Fine-granted Segmentation For Brain Tumor Using Cascaded Convolutional Networks

Brain tumors,with very high lethal rate,seriously cause harm to the health of patients.With different histological structure,brain tumors can be divided into various heterogeneous sub-regions such as edema,necrotic core,enhancing and non-enhancing tumor core,which are with different prognosis,treatment and surgical value.Because the magnetic resonance imaging(MRI)technology has characteristics like non-invasive,three-dimensional imaging,multimode imaging,etc.,it is often used to diagnosis,locate the tumor and to help doctors make surgical plans and other clinical processes.Recently,the research on fine-grained segmentation algorithm of brain tumors based on 3D multi-modal MRI has become a hot topic.It is also a practical need of clinical treatment and medical research.The main work of this paper has three points.First,because the traditional fine-grained segmentation algorithm cannot make full use of medical information,This paper designed a cascaded system based on classical method of deep learning.The fine-granted segmentation system of brain tumor is divided into tree stages: whole tumor(including tumor core and edema)segmentation;tumor core(including enhancing tumor,non-enhancing tumor and necrotic)segmentation;enhancing tumor segmentation.The cascaded system improves the performance.Second,based on the above scheme,the paper proposed a multi-frame 2D deep convolutional network,which further improves the performance.Third,for the purpose of reducing false positive voxels in the segmentation results and improving performance in small tumor cases,this paper proposed a VASPP structure based on the method of Densenet and atrous convolution.The structure improves the robustness of the proposed system.Finally,the proposed algorithm can achieve segmentation with high precision,which is robust in the Brain Tumor segmentation task.The system reaches the clinical requirement basically.In order to analyze the performance of the proposed algorithm,I design several sets of contrast experiments and prove the advantages of the cascaded network,multi-frame convolution network and the VASPP structure,respectively.Besides,the dataset used in this paper is BraTS2018,including several 3D MRI with 4models of gliomas,one kind of brain tumor.