The Diagnosis On Breast Ultrasound Image Sequences Combining With Elasticity Features

Nowadays, breast cancer is the most common malignant tumor of women and attracted more and more attention. Breast ultrasound, also known as sonography is frequently used to evaluate breast abnormalities during a physician performed clinical breast exam. A large amount of research shows that, the accuracy of detection using the sonography can be further improved by combining the computer aided diagnosis (CAD) technology. The breast tumor CAD system usually performs the diagnosis based on the supplementary information of morphology and texture features. However, the performance of the CAD system utilizing the aforementioned features is still not so satisfactory to date. Clinical research has demonstrated that elasticity information of the tumor is also a very important indicator to judge whether the breast tumor is benign or malignant. This dissertation studied the performance of a breast tumor CAD system with the particular inclusion of the tissue elasticity features, which were obtained by ultrasound image sequences collected during a continuous free-hand compression process. The objective of the dissertation is the value of the sonography in differentiating benign and malignant breast tumors. According to the conventional processes of a CAD system, the main research work and contribution of this dissertation are as follows:(1) Preprocessing the sequential ultrasound images. An algorithm based on the anisotropic diffusion equation is presented. The algorithm combines the robust estimation and considers the feature of the speckle noise, so it can suppress the speckle noise effectively and be more robust, thus the edge details of the ultrasound image can be preserved even enhanced, which can provide effective safeguard for the following edge extraction. The dissertation proposed a method to compute speckle scale coefficients automatically, which reduces the influence of the human beings, and enhances the stability of the algorithm.(2) Segmenting the breast ultrasound image sequences to get the tumor boundary. This dissertation proposed an improved C-V model, which could avoid the step of re-initialization, thus the speed of segmentation being accelerated greatly. For the breast ultrasound image sequences, the segmentation results were evaluated by using two difference methods. In the first method, the segmentation result of each frame was used as the initial boundary of the next frame. In the second method, the segmentation result of each frame was evolved by an improved C-V model with a shrink factor before it was used as the initial boundary of the next frame. Experimental results show that the second method got higher accuracy to the first method.(3) Extracting the characteristic parameters of the breast tumor image sequences based on the former operation. The unelasticity features are calculated on the average of image sequences which contained more statistical information compared with a certain image. Considering that the free-hand compression process was very difficult to keep a constant speed, a new relative evaluation algorithm was proposed to quantify the relatively compression coefficient between two consecutive images in the image sequences. The dissertation extracted parameters including 16 unelasticity and 25 innovative elasticity features based on relatively compression coefficient from the breast ultrasound image sequences. They can compensate for each other and improve the validity in the diagnosis of malignant breast tumors.(4) Selecting the feature parameters that were extracted in the previous step and using them to aid the classification of breast tumor. The whole model of feature select and classification was based on the distances between clusters analysis, support vector machines (SVMs) and the sequential backward selection method, combined the advantage of Filter model and Wrapper model. At first, some irrelevant features were filtered by analysis of the distances between clusters. Then, it deleted some features in the Wrapper model by the sequential backward selection method one by one until the predefined dimension. The determining criterion of this model is the accuracy of the features in support vector machines (SVMs). In conclusion, the best combination of features which are selected in the dissertation is used to classify the cases in SVMs. The radial basis function (RBF) and linear basis function (LBF) are used as its kernel function. In the testing process, the very high accuracy diagnosis is achieved by combining two classes of unelasticity feature and elasticity feature.Based on the above steps, 280 pathologically proven cases including 112 malignant tumors and 168 benign ones were tested in this study. The experiment results showed that the performance of the CAD system which used the unelasticity features and elasticity features jointly was much better than the system which only used the unelasticity features or elasticity features. Therefore, it was concluded that the features based on the breast ultrasound image sequences performed well in the classification of benign and malignant tumors, and it could be used as supplementary information for the clinical diagnosis of breast tumors.The research was supported by Provincial Natural Science Foundation of Anhui (2006KJ097A).