Research On Brain Tumor Segmentation And Classification Of MR Image

The doctor's artificial diagnosis is the traditional brain tumor diagnostic method. But this approach will be affected by the doctor's subjective judgement. Automatic classification of brain tumor of MR image need higher accuracy as it can affect the tumor treatment options and the best treatment time. With the developing of CAD, the technique of brain tumor segmentation and classification has got excellent application and development. The main purpose of this paper is to study methods of the accurate segmentation and precise pattern classification of brain tumors. On this basis, we establish computer aided diagnosis system for dealing with brain tumor image.First of all, we studied a method of extracting brain tissue of MR image. Firstly nonlinear parameter estimation is used to match gray level histogram. Then we determine the size of gray threshold of background and skull, and the initial segmentation image by separating skull and brain tissue was obtained. By combining with morphological principle and the knowledge of convex hull,we use the convex hull to surround the upper half of the initial segmentation image, and it is matched with the initial segmentation image.Thereby we obtain the brain tissue without brain skull and background.This laid a good foundation for brain tumor segmentationand and its classification.Secondly, we researched the algorithm of brain tumor segmentation, a method based on level set and FCM which is applied to the extracting brain tumors was proposed.We make necessary improvement of FCM algorithm, adding the spatial information based on phase consistency principle to the FCM algorithm. It can overcome the effects of brain image noise effectively. For malignant brain tumors, because its gray scale information is complex, we use the convex hull to surround the entire brain tumor region; it is used as the initial segmentation results. The parameter values of level set and the initial outline were determined based on the result of fuzzy clustering. It can avoid bad result of segmentation because of the improper parameter initialization.The result of brain tumor image segmentation is the foundation of the brain tumor classification.Finally, by analysing the feature of the MR images, we studied the feature extraction of the ROI. We analyzed characteristics of meningioma and glioblastoma from gray feature, texture feature and morphological feature. In order to meet the criteria of characteristics we proposed the PCA algorithm which is used to reduce the dimension of feature data. We used the feature data of reducing dimension to replace the whole feature data. The model of PSO-SVM is chosen as the classifier. The standard PSO algorithm with adaptable inertia weigh is introduced into the SVM kernel parameter optimization.We studied PSO-SVM algorithm with the single feature and combined features to classify the brain tumor. Compared with other algorithms, the experiment results showed that the algorithm of this paper get higher accuracy and better result.