| Hepatocellular carcinoma(HCC)is one of the most common malignancies.It became the third and fifth leading cause of cancerous death in China and worldwide.In actual diagnosis based on magnetic resonance imaging(MRI),experienced medical professional can easily find the lesion area.However,without quantification and comparison,it’s very difficult for them to identify the accurate grade of HCCs.Therefore,in this paper,we quantified the MRI images and analyzed the relationship between the features and the malignancy degree of HCCs.What’s more,a computer aided diagnosis(CAD)system was constructed to provide better diagnostic assistance for doctors.The main contents of the dissertation are as follows:(1)The CAD system for estimating the malignancy degree of HCCs using MRI images was designed,and the basic technology of each module in the CAD system was introduced in detail.In the module of image preprocessing,the regions of interest(ROIs)are segmented manually according to the enhancement performance of HCCs in MRI.In the module of quantification,the extraction principle and calculation of five feature sets are described in detail.In the module of classification,four kind of common classifiers are introduced.The design of this CAD system establish the foundation for the next work.(2)The differentiation between the malignancy degree of HCCs and the features of ROIs was studied.The feature average gray-level intensity(Mean)and gray-level nonuniformity(GLN)of ROIs could reveal the malignant degree of HCCs.The results of one-way ANOVA and ROC analysis proved that most features had diagnostic value for the diagnosis of malignant degree of HCCs.What’s more,Mean and GLN showed the most significant difference and the highest diagnostic value for HCCs.(3)In order to assist doctors in predicting the pathological information of HCC using MRI,an adaptive weighted multi-classifiers fusion algorithm based on hierarchical strategy was proposed.In addition to training set and test set,validation set was adopted to calculate the adaptive weight of each classifier.At first,five sets of texture features were extracted for each ROI.Secondly,five single-layer classifiers were selected in the common classifiers for the 5-layer features.Thirdly,the adaptive dynamic weights of each classifier were calculated using our algorithm.Finally,majority voting procedure was adopted to realize multi-classifiers fusion.The comparative assessment of the various classification architectures shows that combining five single-layer classifiers of different types with a voting scheme,fed with identical feature sets obtained,may result in an accurate system able to assist differential pathological diagnosis of HCCs from MRI images. |
