Brain Tumor Segmentation Via Dilated Convolutional Networks

Gliomas are the most common type of brain tumor in adults.Early diagnosis and treatment can effectively prolong the survival time of patients.Magnetic resonance imag-ing(MRI)is an important brain tumor imaging technique.The task of brain tumor image segmentation is to mark the corresponding region of brain tumor in MRI.Accurate la-beling can help doctors staging the disease,locate the lesion,develop the appropriate treatment.However,labeling MRI is an extremely time-consuming task.The purpose of the automated brain tumor MRI segmentation model is to resolve this contradiction.The classic brain tumor MRI segmentation techniques include threshold method domain-based segmentation method pixel-based classification methods.With the unprecedented success of deep learning in the field of computer vision,the brain tumor MRI segmenta-tion model has also been gradually developed to adopt deep models.Those models which usually derived from the semantic segmentation model have two problems.One is that the network can not process the native 3D MRI,and the other is the local information loss caused by the existence of the pooling layer.In view of this,this article focuses on the task of brain tumor MRI segmentation,from data preprocessing to the construction of the segmentation model.The content of this article is divided into three parts.First,the main fault of MRI is that it's pixel intensity have no absolutely anatomical meaning.Therefore,a lot of im-age preprocessing models are promote to solve the problem such as bias field correction,intensity range standardize.However,there are no useful metric to measure the perfor-mance of such methods.In this article we promote to utilize the paired t-test and F-test to measure the effects of bias field correction method and intensity range standardize model respectively.Second this paper defines the d dimensional dilated convolution op-erator theoretically and the corresponding simple dilated convolutional network(DCN).This paper also defines the receptive field and the checkerboard patterns(CP)of the DCN,and studies its basic properties theoretically.Besides,this paper presents and proves two necessary and sufficient conditions about the CP of DCN,and gives the formula for the calculation of network receptive fields.On the other hand in order to make more effective use of DCN,this paper analyzes the relationship between input image size and network computing efficiency and computer memory pressure.It is proved that when the input image is d dimensional hypercube,the network has the highest computing efficiency and minimum memory pressure.This conclusion provides a powerful guide for designing the input and output structures of the network.Third based on the first two parts,this paper constructs a 3D dilated convolutional network which is suitable for MRI.The performance of image segmentation between dilated convolutional networks and general convolutional networks is compared experimentally and the effect of checkerboard pattern on the accu-racy of network classification is verified.