Research On Brain Tissue Segmentation Based On Deep Convolution Neural Network

Brain magnetic resonance imaging is often used to diagnose brain diseases.Clinically,the physician evaluates brain disease and develops a medical plan by analyzing the size,shape,and location of the brain tissue in the brain MR images.Manual segmentation of brain tissue by experts is of great workload and subjectivity.Therefore,the study of MR images of brain tissue segmentation is of great significance.Two deep convolution neural networks were used to complete the segmentation of brain tissue.In the first network structure,the inputs of the network capture information around the voxel of interest: 3D and multiscale orthogonal 2D intensity patches;the first few network layer use parameter sharing way in which three orthogonal two-dimensional intensity patches extracted by same method share the network parameter,which not only reduces the network parameters,but also improve the performance of network model;in the middle layer of the network,the residual unit is added to ensure that the network can reach deeper,and there is no problem of gradient vanishing.The second network structure is an improvement over the first network structure,which adds an auxiliary classifier to the middle of the first network structure to improve the performance of the model.Both network structures use the softmax classifier,which learns the category information for each voxel,and classifies the voxels in each brain tissue during the test phase to complete the segmentation of the brain tissue.Compared with the traditional method of brain tissue segmentation,the proposed method does not need to non-rigid registration of MR images,and does not need to extract the image features manually.The method proposed is applied to the MICCAI 2012 challenge on multi-atlas labelling.The experimental results show that the two deep convolution neural networks proposed have a better segmentation effect on brain segmentation.