Research On MRI Brain Tumor Classification And Segmentation Method Based On Deep Learning

In recent years,with the increasing incidence rate and death rate of brain tumor,brain tumor diseases have caused serious harm to human life and health,Doctors rely on personal experience to diagnose and analyze brain tumor MRI images,which not only has low efficiency,but also leads to artificial subjective misjudgment,Therefore,improving the efficiency and accuracy of computer-aided diagnosis of brain tumors has become one of the research hotspots of brain tumor assisted diagnosis.This paper studies the brain tumor classification and segmentation algorithm through the deep learning method,and constructs an effective auxiliary diagnosis method.The specific work is as follows:Firstly,aiming at the low efficiency of brain tumor classification in traditional methods and the problem of over fitting due to the small amount of data,an initial residual network model is constructed to classify brain tumors.The residual structure is integrated into the concept module,so that the network structure can not only obtain more feature information in depth and width,but also reduce the risk of gradient disappearance.Experiments show that the accuracy of using the proposed initial residual network model to classify three types of brain tumors can reach about 99%,which meets the standard of computer-aided classification of brain tumors.Secondly,aiming at the problem of false segmentation caused by high heterogeneity and complex structure of brain tumors,a 2D brain tumor segmentation algorithm based on U-Net is constructed.Integrate the dense block in densenet network into the U-Net structure,strengthen the transmission and reuse of features between different connection blocks,and integrate high-level semantic information and low-level content information.At the same time,a lightweight attention module CBAM is added to the encoding and decoding structure of U-Net,which uses dimensional and spatial attention to focus on effective feature areas.Finally,the end-to-end training model is formed,and the model test is completed on the 2D slice of brats19 data set to verify the effectiveness of the algorithm.Finally,aiming at the false segmentation problem caused by the loss of spatial structure information caused by 2D slice,a 3D brain tumor segmentation algorithm based on v-net is constructed.The algorithm adds residual connection to speed up the convergence of the model,and replaces the pooling operation of up and down sampling with convolution with convolution kernel of 2;The algorithm introduces the hole space pyramid pooling model,uses the variable expansion rate hole convolution to obtain more scale context information,and uses the overlapping segmentation operation to reduce the loss of each data space feature.Through the model test on brats19 data set,the brain tumor region can be segmented better.