Research On Key Techniques In Multi-phase CT Image Based Computer-aided Hepatic Lesion Diagnosis System

The incidence of liver cancer and cancer mortality rates ranks the top 2 of cancer spectrum. The liver cancer patients at early stage often do not have the appropriate signs and symptoms, which always cause the fact that lesion have already been developed into middle or late stage when clearly diagnosed. Currently the most authentic diagnosis of liver cancer is still through tissue biopsy, which would give patients great physical and psychological sufferings. The traditional methods to differentiate normal liver tissues from abnormal ones are largely dependent on the radiologists'experience. Thus Computer-Aided Diagnosis (CAD) system based on the image processing and artificial intelligence techniques gain a lot of attentions, since they could provide constructive diagnosis suggestions to clinicians for decision making.The use of image analysis and image quantity techniques has already been proposed in the discrimination of hepatic tissues. Although a lot of endeavors have been devoted into the discrimination of liver tissue characteristics, the result of current systems is still not satisfactory enough, for the particularity of liver diseases. The lack of shape-prior information makes liver CAD system is not as popular as other CAD systems like breast nodule detection or lung nodule diagnosis systems.Comparing to precontrast CT images, Multi-phase CT image can provide more information for the description of lesions, because it carry the timing characteristics. We use multi-phase CT image as the image medium of our liver diagnosis system. Besides using image texture metric as the feature vector, we also designed a temporal and sacttergram-based lesion enhancement pattern descriptor to quantify the different lesions. In the designing of classifier module, we convert a 4 classes classifying problem into 3 binary classify problems by using a hierarchy layer classify architecture. On selecting the specify classify algorithm, we applied 3 different approaches: artificial neural network classifier, k nearest neighbor classifier and support vector machine classifier. Also we analyzed the diagnosing effect of these classifier algorithms to choose the best fit one. Finally, we obtained the best classification accuracy of 0.955, 0.972 and 0.964 for normal-abnormal, cyst-otherdisease and carcinoma-haemangioma sub problems respectively.The main works and innovation points of this thesis are described as below: 1. Different with the commonly approach that using precontrast CT image as the only diagnostic image medium, we adopt multi-phase CT image when developing our liver lesion CAD method. And we applied image texture analysis as metric to quantify the liver disease.2. We designed some temporal and scattergram-based features to further describe the liver disease enhancement pattern.3. A hierarchy layer classify architecture is adopted to convert a 4 classes classifying problem into 3 binary classify problems to achieve better classify accuracy and stability.4. Three different algorithms are used to do the sub-classify work, and we also analyzed its fitness in the liver lesion computer-aided diagnosis application respectively.