Research On Brain Tumor Image Segmentation And Survival Prediction Based On 3D Convolutional Neural Network

Brain tumor is a common nervous system disease,also one of the most aggressive tumors.If not treated in time,it will seriously threaten the health of patients.Magnetic resonance imaging technology is widely used in tumor image acquisition because of highly imaging quality and harmless to human body.However,the brain tumor image is very complex,and the boundaries of different tumor subsets have certain fuzziness,which makes tumor segmentation very difficult.Now,the main segmentation method is manual segmentation by doctors,which is inefficient and prone to errors.In addition,if we can predict the survival time of patients with brain tumors accurately,doctors can make better clinical decisions and more appropriate treatment plans.At present,the survival time is mainly predicted by various scoring prediction models developed by hospitals or scholars.These models are easy to use,but lack of scientific in setting score and selecting features,also the prediction accuracy is generally low.Therefore,how to use computer methods to achieve automatic segmentation of brain tumor images and predict the survival time of patients has always been a hot research direction in the academic community.To solve the problems of brain tumor segmentation,this paper designs a multi-path TS-U-net model on two-dimensional images.This model has the following innovations: effective data expansion method,weighted cross entropy loss function and multi-path attention mechanism.These innovations solve the following problems respectively: data shortage,brain tumor image class imbalance and multimodal data loss respectively.Traditional 3D u-net,mixup based 3D u-net and total variation based 3D u-net models are designed on 3D images to improve the segmentation accuracy of the network.It is proved that a fully trained 3D convolutional neural network can achieve high segmentatio n accuracy by using some data enhancement and denoising methods,without modifying the model structure.Compared with other most advanced brain tumor segmentation models,the 3D model designed in this paper has achieved a certain competitive segmentation accuracy.Aiming at the problem of lifetime prediction,this paper validates several machine learning models and designs a lifetime prediction model based on feature selection.This model takes tumor segmentation information and age as input features,and s elects the features according to the influence factors of the model,so as to avoid over fitting of the model and improve the prediction accuracy.The data set used in this paper is from the brain tumor image segmentation2019(Brats 2019),which was held by the MICCAI(medical image computing and computer assisted intervention,MICCAI)conference.In the task of brain tumor segmentation,the dice coefficient of 3D TV denoising U-net model in three brain tumor sub regions are 0.90,0.877 and 0.846.Com pared with 0.91,0.867 and 0.823 of the first place in the Brats 2019 segmentation challenge,the segmentation accuracy in the second and third sub regions is improved by 0.01 and 0.023 respectively.Also,in the task of survival prediction,the accuracy o f the model based on feature selection is 0.59,which is very close to 0.61 of the second place(The first place in the competition was incomplete)in the brats 2019 prediction challenge.The above results proved that the tumor segmentation model and survival prediction model designed in this paper are very effective.