Automatic Segmentation Method For Brain Tumor Based On Multi-Modal MR Image Fusion

Brain disease has become one of the most important diseases that endanger human healthy,among which brain tumor is attracting more and more attention due to its high incidence and high risk.Magnetic resonance(MR)imaging is a non-invasive,non-radiation and multiparameter imaging method.Since images of different MR modalities present characteristics of brain tissues and tumors differently,the use of multi-modal images to segment brain tumor regions and provide statistical information of tumors such as shape and size is a main research direction.However,due to the image noises,the gray deviation and the uncertainty of the tumors’ characteristics and locations,the brain tumor segmentation of MR images is more difficult than other medical image segmentation.To this end,taking the detailed segmentation of different brain tumor regions step-by-step as the main clue,we study the semi-automatic and automatic segmentation methods in the context of the application of brain tumor MR images.The main contributions of this paper are as follows:Aiming to segment the brain tumor lesions as a whole,we propose a segmentation method based on the joint constraints of low-rank and sparse by taking use of the multi-modal MR images.Compared with single-modal segmentation,the segmentation results based on the multi-modal fusion image are proved to be more effective.In the method,the brain tumor image is modeled by the joint representation of low-rank and sparse and the whole brain lesion regions including brain tumors and edema are regarded as image abnormalities which exist independently of healthy brain tissues.Then,the brain lesions can be obtained by solving the model residual.Since the low-rank representation is conducive to describing the overall structure of image and the introduction of the sparse term is beneficial to keep the local features of brain tissues,the combination of the two improves the representation accuracy.In addition,segmentation of brain tumors by means of abnormal detection eliminates the needs for clinical priors of brain tumors,thereby providing ideas for batch detection of brain diseases.Aiming to segment the whole brain tumor,we propose a multi-modal segmentation method based on graph-cut and Softmax regression.The segmentation results are compared to establish the way of multi-modal fusion and feature fusion.Since the image spatial features are not taken into account,the classification results based on Softmax regression are sensitive to noise and grayscale deviation.Therefore,the prediction probabilities of pixel belonging to each category obtained by Softmax are introduced to the graph cut model as regional energy,so that it can be optimized under the spatial constraints of the boundary term.The combination of the pixel characteristics and the neighborhood information ensures the segmentation completeness of brain tumor.In addition,the problem that the graph cut method depends heavily on the features of manually marked points can be solved.Aiming to extract brain tumor and edema respectively,we propose a brain tumor segmentation method based on multi-scale superpixel and multi-kernel collaborative representation classification.Due to the complex structures of brain tissues and tumors,the spatial features of image based on multi-scale superpixel are extracted to construct the kernel function and the image is modeled by the kernel collaborative representation.The method enables the classification to be performed in the high-dimensional spectral kernel space in which the brain tumor image is linearly separable and the constructed spectral-spatial kernel improves the accuracy of pixel similarity measurement.Moerover,the spatial features based on the multi-scale superpixels are in line with the structures of brain tissue and tumor,which is conductive to maintain the local features such as edges and details in the image.Aiming to extract enhanced tumor,necrosis and edema respectively,we propose two brain tumor segmentation methods based on the multi-kernel learning and the superpixel kernel lowrank representation as well as the multi-kernel learning and the superpixel kernel low-rank sparse joint representation.In order to overcome the fuzzy boundary of brain tumor and the gray deviation,the image is modeled by the kernel low-rank representation and the kernel lowrank sparse joint representation in the high-dimensional feature space induced by kernel function.The method improves the accuracy of image representation by combining the kernel method with the advantages of low-rank representation and sparse representation in maintaining image structure.In addition,the optimal superpixel kernel generated by the multi-kernel learning learns high-dimensional manifold features of samples of each category adaptively.