Brain Tumor MRI Automatic Segmentation Method Combining 2D And 3D Networks

Brain tumor is one of the diseases that seriously endanger human health with high morbidity and easy death.Glioma caused by the canceration of glial cells is the most common brain tumor.With the continuous development of medical imaging technology,high-performance computers and deeplearning algorithms,medical image processing plays an increasingly important role in the diagnosis and treatment of brain tumor diseases.Magnetic Resonance Imaging(MRI),as one of the most important imaging methods in medicine,not only has no ionizing radiation in the imaging environment and is non-invasive,but also has the advantages of high imaging resolution and good soft tissue contrast.Accurate brain tumor segmentation through MRI can assist doctors in diagnosis and guide subsequent operations,but traditional manual segmentation methods not only require professional knowledge background,but also time-consuming and labor-intensive.Therefore,the automatic segmentation method of brain tumor combining medical image processing and computer deep learning technology has high research and application value.Starting from the optimization of deep learning network and segmentation method,this paper mainly does the following work:1.We proposed improved Att＿Dense＿UNet based on U-Net.Firstly,Dense Blocks and transition layers are introduced in the encoder part of U-Net to enhance the feature extraction ability of the net.At the same time,a self-focusing pooling method is designed on the basis of the pooling layer to improve the information integration ability of the downsampling module;Inside the Dense Block,group convolution combined with pointwise convolution is used instead of traditional convolution to reduce training parameters and network redundancy as much as possible without losing network representation performance;Finally,in the encoder and decoder A spanning connection with an attention mechanism is realized between them,which alleviates the semantic gap problem faced by the direct splicing and fusion of features at different levels,thereby improving the utilization efficiency of features at each level and enabling the network to perform more accurate segmentation.2.Based on the above Att＿Dense＿UNet,a brain tumor segmentation method combining 2D and 3D method is designed and implemented to reduce the limitations of one single method.First,build the Att＿Dense＿Unet network with the same structure of 2D and 3D,and then train the 2D network with the images obtained from the slice.Then,the 2D trained convolution kernel parameters are extracted,and their dimensions are expanded to three dimensions and used to initialize the 3D network convolution kernel parameters so that they can converge better and faster when training with relatively few 3D samples.At the same time,the region of interest containing the brain tumor target is obtained from the prediction results of the 2D network,and its position information is used to guide the 3D network to perform more refined segmentation of the target region.Finally,the weighted average method is used to integrate the predicted probability maps of 2D and 3D network to obtain the final segmentation result.This paper validates the performance of the proposed network and joint segmentation method by experiments on the Bra TS 2020 dataset.In order to alleviate the serious imbalance of positive and negative samples in actual training,we use a balanced sampling method to generate a training set to include as many brain tumor regions as possible,and use a generalized Dice loss with adaptive class weights and an improved focal loss.The sum is used as the loss function.At test time,in order to eliminate the generalization error caused by inconsistent test and training input sizes,we use the method of partition prediction and re-integration to obtain the final segmentation probability map.Finally,the experimental analysis and comparison results show that the network in this paper has good performance and the combined segmentation method of 2D and 3D network can further improve the accuracy of brain tumor segmentation.