Research On Brain Tumor Segmentation Algorithm In MR Image

Early diagnosis of brain tumors plays an important role in improving the possibility of treatment and improving the survival rate of patients.But brain tumors are difficult to diagnose by invasive examination.Therefore,medical imaging has become the most direct way to assist brain tumors and brain tumors.In various imaging methods,MRI,as a widely used medical imaging technology,has high resolution and no bone artifact for soft tissue.It can generate various sections or even three-dimensional images,which have great advantages in the diagnosis of brain tumors.The segmentation of MR images of brain tumors refers to the segmentation of brain tumor tissue on MR images from normal tissues(including cerebral gray matter,white matter and cerebrospinal fluid),and further segmentation of its substructure(including edema,necrosis,enhancement and non enhancement of tumor nuclei).Segmentation of MR images of brain tumors and their substructures plays a very important role in the diagnosis,treatment and evaluation of brain tumors.With the increasing incidence of brain tumors,more and more MR images need to be segmented.However,manual operation is time-consuming and laborious,and there is no repeatability.Therefore,the research of semi automatic or even automatic segmentation of brain tumor MR image algorithm is of great significance.In this situation,based on the in-depth understanding of the MR image segmentation algorithms for many brain tumors,this paper focuses on the CNN based brain tumor MR image segmentation method,and proposes an improved DCNN method combined with CRF.The critical work of this article has the following points:(1)This paper introduces some common brain MRI tumor segmentation algorithms,including non learning segmentation methods based on threshold,boundary,region,deformation and wavelet transform,as well as fuzzy C meanclustering based on machine learning,Markov random field,support vector machine and random forest.On the one hand,we should grasp the basic principles of these methods,and understand the latest progress on the other hand,and summarize the advantages and disadvantages of various methods.(2)A deep study of CNN for image segmentation: some analysis of the parameters initialization,activation function,pooling,loss function,optimization of parameters and other details,and studies the FCN which is suitable for image segmentation.(3)A segmentation method of brain MRI tumor combining CRF and DCNN is proposed.First,Atrous Spatial Pyramid Pooling structure is used to construct a segmented network structure to obtain multi-scale information;and then the hierarchical structure is segmented through cascaded structures;finally,the relationship information between pixels is obtained through a CRF.Therefore,this method can better get information between multi-scale and pixel.The experimental results show that compared with the classical brain MRI tumor segmentation technology based on CNN,the segmentation accuracy is significantly improved..