Feature Extraction And Segmentation Of Cerebral Tumorous Tissues In MR Images

In real magnetic resonance (MR) brain images, tumors may vary in size, shapeand location. Gliomas are usually accompanied with edema, and present ratherdiffuse boundaries. Hemorrhage, necrosis and cystic component may also appear inthe tumorous region. These make the segmentation of pathological region andpathological tissues in MR images greatly challenging. However, automaticmeasurement of the volume and its variation of these tumorous tissues is verysignificant for assisting the clinical diagnosis and evaluating the therapy results.Through studying the diffuse growth characteristics of gliomas, the tissuediversity within the tumorous region and the signal intensity contrast of differenttissues in the cerebral region, a two-step segmentation frame is proposed for thesegmentation of different tumorous tissues in this thesis. First the whole pathologicalregion is captured with considering the intensity difference of normal brain tissuesand tumorous tissues. Then, delicate analysis of the signal characteristics inmulti-sequential MR images for different pathological tissues region are performed inthe obtained tumorous region, and further segmentation of different tumorous tissues,including solid tumor, edema, necrosis, cyst etc., can be realized.Three methods are used for the segmentation of the whole tumorous region: (1)A new feature space is constructed by weighting normalized intensity values frommultiple sequences of MR imaging and normalized 3-D coordinate for each voxel.Then support vector machine (SVM) is applied in the feature space to classify thetumorous tissues and normal brain tissues. (2) Based on the signal difference ofdifferent tissues on the different sequence, the relative intensity contrast betweennormal tissues and tumorous tissues can are enhanced by applying an intensityprocessing, named "gap transform" in this thesis to the normalized T1- and T2-weighted images. The the whole tumorous region can be obtained through thresholdanalysis. (3) A segmentation method based on fuzzy connectedness and the featuresimilarity validation on the optimum paths is used on the contrast-enhanced imagefrom the gap transform to realize the extraction of the tumorus region. All of thesethree methods fully utilize the characteristics of the tumorous region and theknowledge provided by neuro-radiologists, and can produce accurate segmentation oftumorous region with little human interaction.Nearest neighbour (NN) algorithm and Fuzzy c-means (FCM) clustering areused for the further segmentation of each pathological tissue in the tumorous region.The results of NN algorithm depend on the selection of the training data, while FCMclustering, which uses the fuzzy membership to describe the possibility of eachsample belonging to each kind of tumorous tissues, can achieve more success.The two-step segmentation frame is evaluated by applying it to the real MRimages of three patients. The experimental results prove that the automaticsegmentation of the two-step framework can produce very satisfactory results, whichcorrespond well with the manually labeled results by neuro-radiologists. Moreover,this method also builds an excellent open framework for the segmentation of allpossible pathological tissues in the abnormal region with any available imagesequence providing additional information.