Research On 3DMRI Brain Glioblastoma Segmentation Method Based On Deep Learning

Glioma is a primary brain tumor with high incidence rate and high mortality rate.It can cause great harm to human life and health.At present,brain glioma segmentation based on MRI(magnetic resonance imaging)image can help doctors observe and analyze the external morphology of brain glioma for diagnosis and treatment;however,gliomas have high heterogeneity,and their multi-modality MRI brain images show the characteristics of uneven gray scale and irregular shape,and the imbalance of data samples of gliomas is a long-term problem to be solved.Therefore,it is a challenging work to develop an automatic glioma segmentation algorithm with high accuracy.Aiming at the problem that it is difficult to capture long-distance dependencies in glioma segmentation,an improved glioma segmentation model based on graph convolution network was proposed.The model maps the feature graph from the encoder to the graph data,and makes full use of the multi-scale global reasoning module for global semantic information reasoning,so that the network can better model long-distance dependencies,then the global information is mapped back to the original space,and then decode and up-sampling to realize the accurate segmentation of glioma.At the same time,the sum of dice loss and cross entropy loss is used as the loss function to optimize the model.The loss function can not only alleviate the imbalance between the foreground and background voxels of glioma images,but also accelerate the training convergence speed of the network and maintain the stability of training.In order to further improve the segmentation performance of the network model and enhance the sensitivity to small samples,this thesis further optimizes the glioma segmentation model based on the improved graph convolution network,and proposes an improved glioma segmentation model based on attention mechanism.The network effectively combines the context information of glioma and enhances the attention to the glioma region.At the same time,GDL(generalized dice loss)loss function is used to optimize the model.GDL loss function can assign greater weight to the loss of small target enhanced tumor and balance the differences between different glioma categories,so as to improve the segmentation effect of the network on enhanced tumor.This thesis makes the 5-fold cross validation of the model on Bra TS20 data set.Firstly,the effectiveness of each work is verified by ablation experiments,and then the glioma segmentation method proposed is compared with other mainstream methods.The experimental results show that the average dice scores of the improved glioma segmentation model based on attention mechanism in the whole tumor region,the tumor core region and the enhancing tumor region are 0.91,0.87 and 0.81 respectively,which are 0.67,0.31 and 1.36 percentage points higher than the improved glioma segmentation model based on graph convolution network.At the same time,it is superior to other mainstream methods and can accurately segment glioma.