Research And Implementation Of Liver Cancer Identification Based On Improved SVM Model

Liver cancers are the most serious malignant tumors in China, and the mortality rate the annual growth rate has reached the top in the world. Early diagnosis and treatment is one of the most important methods to reduce the mortality rate of patients. At the early stage of liver cancer diagnosis, CT images are widely favorite for doctors because of their high resolution, minor injuries to patients and because they can locate the liaisons visually and correctly. However, the numerous CT images are also a huge burden for radiologist. Sometimes, even experienced doctors may also make wrong diagnosis because of fatigue or other subjective factors. For this reason, the medical profession threw out an urgent demand on the computer aided diagnosis (CAD) technology which can make quantity analysis and give further reference advices. CT-Based CAD is a hot point and difficult issue in the CAD field.By widely consulting domestic and foreign research achievements on liver cancer CAD, this thesis made a deeply study and research on some key algorithms on feature extraction, feature selection and classifier design, and then proposed a new approach of liver cancer CAD based on abdomen CT images. This approach is consisted of four steps:region of interests (ROI) selection, feature extraction, feature selection and classifier design. In the first step, we convert the format of CT images from DICOM to BMP, then we identify the ROIs. In order to acquire complete and accurate information, we take various methods for extraction of texture features. And to reduce redundant features, we also use genetic algorithms for feature selection and make the output as the input of classifiers. Finally, on the basis of research on support vector machine (SVM) classification technologies, we design an improved series SVM network model, which is serially connected by three SVMs, and can extend the performance of SVM on multi-classification.We made some liver cancer identification experiments based on actual medical images using the approach proposed in this thesis. The result shows the feasibility of the approach, the effectiveness of feature extraction, and the accuracy of liver cancer identification based on improved SVM network model. We also made a comparison with Bayesian classifiers, BP neural network, and traditional SVM, which shows the apparent advantages of our approach. The identification achieved a satisfying result and can give reference advices to doctors for liver cancer diagnosis.