Research On Semantic Segmentation And Overall Survival Prediction For MRI Brain Tumor

Brain tumor,also known as glioma or brain cancer,is one of the most destructive cancerous tumors that can affect the human body.Its growth within the brain parenchyma can cause functional impairment and even life-threatening conditions,resulting in significant impact on the physical and mental well-being of the patients.Magnetic resonance imaging(MRI)provides high-quality images without causing harm to the human body,and is widely used to capture tumor images.However,brain tumor images are highly complex and the boundaries between different tumor cells are often blurred,making brain tumor segmentation challenging.Currently,brain tumor segmentation is still mainly performed manually,which not only results in low efficiency but also causes work fatigue,leading to decreased accuracy.With the rapid development of computer technology today,brain tumor segmentation and survival prediction have received more attention from the academic community.Segmentation of brain tumors can help predict patients’ survival periods and provide evaluation criteria for treatment planning.By predicting survival,doctors can choose more optimistic or cautious treatment plans to better treat patients.Therefore,survival prediction plays a crucial role in the clinical treatment of brain tumors.Accurate brain tumor image segmentation can effectively help clinicians in the early diagnosis of glioma,the formulation of treatment plans,and the follow-up of treatment effects.In this thesis,brain tumor segmentation technology is studied,the shortcomings of existing methods are analyzed,and a neural network for brain tumor segmentation——HPLO-Net is proposed.The network is characterized by small model size,high segmentation efficiency,and good segmentation accuracy.The algorithm optimizes the design of the encoding end,and at the same time incorporates Transformer to model long-distance dependencies,and optimizes and improves it.The test results show that the proposed algorithm achieves better segmentation results compared with some representative algorithms,and the model Significantly reduced complexity and memory footprint.This thesis also conducted research on survival prediction,which can be regarded as the post-task of brain tumor segmentation.In survival prediction,radiological features are usually used as predictors,but using radiological features for prediction can easily lead to overfitting.This thesis processes the extraction of radiological features,uses neural networks to extract deep abstract features,and finally extracts tractographic features based on connectomics.Using these three types of features for survival prediction improves the accuracy and generalization ability of the algorithm.This thesis conducts research on brain tumor segmentation and survival prediction using two datasets,BraTS 2018 and BraTS 2019,from the Brain Tumor Segmentation Challenge.For the segmentation task,the HPLO-Net model proposed in this study achieved Dice scores of 91.41%,81.20%,and 84.36% for the whole tumor,tumorenhanced region,and tumor core region,respectively.The proposed method was compared with mainstream algorithms,and achieved higher segmentation accuracy.For the survival prediction task,the proposed model achieved an accuracy of 64.24%,which is higher than most representative methods.