Brain Tumor Segmentation Using Multimodal MR Brain Images And Fully Convolutional Network

Glioma is a relatively common primary brain tumor in adulthood,with an incidence rate of about 80% of malignant brain tumors.Its high rate of disability and mortality seriously threaten people’s lives and health.Magnetic Resonance Imaging,MRI is an indispensable tool in the diagnosis and treatment of glioma.According to the tumor region in MRI,physicians can get the location,pose and malignancy of glioma.Currently,the tumor region is segmented manually,which is a hard and labor-intensive task.Moreover,manual segmentation is non-repeatable and susceptible to subjective factors.Recently,deep learning(Deep Learning,DL)has been widely used in the field of brain tumor segmentation.However,due to the infiltrative growth of glioma,the tumor boundaries are blurred,and the location,shape and size of the tumor are high diversity,making accurate brain tumor segmentation a difficult task.To tackle the above issues,in this thesis,we proposed two DL based models: the end-to-end multi-task cascade guidance model and the multi-task&multi-view convolutional network model.The first model divides the tumor into three mutually contained hierarchical regions in a multi-task cascade(the entire tumor,the tumor core and the enhanced tumor),and each task is responsible for segmenting one of the three regions.Each task is based on the segmented regions from the previous task,i.e.,the segmentation is refined in a cascaded manner.Since priori knowledge from the previous task is used to guide the current task,the segmentation accuracy of each tumor sub-regions can be improved.The second model is based on multi-view information from MRI.Specifically,by constructing three parallel branch networks containing attention mechanisms,each segmentation task of one perspective is processed,and multi-view information fusion is performed,each of which works on one view(axial,coronal or sagittal).The segmentation results are obtained by fusing the information from all three views.Since multi-view information are used,the segmentation accuracy can be enhanced.The proposed model were evaluated using public brain tumor dataset BraTS 2017,and the experimental results demonstrated that our models can improve the segmentation accuracy as compared with the state-of-the-art methods.