Research Of Brain Tumor Segmentation Based On MR Images

The segmentation of brain tumor plays an important role in diagnosis,treatment planning and surgical simulation.Some patients with the most aggressive tumor usually have very short life expectancy,and the development of appropriate treatment planning is the key to improve the quality of patients' life.Magnetic resonance imaging is widely used to evaluate tumors because of its better soft tissue resolution.And automatic and reliable brain tumor segmentation methods are required with the emergence of large amounts of data and the empirical error of manual segmentation.In this paper,we propose a brain tumor segmentation method based on kernel sparse coding and dictionary learning.In this paper,we first introduce the development,the latest research status and the difficulty of the MRI image brain tumor segmentation.Then we introduce and deduce the sparse coding and dictionary learning theory.Considered the non-linear features of MRI images,we present a kernel sparse coding based fully automatic brain tumor segmentation method from 3D FLAIR-Weighted contrast-enhanced MRI images.In order to capture the non-linear features of MRI images,kernel dictionary learning is adopted to learn two adaptive dictionaries using tumor and non-tumor tissues.Finally,the full MRI image was divided into tumor region and non-tumor region,and the performance of our method is analyzed and evaluated.In order to study the effect of different features on the segmentation method based on kernel sparse coding and dictionary learning,this paper proposes an automatic brain tumor segmentation method based on high-order features.In contrast to,the first and second-order features extracted from the voxel and its nearly information are used instead of the gray-level feature as the feature vector.At last,in order to meet the requirements of the BRATS competition and to compare with the latest methods of ranking,we propose a multimodal brain tumor segmentation method in this paper.The multimodal brain tumor segmentation competition(BRATS)provides multiple modal MRI data,which has been linearly co-registered and interpolated to 1mm isotropic resolution,the ground truth images provided by the competition divide the tumor area into five categories and provide 3 different tumor sub-compartements: complete tumor region,tumor core and enhanced tumor region.Due to the fact that different tumor types have different performance in different image modalities,the multimodal magnetic resonance images after preprocessing are used to extract the probable tumor areas using the K-mean clustering algorithm and then the five dictionaries were constructed respectively.In order to evaluate the segmentation performance of this method,the segmentation results are uploaded to an online assessment system,where four evaluation matrixes(Dice Score,PPV,Sensitivity and Kappa)are used to evaluate the segmentation results.There is a high degree of correlation between adjacent pixels and its surrounding areas in MRI images,and the modeling for tumor regions and non-tumor regions using local image patches has been proved to be successful.However,each MRI modality has its different sensitivities to the pathological tissues.There is still a certain correlation between the multimodal MRI image data.Increasing the number of features with multimodal MRI data can theoretically improve the brain tumor segmentation characteristics.Besides,the optimal segmentation effect is achieved when the number of dictionary atoms needed for high-order statistical feature is much less than the method using the gray feature,but more time consuming.Generally,the method that adopts gray feature works better.The experimental results show that the proposed method in this paper has a good performance in brain tumors segmentation.