Research On 3DMRI Brain Tumor Segmentation Algorithm Based On Deep Learning

Glioma is the most common primary brain malignant tumor,with high invasiveness and various histological sub-regions.Segmenting the brain tumor area from the multi-modal magnetic resonance scan image can help patients Perform early abnormal diagnosis,prognosis monitoring and treatment evaluation.Due to the inherent high heterogeneity of gliomas,brain tumors are highly heterogeneous and irregular in multimodal MRI images.Therefore,accurate and reliable segmentation of brain tumors remains a challenging task in medical image analysis.In recent years,deep learning has shown good performance in the field of brain tumor segmentation.However,the deep learning-based segmentation method requires large-scale annotation data,and brain tumor data is generally small in size and complex in single-case data.The inherent high heterogeneity of brain tumors itself will cause intra-subclass regions between tumor regions Differences and inter-class differences between tumor areas and non-tumor areas.Therefore,the use of deep convolutional network models to model data needs to be considered from two aspects:(1)The size of the data block input to the network should be as large as possible,in order that the sample data block of a single sample can more fully represent the distribution of data.(2)The complexity of brain tumor data requires a more powerful convolution model to model it,increase the network's ability to understand the data,and better complete the task of pixellevel segmentation.However,in practice,it is often limited by physical computing resources.When using deep convolutional neural networks to model brain tumor data,it is necessary to find a balance between input data size and network complexity.Based on this,designing a lightweight and strong expressive deep convolution to complete the brain tumor segmentation task has become the research motivation of this paper.In response to this problem,this paper explores and designs network architectures for brain tumor segmentation from different spatial dimensions,and proposes a network structure(Brain Tumor Attention Network)BT2Net that combines multiple attention mechanisms and feature fusion mechanisms.In the BraTS19 data set The scores are 0.8931,0.7936,and 0.706.In order to make up for the loss of space,this paper also proposes the(3DBrain Tumor Attention Network)3DBT2Net fused attention mechanism.In the 3D network structure,in order to solve the problem of shallower convolutional layers and smaller experience fields,this article Avoid splitting data into small pieces randomly when inputting data.Instead,divide the data according to prior knowledge,and add a global average pooling layer to increase the network's understanding of the global semantic information of the data,without significantly increasing In the case of memory,increasing the attention mechanism module can accelerate the convergence rate of the network.By designing the Loss function with intra-class and inter-class weights,the effect of area imbalance between multiple target areas on training is improved.In BraTs19 The data tested on the data set shows that our method obtains Dice scores of 0.70,0.85,and 0.80 for the enhanced tumor,the whole tumor,and the tumor core on the detection data set,respectively.