Research On Texture Analysis And Its Related Technologies Of Medical Images Based On Liver CT Images

Focal liver lesions mainly include malignant Hepatocellular Carcinoma(HCC)and benign hepatic Hemangioma(HEM).These two lesions are very similar on CT images,which greatly hinder the diagnosis of doctors.Research shows that CT images contain a large number of potential texture statistics law.Therefore,quantitative texture data analysis can be used to distinguish these two lesions effectively.These studies are collectively called Radiomics studies.Following the Radiomics technology route,this study segments tumor regions and extract regions of interest(ROI)from hepatic CT images.Texture features are constructed from ROI using various texture analysis methods,and texture parameters with significant differences between the two classes of samples are selected from these features.Texture classification model is constructed to realize automatic differential diagnosis of HCC and HEM.The details are as follows:1.For the problem of tumor segmentation with low accuracy in liver CT images.A texture clustering method based on co-occurrence matrix(GLCM)and a region growing method based on local binary pattern(LBP)texture are proposed in this paper.Compared with many classical methods,the segmentation method based on GLCM achieves the best result,and achieves an average segmentation accuracy of 91.76%.Then,two-dimensional ROI is extracted based on the segmentation results,and the three-dimensional ROI of tumors is reconstructed by voxel reconstruction.2.Traditional texture analysis methods are usually carried out on a single CT image,which is difficult to fully express the texture features inside the whole tumor.So in this paper,a multi-scale three-dimensional co-occurrence matrix texture analysis method is proposed to extract three-dimensional texture from ROI,and three kinds of two-dimensional texture analysis methods are discussed,including co-occurrence matrix,histogram and combined wavelet transform.Mann-Whitney U-test was used to analyze and select the features with significant differences between the two classes of samples.The correlation between the commonly used texture parameters and the malignant degree of liver tumors was discussed in depth with pathology.It is concluded that HCC has higher contrast,lower correlation,uniformity and energy than HEM.3.The diagnosis of HCC and HEM depends too much on doctors' experience.So in this paper,the selected texture features are separately or combined to constitute multiple feature sets.Support vector machine and 5 fold cross validation are used to train the texture classification model.The multi-scale three-dimensional co-occurrence matrix feature set of this paper achieves the best experimental results among the four types of texture features mentioned above,which achieved 77.63% accuracy and 0.8 AUC value.In this paper,a multi-dimensional texture feature group is proposed,which combines all the selected features.Its texture classification model achieves 88.19% accuracy and 0.91 AUC-value,which is higher than the methods of Chang and Kumar are realized based on data set of this study.4.Based on the above research,an auxiliary diagnosis system of HCC and HEM is implemented in this paper,which can construct ROI visualization model for input CT sequence and automatically give the results of differential diagnosis for HCC and HEM.This system can help doctors to define the scope and nature of the lesion,and help them make early diagnosis based on CT images.