Feature Extraction And Selection Of Ultrasound Images Of Live Cancer

Since 1990's, Liver cancer has been ranked in the second position of malignant tumors according to the Ministry of Health's Statistics of China. B-mode scanner is the first selection method of popular diagnostics. Ultrasound images have been widespread accepted as an effective diagnostic tool. Affected by the quality of ultrasound images of liver cancer, benign expression of malignant tumors and observer's visual fatigue, careless mistakes, and diagnostic ability of physician, the results of the diagnosis are influenced by some uncertain factors. Thus, it is necessary to provide medical operators a computer aided diagnosis system which is helpful to reduce the possibility of wrong diagnosis or missed diagnosis of liver cancer. We aimed to get a most sensitivity and stable group of features by summing up the features of liver cancer showed on ultrasound images and by mathematical analysis and comparison these features.According to the ultrasound image representation of hepatocellular carcinoma (HCC) and echo type in the tumors, this thesis made large studies on feature extraction and selection, especially on the texture features and shape features. The main achievements of this thesis are organized as follows:(1) Deal with the texture feature extraction, and analyze the texture features of various types of images of liver cancer, which can be defined by their echoes in the tumors. Then, analyze the texture features of all liver cancer and other confused diagnosis (such as angioma, hepatic abscesses). Including to the experiment result, we find one feature is not suitable for discrimination the liver cancer and normal ones. For the different type of echoes in the tumors, there are some respective different features to discriminate them. Thus, if there are enough samples, according to the HCC image representation, making a division of liver cancer will improve the accuracy of classification.(2) Perform the research on shape feature extraction. Due to the result of segmentation of snake methods, this thesis extracts some shape features, and provides doctors certain information about the tumors, such as size, shape, and location.(3) Study some common methods of feature selection. And take Sequential Forward Selection (SFS) method, Relief method and Relief-SFS method to choose the features we have extracted, and then use the SVM classifier to evaluate them. The experiment showed that this Relief method is effect to choose the suitable group of features extracted from images of liver cancer.