Brain Magnetic Resonance Image Segmentation Based On Supervoxel And Fully Convolution Network

The brain is one of the most important organs in the body,and its health has always been the focus of attention.With the development of medical imaging technology,Magnetic Resonance Imaging(MRI)has become an important tool to diagnose the condition of the brain.The correct segmentation of brain MRI images is of great significance to the diagnosis of diseases,the localization of lesions and the treatment of diseases.However,it is a timeconsuming and tedious task to divide the MRI images manually.Automatic and accurate segmentation of brain MRI images has become an important topic in medical auxiliary diagnosis and treatment.The Convolutional Neural Network(CNN)is an important part of the deep learning field.It optimizes the distribution from the data by the optimizing algorithm,and then applies the distribution to the target data to accomplish specific tasks.Traditional CNN cannot be applied directly to image segmentation,until Fully Convolutional Network(FCN)implements the endto-end classification.However,the end-to-end convolutional neural network model such as FCN cannot avoid the loss of detailed features during network transfer such as downsampling,and these details cannot be recovered easily while upsampling.The texture of the brain tissue is complex,and each tissue is distributed and dispersed with many details.The loss of detailed information in the convolutional neural network will affect the segmentation of the details in brain tissue.In order to overcome the deficiencies of the end-to-end CNN which is similar to FCN in the processing of MRI image segmentation,this thesis designs a 3D brain MRI image segmentation algorithm that combines supervoxel and convolutional neural network.The classical Inception module can provide denser feature information through four parallel branches,which can effectively compensate for the loss of excessive detail information in FCN.Therefore,the Inception module is introduced into FCN.In addition,since the supervoxel has characteristics such as boundary fitting and homogeneity,it can provide border and detail information very quickly and conveniently.Therefore,in the three-dimensional space,the detailed information provided by the super voxel and the local and global features learned by the convolutional neural network are combined to achieve effective segmentation of the brain MRI image.For assessing brain MRI image segmentation algorithm proposed in this thesis,we chose the general IBSR18 and BrainWeb20 experiment data sets.The effects of the number of supervoxel and the parameters on the performance of the algorithm are investigated in this thesis.The performance evaluation of the algorithm adopts four evaluation indexes: Dice,Mean IU,AVD and Hausdorff.Experimental results shopaw that the proposed algorithm can achieve better segmentation results.