MR Imaging Diagnosis And Computer Aided Diagnosis Of Hepatic Lesions

Primary hepatocellular carcinoma (HCC) is a malignant primary liver cancer, and HCC usually occurs as a complication of cirrhosis. Small HCC detected from cirrhotic background makes the successful surgical treatment of HCC and cure being possible. Fibrosis is an early stage of cirrhosis and fibrosis is expected to recover through medicine treatment if it is diagnosed accurately. In recent years, Magnetic resonance imaging (MRI) has been one of the most sensitive methods for diagnosis of cirrhosis and HCC. But it is difficult to diagnose of fibrosis by MRI because morphological changes of fibrosis are often mild. Successful application of novel MRI contrast agent-superparamagnetic iron oxide (SPIO) improves the ability of differential diagnosis regenerative nodule (RN) from HCC nodule, but the SPIO-enhanced imaging can not show the hemodynamic characteristics of hepatic lesions and the false positive rate of SPIO-enhanced imaging is high relatively. It is difficult to diagnose RN from HCC nodule by MRI now.For the problems of imaging diagnosis of fibrosis and hepatic nodules correlation cirrhosis, this reseach integrates animal expetiments and clinical imaging with imaging processing and imaging analysis technology. Possessing advantages of multiple subject crossing, it carries through the investigation of MR imaging diagnosis and computer aided diagnosis of hepatic lesions. Firstly, this dissertation studies the diagnostic value of SPIO-enhanced imaging and combined SPIO-enhanced and Gadolinium DTPA (Gd-DTPA) enhanced imaging for differential diagnosis of RN and HCC nodule through establishment of rat HCC model. Secondly, a neural network (NN) classifier based on texture feature is designed to distinguish rat hepatic nodule in SPIO-enhanced imaging. Thirdly, clinical patient cirrhosis MR imaging is classified by the NN classifier based on texture feature. Lastly, the initial value of the NN classifier for clinical patient cirrhosis MR imaging is optimized by Ggenetic algorithm (GA) to improve the performance of the classifier. The main research work and contributions of this dissertation can be summarized as follows:(1) Analysing rat RN and HCC nodule SPIO-enhanced imaging and combined SPIO and Gd-DTPA enhanced imaging and the ability of rat RN and HCC differentiated diagnosis is improved. Rat HCC model is induced by low dose diethylnitrosamine, a total of 106 nodules including 24 RN and 82 HCC are pathologically confirmed. SPIO-enhanced imaging and combined SPIO and Gd-DTPA enhanced imaging of rat 106 modules are analysized compared with pathology. The results show SPIO-enhanced imaging contrast noise ratio (CNR) of HCC nodule is higher than the pre-SPIO enhanced imaging (P< 0.05), SPIO-enhanced imaging CNR of cirrhosis tissue has not significantly changed (P>0.05). The sensitivity of the combined SPIO and Gd-DTPA imaging (96.34%) is higher than the sensitivity of Gd-DTPA imaging alone (89.02%), but the statistical difference is not significant (P>0.05). It is concluded that SPIO-enhanced imaging can improve rat HCC nodule detection and diagnosis capability compared with T2WI, and SPIO-enhanced imaging is a useful complement to the Gd-DTPA enhanced image for rat RN and HCC differentiated diagnosis.(2) Classifying rat and patient hepatic MR imaging using BP NN classifier based on texture feature and higher classification rate is obtained. Firstly, rat hepatic edge is automatically segmented by ISODATA algorithm, Otsu threshold algorithms and Iterative threshold algorithms on T1WI. The result shows that ISODATA algorithm is suitable for unknown category imaging automatical segmentation, and it is preparative for lesion region of interest (ROI) searching next step. Secondly, ROI of rat RN and HCC nodule SPIO-enhanced imaging and patient cirrhosis and normal liver parenchyma T1WI are cut manually and feature characteristics are extracted by Gray Level Coourrence Matrix (GLCM). Feature extraction program based on texture is chosen after comprehensive comparision. Six texture characteristic parameters including Angular Second Moment, Contrast, Correlation, Inverse Difference Moment, Entropy, and Variance are extracted according to the medical experience. For rat RN and HCC nodule SPIO-enhanced imaging, six texture characteristic parameters which are significantly different between two classes are chosen after statistical analysis. The result is same as clinical doctor's subjective visuality. Then a BP NN classifier is designed and classified.The independent test classification accuracy of the BP NN classifier is 91.67%. For patient cirrhosis and normal liver parenchyma T1WI, four texture characteristic parameters which are significantly different between two classes are chosen after statistical analysis. Then a BP NN classifier is designed and classified.The independent test classification accuracy of the BP NN classifier is 87.60%. It is concluded that the BP NN classifier based on texture feature is suitable for classification rat RN and HCC nodule SPIO-enhanced imaging and patient cirrhosis and normal liver parenchyma T1WI, but classification accuracy of the patient cirrhosis and normal liver parenchyma T1WI is low relatively.(3) For classification clinical patient cirrhosis T1WI, increasing the number of the texture feature parameters and optimizing the NN classifier by GA. Firstly, direction of GLCM is improved from one to four for extracting texture feature and 14 texture characteristic parameters of every direction are determined. Then feature is chosen by Box Plot. Among the 54 texture feature parameters,24 texture feature parameters which are preferably separability between cirrhosis and normal liver tissue are chosen. In the case of more input texture feature parameters, the BP NN classifier is optimized by GA and GA-BP algorithm is obtained. GA-BP algorithm is a mixed algorithm that the initial value of the BP NN classifier is optimized by GA. On ensuring the classification accuracy, stability and average convergence rate of GA-BP algorithm classifier is improved and the independent test classification accuracy of the GA-BP algorithm classifier is 95%.