Brain Tumor Segmentation Based On Fully Convolutional Neural Network

Glioma is one of the most common tumors that exists in primary central nervous system.It is at high risk of morbidity and mortality.Magnetic Resonance Imaging(MRI)is widely used in clinical because of its high spatial resolution and high contrast imaging of brain soft tissue,making it the best choice for doctors to analyze brain structures.Image segmentation is a key step in brain tumor research,which provides guidance for the follow-up diagnosis and treatment.However,manual segmentation is a very time-consuming and laborious work,and it relies heavily on the experience of doctors.Because of the various shapes and complex structure of brain tumor and the serious class imbalance problem in dataset,traditional image segmentation algorithms such as region growing method and pixel classification are often difficult to obtain satisfactory segmentation results.Therefore,the study of automatic brain tumor segmentation is still a very challenging task.In recent years,with the rapid development of deep learning technology,artificial neural network has been widely used in medical image segmentation and achieved satisfactory results.In this thesis,the fully convolutional network is studied and taken as the basic framework to design a new network model suitable for brain tumor segmentation.The main work is summarized as follows.1.Inspired by the distinct hierarchical structure of brain tumor,chapter 3 proposes an efficient cascaded u-shape net(CU-Net),in which the whole tumor is segmented firstly and then the tumor internal substructures are further segmented.Cascade design will greatly increase the depth of network,which is harmful to the propagation of the information and gradient.In this thesis,residual structure and between-net connection are introduced to alleviate such a problem.To solve the serious class imbalance problem in brain tumor data,Dice loss function is used to train the network.2.In chapter 4,in view of the 3D imaging characteristics of MRI,3D convolution is introduced and the cascade network is expanded into 3D structure to effectively utilize the 3D spatial information of MRI image.Experiments show that compared with 2D network,3D network has obvious advantages in time efficiency and segmentation accuracy.3.In chapter 5,to further improve network performance,a novel attention module called Cross-net Guided Attention(CNGA)is proposed.The first stage network locates the brain tumor area and adaptively guides the attention in spatial domain for next stage network through CNGA,thus segmenting the substructure of the tumor more precisely.In addition,aiming at the instability of the Dice loss function,a weighted Dice loss function(WDL)is proposed to train the network.And we also proposed a structure regularization(SR)method for brain tumor segmentation based on the hierarchical priori of brain tumor to constrain the prediction process of the network,so as to reduce the false positives.The above models and algorithms are systematically tested on the Bra TS18 and Bra TS19 datasets.Results show that the algorithms presented in this thesis have achieved very competitive performance.