Research On Brain MR Image Segmentation Algorithm Based On Deep Neural Network

The analysis of magnetic resonance imaging(MRI)has gradually become the mainstream trend of brain image analysis.A large number of clear image data not only provide accurate and detailed diagnosis information for doctors,but also increase the number of their tasks and visual fatigue.Therefore,it is very important to process and analyze brain MR images automatically and semi automatically,and then to diagnose diseases by computer.Because the low resolution of brain MR image,the low contrast between brain structure and adjacent structure,the distribution of brain tissue is uneven,and the edge contour is fuzzy,so the segmentation of brain MR image has some challenges.In this paper,two-dimensional brain MR image and three-dimensional brain MR image segmentation are studied and improved.The specific research work and innovations are as follows:(1)In order to solve the problem of underover-segmentation,the LBP texture prior information and the Hog contour prior information of gray-scale image that are used as the significant prior features of brain MR image which and the traditional gray-scale features are used as the input of neural network together.In a certain degree,it can solve the problem that the number of layers is shallow,the logic structure is simple,and the traditional neural network can not extract the deep-seated information.At the end of this paper,an improved convolutional neural network combined with prior information is proposed to significantly improve the segmentation performance of 2D brain MR images.Compared with the original CNN,the LBP texture priori method increased 2.7%,2.9%,3,1%,and IOU by 2.8%,5%,and 5.1%respectively in white matter,gray matter,and cerebrospinal fluid,and the hog contour priori method increased 2.8%,3%,4.6%,and IOU by 2.3%,5.8%,and 7.5%respectively in white matter,gray matter,and cerebrospinal fluid.(2)In view of the fact that the gray value of brain structure of MR brain image is close to this feature,the method of contrast limited adaptive histogram equalization is used to enhance the image data,so as to improve the quality of MR brain image.At the same time,the traditional residual u-net is applied to the field of MR brain image segmentation.According to the feature that the distribution of brain structure is related to the location,based on the residual u-net structure,combined with the spatial domain attention mechanism,more attention will be paid to feature areas that are more helpful for task segmentation in the process of training.In addition,in order to pay more attention to the important features in the training process,channels are added in the training process.Domain attention mechanism module.Compared with the original residual u-net,the improved u-net algorithm based on spatial domain and channel domain attention mechanism is 7.4%higher in average dice,7.5%higher in IOU,3.5%higher in average dice and 4.4%higher in IOU.(3)In view of the complexity of head and edge structure of brain structure in brain MR image,the weight of the brain structure and its edge pixels is increased in the training set,and the network is forced to learn how to segment the edge part of brain structure,so as to improve the accuracy of the whole brain structure segmentation.Secondly,we use the Inception feature extractor to extract the multi-scale features of the image,and get the more easily statistical correlation feature information.Aiming at the problem of over fitting and too many parameters that may appear in the common convolution structure,we use the residual structure and the depth separable convolution structure in the network structure.Finally,according to the characteristics of brain MR images,we combine the feature information of different stages together to shuffle the channels,and get the enhanced information features which contain the deep and shallow level information at the same time,and then add them to the network for training.The input feature information of each stage is richer,and the speed of learning features is faster,and the convergence is faster,and the segmentation performance of the network is significantly improved.The final experimental results show that the results of IB SR data set show that the method proposed in this paper is better than that of CNN,and in the part with complex edge,the segmentation effect is better.Dice and IOU values are increased by 0.9%-6.6%,1.3%-9.7%respectively.In IBSR,Hammer67n20,LPBA40 three data sets,the segmentation results of the method proposed in this paper are better than other mainstream methods.