Cervical Lymph Node Ultrasound Image Analysis And Applications

Lymph node is one of the important organs as the immunologic filter in the body. The abnormity of lymph nodes often denotes pathological changes within their attributive areas. Therefore, the evaluation of lymph nodes is meaningful for the disease diagnosis. Since cervical lymph nodes are the prone regions of the lymphadenopathy, they are especially important in the clinic. Among all kinds of imaging modalities, the ultrasound is the preferred technique for the evaluation of cervical nodes, due to its good qualities of non-invasion, real time, cheapness, and convenience. However, the sonographic examination of cervical lymph nodes is highly subjective. It will be helpful for the radiologists to diagnose the cervical nodes on sonograms more objectively by the quantified sonographic features and analyses provided by a computer. In this dissertation, a computer-aided diagnosis (CAD) system has been developed, which has the main functions of image segmentation, feature extraction and classification of cervical nodes on sonograms.Firstly, an improved gradient vector flow (GVF) snake model is proposed to segment the nodes on sonograms. The traditional GVF snake model segments an image only based on the intensity attribute. Since the node sonograms exhibit poor quality with speckle noises and low contrast, the performance of the traditional GVF snake model is not satisfied for the sonograms. To fully utilize the texture information in the sonograms, the GVF is formed with the edge flow that synthesizes the intensity and texture information. A semi-automatic segmentation of sonograms of the lymph node is accomplished under the condition that the four marks are provided by a radiologist. Both the qualitative and the quantitative evaluation of the segmentation results indicates that the segmentation quality is obviously improved with the integration of the edge flow into the GVF snake model for the segmentation of nodes on sonograms.Based on the segmented contours of the nodes, a computerized scheme is designed to extract the sonographic features. Totally 10 kinds of sonographic features are quantified including gray scale sonographic features related to size, margin, nodal border, shape, medulla ratio, medulla distribution, echogenicity and echogeneity, and power Doppler sonographic features related to vascular density and vascular pattern. For each of 10 kinds of sonographic features, correlations between the quantified feature parameter and two radiologists' consensus grading are computed to assess the values of these parameters in the diagnosis. The most correlated parameter for each of 10 kinds of sonographic features will be provided to the radiologists to help them evaluate the cervical lymph nodes on sonograms.Finally, the rough set notion is incorporated into the support vector machine (SVM) to deal with the overfitting problem due to the noises or outlies in the training set. A rough margin is defined in this dissertation. The optimal separating hyper-plane is found by maximizing the rough margin. In the learning process, since more training samples are adaptively considered with the rough margin, the effect of noises or outliers may be reduced. The testing results from the three standard databases demonstrate that the generalization performance of the rough margin based SVM is higher than that of the traditional SVM classifier when there exist noises or outliers.Assembling the above-mentioned three steps, a computer-aided diagnosis system is constructed to analyze the cervical lymph nodes on sonograms. Besides the functions of the image segmentation, feature extraction, and classification, with the system the users can manually modify the segmented results, add or delete a sample in the training set, and display the information of the known samples. A total of 210 cervical lymph nodes (including 96 benign nodes and 114 malignant nodes) on sonograms have been analyzed and differentiated by the developed system. The system produced the accuracy, sensitivity and specificity values as 84.29%%, 84.21%, and 84.38%, respectively. Compared with the performance of the accredited radiologist (with the accuracy, sensitivity and specificity as 78.57%, 80.70% and 76.04%, respectively), the developed system has the potential to improve diagnostic accuracy of the radiologists in the task of distinguishing between malignant and benign cervical lymph nodes on ultrasonography.