Research On Brain Tumor Diagnostic Segmentation And Survival Prediction Method

Brain tumors are classified into malignant brain tumors,which are cancerous and tend to spread to other parts of the brain,and benign brain tumors,which are not.However,in both cases,the growth of brain tumors in the rigid brain space can lead to dysfunctional and even life-threatening conditions in the body.The incidence of brain tumors is increasing year by year and poses a significant public health burden.The primary method of diagnosing brain tumors is to segment the brain tumor area using Magnetic Resonance Imaging(MRI)of the brain,however,segmentation of brain tumors is still performed manually by brain surgeons.Segmentation in this manner takes a lot of time for the brain surgeon to label,and the heterogeneity of brain tumors makes segmentation difficult,resulting in inefficient manual segmentation.In addition,brain tumor segmentation can be used as a basis for predicting the patient’s postoperative survival period,assessing the patient’s subsequent treatment plan,and using the survival period prediction to develop an aggressive or conservative treatment plan.Therefore,in order to assist brain surgeons in brain tumor diagnosis,a systematic study on brain tumor segmentation and survival prediction is conducted in this paper.The current dominant approach on brain tumor segmentation is based on deep learning methods.However,due to the sensitivity and privacy of medical data,there are few existing relevant public datasets with small data volume,and serious over-fitting problems will occur if deep learning methods are used directly.On the other hand,deep learning models involving3 D images have the problem of too many parameters.In this paper,by introducing a training strategy combining knowledge distillation and adversarial training for brain tumor segmentation,we are able to obtain brain tumor segmentation maps with high accuracy.Knowledge distillation pre-trains multiple teacher networks to obtain soft labels,and then soft labels are used to supervise the training of student networks and compress the model volume.Soft labels enable more diverse mapping of student networks,thus alleviating the over-fitting problem.The combination of knowledge distillation and adversarial training makes the adversarial training more effective in supervising the global network and achieving regularization.After that,this paper compares experimentally with the dominant brain tumor segmentation algorithms,and the results show that the present method can obtain high segmentation accuracy on the compressed model.In addition,existing survival prediction methods mainly extract radiological features on brain MRI and segmentation maps,and use machine learning methods to make predictions.However,the manual extraction of surface features is limited by the correlation between radiological features and survival period,which makes it difficult to further improve the accuracy of survival prediction.In this paper,we use neural network to automatically extract depth features and adopt feature fusion method to fuse radiological features and depth features for survival prediction,which further improves the prediction accuracy.In the experimental stage,this paper designs several comparison methods,including the machine learning method based on feature extraction and the deep learning method without feature fusion,and the results show that the method in this paper can get better prediction results.Furthermore,the performance of the brain tumor segmentation model proposed in this paper was laterally validated by comparing the impact of brain tumor segmentation maps on the patient survival prediction task.