Research On Brain Tumor Segmentation Algorithms For MRI Via Deep Convolutional Neural Network

Brain tumor is a kind of cancer with high mortality in human brain.Early detection of brain tumors plays an important role in disease prevention and follow-up treatment.Magnetic Resonance Imaging is a non-invasive and radiation-free imaging technology,which has the advantages of high contrast of soft tissue and becomes the mainstream imaging technology for the diagnosis of brain tumors.With the advent of the era of medical big data,the booming artificial intelligence provides strong technical support for the rapid,accurate and automatic segmentation of brain tumor lesions,reduces the burden of doctors manually marking tumor areas,and greatly improves the efficiency of tumor detection.The existing mainstream methods usually use machine learning theory and technology to construct the automatic segmentation process of brain tumor.However,using machine learning method needs to extract a certain pattern which can represent the data distribution from the massive source data.The quality of patterns directly affects the final performance of segmentation model.To design an excellent pattern extraction method manually requires a wealth of interdisciplinary experience and knowledge.This hinders the popularization of artificial intelligence technology in the medical industry,and too many human intervention pattern extraction methods often have certain limitations.In view of the above problems,the idea of deep learning integrated with autonomous pattern extraction and analysis provides an important research direction for the automatic segmentation of brain tumors.However,the problems such as uneven intensity,noise interference,weak contrast,irregular appearance and different size of tumor focus in Magnetic Resonance Imaging data brings challenges to brain tumor research based on deep learning technology.This thesis takes Magnetic Resonance Imaging data of brain tumor as the research object.Aiming at the problems of intensity inhomogeneity,noise interference,weak contrast,irregular appearance and different size of tumor focus,we use deep convolution neural network to build a powerful and high-precision automatic segmentation method of brain tumor.The main contributions of this thesis include the following aspects,where the first two constributions are to design new network structures to achieve automatic brain tumor segmentation,which belongs to structural innovation;the third contribution is to combine methods in different fields to segment tumor lesions,which belongs to method innovation:(1)There are two problems in the traditional automatic brain tumor segmentation method based on the deep convolution neural network: spatial information loss caused by repeated pooling and sliding operations,and the weak processing ability of multi-scale lesions.In order to solve these two problems,this work first proposes a single-stride 3D atrous convolution instead of pooling and sliding operations to construct the backbone network for feature learning.And then a 3D atrous convolution feature pyramid is added to the end of the backbone network to improve the ability of the whole model to segment tumors of various sizes by means of integrating useful context.Finally,the 3D fully connected conditional random field is used as the post-processing method to improve the output of the network to obtain a structured segmentation of appearance and spatial consistency.A large number of ablation experiments and comparisons with the state-of-the-art methods show that the proposed method has certain advantages and can effectively solve the problems of the traditional convolutional neural network in brain tumor segmentation.(2)The existing methods based on deep convolution neural network using Magnetic Resonance Imaging data have made great progress in automatic brain tumor segmentation.However,the deep convolution neural network constructed by simply stacking convolutional layers will produce some problems that affect the segmentation effect.These problems can be attributed to two aspects: the gradient explosion/disappearance and the limited nature of feature computation.In order to solve these two problems,this work proposes a new framework of feature reuse and multi-scale information fusion for automatic brain tumor segmentation.Firstly,a 3D dense connection architecture is used to construct the backbone network for feature computation.Secondly,a new pyramidal module of constant size feature is designed by using 3D dilated convolution,which is added to the end of backbone network to fuse multi-scale context information.Finally,a kind of 3D deep supervision mechanism is proposed to improve the training of the network.The experimental results show that the network structure proposed in this work can effectively alleviate the training difficulties caused by deep structure and the problem of imcompletely reusing the features of each layer.(3)In view of the irregular appearance and immersive growth of tumors in brain Magnetic Resonance Imaging data,the tumor lesions are not distinguishable from normal tissues.In addition,the heterogeneity of tumors will lead to the similarity between normal region features and tumor features obtained by traditional feature extraction methods.Therefore,we defines tumor segmentation as a maximum posterior probability problemn from the perspective of statistics.Using the theory of partial differential equation and combining the expectation maximization algorithm,we proposes a variation level set algorithm that combines a Gaussian mixture model with Markov Random Fields.At the same time,a deep hybrid network integrated with convolutional neural network and gating recursive unit is designed to implement the proposed method.The sufficient ablation comparison experiments verify the effectiveness of the important components of the method in this work.The comparison with the state-of-the-art method shows that the hybrid network proposed in this work can obtain better segmentation results in public data sets,and has certain advantages.