MRI Image Segmentation Of Brain Tumors Based On Improved Unet And Polyp-PVT

Brain tumor is a very common neurosurgical disease,as a high-risk disease of the brain,it has caused serious harm to human life and health.Inorder to make full use of medical imaging information and help doctors make rapid and accurate disease diagnosis,deep learning models show powerful feature extraction capabilities in brain tumor segmentation tasks.However,due to the various shapes of brain tumors and the variability of brain images,there is still room for improvement in the segmentation effect of brain tumors.（1）This dissertation aims at the problems that the previous Unet network structure is complex,the parameters are too many,and the calculation amount is too large,which leads to the inability to meet people’s requirements for calculation speed and accuracy.Inorder to improve the calculation speed and accuracy,and improve the quality of detailed features in the training process,an Inverted Residuals Block is introduced to replace the convolutional modules in the encoding and decoding stages of the Unet network,and an improved residuals Convolutional Block Attention Module（ResCBAM）is added between encoding and decoding.Combining the above two points,an improved InR-ResCBAM-Unet model is proposed.Based on the Brats2019 data set,this dissertation conducts ablation experiments on the proposed Unet network improved model,and compares it with the TrUE-Net network,ConResNet network and OM-Net network,using the Dice Similarity Coefficient and Hausdorff distance（HD）as the evaluation index to analyze the segmentation effect of the model.The experimental results show that the InR-ResCBAM-Unet network not only improves the running speed of the network,but also greatly reduces the amount of parameters and improves the network segmentation effect.（2）Since a large number of convolution operations are used in the InRResCBAM-Unet network proposed in this dissertation,the network will show limitations in explicitl modeling long-range dependencies,and the network cannot be fully parallelized,which will affect the training speed of the network.Therefore,this dissertation based on the polyp segmentation pyramid vision Transformer（Polyp-PVT）network,it has the advantages of fast running speed and high segmentation accuracy.Using the Polyp-PVT network as the basic model,in order to improve the network’s ability to feature fusion,and completes the information mixing between channels without increasing the amount of computation and parameters,and enhances the segmentation effect,introduces selective fusion attention（ASF）module replaces the Cascaded Fusion Module（CFM）in the Polyp-PVT network,and the Channel shuffle module is introduced before the ASF module.Combining the above two points,a PVT-ASF-CS network is proposed.Based on the Brats2019 data set,this dissertation conducts ablation experiments on the PVTASF-CS network,and compares it with the model TransBTS network,VT-Unet network and InR-ResCBAM-Unet network,using Dice coefficient and HD as evaluation indicators to evaluate the model segmentation effect.analyze.The experimental results show that the improved model of the Polyp-PVT network proposed in this dissertation can extract more and better features,and has a better segmentation effect on brain tumors.