MRI Segmentation Of Brain Tumors Based On Convolutional Neural Network

As one of the high-risk diseases of the brain,brain tumors,also known as intracranial tumors or space-occupying lesions,has expansive and invasive growth in nowadays.Once occupied a certain size of space in the brain,it will inevitably lead to increasing intracranial pressure,compression of adjacent brain tissues,result in damaging to the central nervous system,which will endanger the life of patients in serious cases.As high sensitivity and multiparameter images,MRI can show the anatomical structure of any section of the human body.It has high resolution of soft tissue and no bony artifacts.Therefore,it has great advantages in the diagnosis for brain tumors.It is a common equipment for doctors to diagnose brain tumors.The use of computer can effectively and accurately realize the automatic segmentation of brain tumors' MRI images,which not only saves doctors' working time and improve the diagnostic efficiency,but also is becoming the focus of researchers in the medical field and image processing.The research of the brain tumor segmentation in this paper is based on the deep learning network structure of convolutional neural network with feature fusion.Inputting the original image,the features of the image is extract firstly.In the process of the processing,the filtering methods include Laplace,Gauss and Gabor.Then the processed image is segmented by convolutional neural network,and then fuse the results of four different features after segmentation.The fusion results are clustered by the Gauss Mixture Model(GMM),and the final results are obtained.This network structure can directly segment the tumor image at the pixel level.Based on the proposed segmentation method,this paper carries out quasi-real experiments of the algorithm,chooses different thresholds for the output value of CNN,chooses the optimal thresholds according to the results obtained,and compares the results of different feature combinations on the basis of determining the optimal thresholds.The data show that the segmentation results of the four feature combinations are more accurate than that of the single feature image,and the three evaluation coefficients are increased in some content respectively.Compared with several common and known segmentation methods,the image segmentation method based on feature fusion is not only more accurate,better performance,but also greatly reduces the training speed of the network.This method has obvious deficiencies while embodying its own advantages.The basic convolutional neural network,namely CNN,is backward in structure and can not effectively simulate the direct dependence between the spatial closure tags.In view of this,this paper improves the segmentation part of convolution neural network,and uses the depth convolution neural network(DCNN)with cascade structure for image segmentation,including three residual layers and the pooling method based on the cavity space pyramid pooling model(ASPP).At the same time,in order to better solve the contradiction between classification accuracy and segmentation accuracy on neural network depth and pooling times in traditional DCNN,the conditional random field(CRF)is adopted the post-process the segmentation results.Similarly,this study also carried out simulation experiments on the improved algorithm.On the premise of consistent external environment,the improved network structure is compared with that before the improvement,and the three evaluation criteria are improved respectively,which confirmed the advantages of this network structure.This experiment also selected comparison network 1 without ASPP structure and comparison network 2 without CRF.The conclusion is validated that the segmentation effect of the deep convolution network combined with ASPP and CRF is better than that of the traditional convolution network.And the advantage of the brain tumor segmentation method combined with feature fusion of DCNN is verified.