmentation of B-mode ultrasound image affects the segmentation of the SWE image. We proposed an improved reaction-diffusion level set model for segmentation of breast tumor ultrasound images, by embedding an edge stop function based on GAD into the classic reaction-diffusion model. We compared our improved model with the original reaction-diffusion model by using the manually traced borders of the lesions as the gold standard. The true positive rate and accuracy of the improved method were 86.65% and 96.33%, increased by 9.27% and 0.79%, respectively. The root mean square error was 7.26 pixel, reduced by 14.8%.Third, the breast tumor feature extraction. The extraction of quantitative parameters for breast tumors is crucial for accurate classification of benign and malignant tumors. Due to the biological heterogeneity of tumor cells, the pixels in both the SWE and B-mode ultrasound images present non-uniform distributions. The non-uniformity can help differentiate between benign and malignant tumors. Therefore, in addition to the extraction of traditional morphological features, we focus on extracting texture features of the tumors, in order to represent the homogeneity and heterogeneity of the elastic modulus(shear-wave velocities) and B-mode gray values in SWE and B-mode images of the benign and malignant tumors. Texture features extracted from SWE and B-mode images include the mean, standard deviation, maximum, minimum, and other first-order statistics and features of the gray-level co-occurrence matrix(GLCM), such as contrast, correlation, energy, uniformity and entropy. These texture features are computed from the images and thus are called the original-domain texture features. Furthermore, we use the contourlet transform in a multi-direction and multi-resolution framework to convert an image to the contourlet domains, from which we extract the contourlet-domain texture features in SWE images and B-model images. The experimental results show that the texture features calculated from the two types of domains could more comprehensively quantify the properties of breast tumors.Fourth, the differentiation between benign breast tumors and malignant tumors. We combine various types of features extracted from the SWE images and B-mode images as the input variables of the support vector machine and the deep learning basis learner method, to compare the performance of different feature combinations and different classifiers in breast tumor classification. The experimental results show that when combining the texture features extracted from the first contourlet level of the SWE and B-mode images with the morphological features, the classification sensitivity, specificity, accuracy and Youden index of the basis learner method were 91.30%, 99.26%, 96.04% and 90.56%, respectively, demonstrating the high precision of the computer-aided diagnosis for breast tumors.