Design And Implementation Of Brain Tumor Segmentation Algorithms Based On Multimodal MRI Images

Brain tumor is an abnormal tissue hyperplasia in the brain,which seriously affects the normal physiological functions of the brain,even endangers the patient's life.Based on the particularity of brain,the brain tumor treatment program requires to protect the surrounding normal tissues,especially the important organs and functional areas,from damage while treating brain tumors.At present,the main therapy of brain tumor diagnosis and localization is tumor segmentation based on Magnetic Resonance Imaging(MRI).Accurate tumor segmentation results provide a reliable basis for preoperative planning of neurosurgery,ensuring more thorough tumor resection,and guiding the operation process to reduce the damage to normal tissue.Therefore,accurate segmentation of the brain tumor and normal brain tissue has become a key component in the treatment plans.However,the diversity of noise in MRI images and the variety of brain tumor location,shape,texture,and structure,have made accurate brain tumor segmentation a challenging task.The most widely used segmentation methods are time-consuming,laborious,and relies heavily on the expert's professional knowledge and experience.Conventional segmentation methods suffer from the difficulty of designing appropriate handcrafted features.The methods based on convolutional neural network cannot localize tumor boundaries accurately.Since single modal MRI images cannot represent all the information of brain tumor.In order to address these problems,this dissertation investigates the brain tumor segmentation based on multimodal MRI images.The main contents are listed as follows:(1)In order to address the problems of various noise in MRI images,low representative handcrafted features,and limited ability to analyze global features of images,this dissertation proposes a segmentation method based on convolution graph cut.The method first employs image convolution to extract the robust features of tumor segmentation from the complex noise of MRI,obtaining the initial segmentation of the tumor.Then,a multimodal MRI brain tumor image collaborativesegmentation graph model is constructed,which transforms the segmentation problem into a minimization loss optimization,which improves the accuracy of segmentation of tumor boundaries and achieves rapid and accurate positioning of the tumors.(2)In order to take full advantage of the complementary information among multimodal MRI images and improve the accuracy of convolution network for tumor segmentation,a brain tumor segmentation method based on multi-channel 3D enhanced convolutional neural network is proposed.This method takes as input multi-modal MRI images and extracts the features of different modals of MRI respectively by convolutional neural network.Then,a mid-term fusion strategy is used to analyze the features of different modals.An enhanced module is employed to enhance the effectiveness of the features in the segmentation process.Finally,experiments have been conducted on the MICCAI BraTS2017 dataset to verify the proposed segmentation method,which demonstrates the effectiveness of the proposed method in comparison with the segmentation results of existing convolution network models.(3)In order to meet the urgent need of the preoperative planning system of neurosurgery for the automatic and accurate segmentation of brain tumors,this dissertation develops a brain tumor automatic segmentation function module for neurosurgical path planning.In addition to the implementation of the above segmentation algorithms,the module integrates the proposed algorithms into the development process of tumor segmentation module of neurosurgery preoperative path planning system based on VTK,QT platform.The segmentation module features automatic segmentation,simple operation,and good visualization effect.In conclusion,in order to address the problem of brain tumor segmentation,this dissertation proposes a multimodal MRI tumor segmentation method based on convolutional graph segmentation and a multi-channel 3D enhanced convolutional network tumor segmentation method,respectively,and the implementation of the two methods for neurosurgery path planning.The work in this dissertation has made some contribution to the theoretical research of tumor segmentation and preoperative path planning,which has great practical value in the diagnosis and treatment of brain tumors.