Research On Method And Application Of Multi-modal Magnetic Resonance Brain Image Segmentation Based On Deep Learning

Brain-related disease has characteristics of high incidence,high disabil-ity,and multiple complications.In particular,magnetic resonance(MR)brain images generate multi-modal images that reflect information at different levels of the brain by adjusting imaging parameters.It mainly includes:T1-weighted imaging(T1)reflects brain anatomy,T2-weighted imaging(T2)enhances the signal display of the lesion area,and fluid attenuated inversion recovery(FLAIR)behaves high sensitive to pathological changes.Clinically,MR brain images have become the main tool for the diagnosis of brain-related diseases.Specif-ically,segmenting MR brain images carefully can not only describe the brain structure variation of the brain-related disease,but also determine the degree of disease progression.At present,the segmentation of MR brain images can be divided into three levels:local brain tissue segmentation,brain structure segmentation and lesion segmentation.Deep learning technology has become the mainstream approach in the field of medical image analysis owing to its powerful autonomous fea-ture extraction capabilities and strong flexibility.However,the complexity of the anatomy of the brain,the diversity of images,and the grey unevenness of MR brain images not only result in blurred boundaries between tissues but also lack of sharp outlines between structures,which limits the segmentation accu-racy of MR brain images.In view of the above limitations,this paper analyzes MR brain images comprehensively by researching deep learning technology and improve the accuracy,applicability,and robustness of segmentation,which provides objective evidence for the diagnosis of brain-related disease.The specific research content and main innovations are summarized as fol-lows:(1)A multi-target interactive neural network framework is proposed aiming at the blurred boundary in local brain tissue segmentation.To solve the problem of gradient dispersion caused by model deepening,an interactive neural network(Inter-Net)is designed.On the one hand,identity mapping is introduced to realize the autonomous selection of the receptive field of the net-work at different stages.On the other hand,skip connection is also employed between modules to achieve efficient extraction of characteristics by fusing the high-level and low-level features.In addition,three various loss functions are designed from the perspective of evaluation index,information theory,and data distribution.And they are integrated at the decision-making level to improve the robustness of tissue segmentation.This article validates the proposed model on the public ADNI dataset.Compared with other representative methods,this method solves the problem of the blurred boundary in local brain tissue segmen-tation better than the other representative methods.Furthermore,It achieves the purpose of more accurate brain tissue segmentation.(2)A cascaded convolutional neural framework based on boundary correction is proposed aiming at the problems of blurred gray-white mat-ter boundaries and excessively deep sulci caused by disease.The challenge of segmenting brain structures is attributed to the determination of boundaries between brain structures.A cascading manner is used to construct the coarse-to-fine brain structure segmentation framework by simulating the working pipeline of human vision.Firstly,densely connected fully convolutional neural network(DC-FCN)is proposed for coarse segmentation.DC-FCN takes the whole MR brain images as input for the purpose of making full use of anatomical structure and global position information.Additionally,densely connection block is em-bedded in the classic encoding-decoding architecture to improve the utilization of spatial information.Secondly,the context information of boundary voxels is extracted based on the results of coarse segmentation.The boundary voxels have iteratively corrected the category of boundary voxels to further improve the accuracy of brain structure segmentation.The proposed model not only per-forms in the fold cross-validation on the internet brain segmentation repository(IBSR)dataset,which proves the effectiveness of the boundary correction strat-egy but also participates in the 2013 magnetic resonance brain image segmen-tation(MRBrainS13)challenge.This method ranked second when the results were submitted.(3)A cross attention densely connected full convolutional neural net-work(CA-DCN)is proposed aiming at the variation of spatial location,shape and size of lesions.On the one hand,in view of the spatial multiple-ness of lesions,a cross attention block is designed combing the advantages of spatial domain and channel domain attention.It is embedded in the above DC-FCN to learn discriminative convolution characteristics and enhance the intra-class consistency.On the other hand,the focal Tversky loss function and generalized Dice related loss function are designed based on the Dice simi-larity coefficient for the unbalanced classes.At last,an integrated model is used to optimize the effect of lesion segmentation task,and explore the trade-off between sensitivity and specificity.The proposed method is evaluated on the public 2017 white matter hyper-intensity(WMH2017)segmentation in the fold cross-validation,which shows the effectiveness of multi-modal images and cross attention block.In addition,the method also participates in the online assessment of 2015 international symposium on biomedical imaging segmen-tation challenge(ISBI2015)and performs amongst the top-performing solution at the time of submission.