Research On Segmentation And Prognosis Model Of Glioma Based On Deep Learning

Glioma is a collective term for tumors derived from glial cells and neuronal cells of the nervous system.It is the most common primary brain tumor caused by the canceration of glial cells in the brain and spinal cord,accounting for 35.26 %～60.96% of intracranial tumors.Glioma is difficult to treat and easy to relapse,and the mortality rate of patients is also high.The use of multimodal MRI images can effectively find the glioma lesion area,but at the same time,a large number of MRI sequence images will be generated,making the efficiency of artificial segmentation of the glioma area very low.The segmentation of glioma based on MRI helps doctors conduct quantitative analysis of medical images.At the same time,accurate segmentation of the glioma lesion area is also conducive to improving the efficiency of the computer-aided diagnosis system.For example,using deep learning technology to construct a prognostic model for glioma patients,so as to accurately identify individuals with glioma patients.The help provided by chemical treatment is also conducive to prolonging the survival time of patients and improving the quality of life.Therefore,the precise division of different areas of glioma and the construction of a prognostic model using the lesion area have become the focus and difficulty in the computer-aided diagnosis of glioma.This thesis focuses on the use of deep learning technology to segment brain gliomas and the construction of prognostic models to discuss and research.The main contents of the research include:(1)Reliable segmentation of glioma tissue from multimodal MRI is of great clinical value.However,due to the complexity of the glioma itself and surrounding tissues and the blurred boundaries caused by infiltration,the automatic segmentation of glioma is difficult.To solve this problem,this paper proposes a 3D U-Net++ glioma segmentation network based on fusion loss function.First,perform preliminary processing on the MRI data of the four modalities for smoother deep learning network training;then the 3D U-Net ++ network is used to segment different regions of the glioma,which uses different levels of the U-Net model performs dense overlapping connections,uses the four alternating output results of the network as deep supervision to better combine the deep and shallow features for segmentation;and combines the dice loss function and cross entropy as the fusion loss function.Compared with other methods,the method in this article effectively improves the segmentation accuracy of tumor regions,especially the segmentation accuracy of small areas.(2)Aiming at the problem that the amount of network parameters in the 3D U-Net++network is too large,and the way to obtain different levels of features is relatively indirect,this article uses the 3D U-Net3+ glioma segmentation network,which changes the 3D U-The overall structure of the Net++ network,after each down sampling in the encoder stage,the decoders of different levels in the decoder stage are directly connected,and each decoder stage will also be densely connected.The loss function also uses a loss function that is a fusion of the Dice loss function and the cross-entropy loss function.The use of deep supervision can improve the segmentation performance of the 3D U-Net3+ network to a certain extent.Compared with3 D U-Net++,3D U-Net3+ can effectively improve the segmentation accuracy of each region in glioma,and the amount of network parameters is also significantly reduced.(3)The prognostic model for glioma needs to be constructed using a large amount of individual patient data,and when the data is too complex,traditional statistical techniques cannot effectively and quickly construct the model,deep learning technology has unique advantages.In the construction of the prognostic model of glioma,this article uses a convolutional denoising autoencoder to extract features,a reconstruction loss function for feature extraction,and a cox proportional hazard regression model as a loss function for feature extraction.The optimization has greatly optimized the extraction of the predicted features for the survival period.Use Kaplan-Meier,time-dependent AUC and other methods to verify the effectiveness of the model.Finally,the consistency index(C-Index)is used to evaluate the prognostic model.The experimental results show that the method in this paper has higher accuracy than other methods.In this study,aiming at the automatic segmentation of glioma from multimodal MRI,3D U-Net++ and 3D U-Net3+ networks using mixed loss functions were constructed,and the segmentation performance comparison between different branches in the network and between networks was performed.After obtaining good segmentation results,a prognostic model of glioma based on DAE and proportional hazard regression loss was constructed,and the validity and accuracy of the model were verified.These results not only provide a higher-precision segmentation method for the automated segmentation of glioma,but also provide a model with better prognostic performance using multimodal MRI data of glioma,which can help doctors treat glioma Tumor patients can better perform prognostic diagnosis and provide more accurate individualized treatment.