Research Of Key Techniques In Computer-aided Diagnosis Of Gliomas

Gliomas accounts for about 45% of main brain diseases,which are caused by brain and spinal cord glial cells carcinogenesis,so they are also called brain cancers.Based on the definition of World Health Organization(i.e.WHO),the most deadly gliomas are classified as grade III and grade IV,which are collectively called high grade glioma(i.e.HGG).The percentage of HGGs in gliomas is about 75%,and the mean overall survival time of patients in HGG is about 2 years.Currently,there is no efficient clinical method to treat HGG.Because of the characters of high dead rate and hard to treat of HGG,it is very important to assistant doctors to diagnose earlier and prognose more accurately with computer-aided diagnosis techniques,which can also help therapists to make better treatment protocols,prolong life-span,improve living quality and palliative care.The studies in our thesis mainly focus on the key techniques about computer-aided diagnosis of gliomas,which include brain tumor segmentation,outcome prediction and prognosis related genetic molecular(O~6-mehtylguanine-DNA methyltransferase,MGMT and Isocitrate Dehydrogenase 1,IDH 1)status prediction.Accurate brain tumor segmentation techniques can help clinicians detect and diagnosis gliomas effectively and early,which are also one of important steps during imaging pre-processing.In the study of brain tumor segmentation,we proposed an effective,strong generalization ability but simple framework,deep learning based multi-task based brain tumor segmentation,which divide one segmentation task into two tasks(e.g.detection and initial segmentation)and combine the results of these two tasks to get the final segmentation result without post-processing.It is simple,easy to implement,and robust segmentation method with strong generalize capability.In the study of overall survival(OS)time prediction of HGG patients,we proposed to predict OS from two different ways that are radiomics based and connectomics based methods.With regard to the radiomics based outcome prediction,we firstly use the brain tumor segmentation method mentioned above to extract the tumor region,then we designed a 3D multi-model multi-channel deep learning network to learn prognosis related deep features from tumor regions of multi-model multi-channel medical images to predict outcome individually.Since the radiomics features only focus on tumor related regions and ignore the effect to normal appearance brain regions,therefore,we proposed to predict OS using connectomics features.As for connectomics based outcome prediction,we firstly construct two brain functional connectivity(FC)network(e.g.static low order and dynamic high-order FC networks)which reflect brain activities and interaction patterns in two different views and extract network properties using graph theory analysis as features;finally,a machine learning based outcome prediction framework is proposed to predict OS individually.Except for using medical images to diagnose gliomas auxiliarily,the mothod of combing medical imaging data with genetic meloclar information to treat gliomas individually has been accepting by more and more neurosurgeons and reseachers.There were a lot of researches reported that the O~6-methylguanine-DNA methyltransferase(MGMT)promoter methylation and isocitrate dehydrogenase 1(IDH1)mutation are two important molecular predictors related to better prognosis of HGG patients.However,in clinical routine,the identification of MGMT and IDH1 statuses depend on melocular pathological analysis of invasively acquired tumor tissue specimen,which may cause severe brain injury and sometimes even increase the risk of infection.In order to help these melocular biomarkers extensively applied on clinical routine,we proposed a novel iterative canonical correlation analysis based feature selection method,which uses connectomics features from pre-surgical medical images,to predict MGMT and IDH1 statuses.Nevertheless,this method cannot learn complex nonlinear relationships between feature matrices and label space,and cannot deal with data with missing melocular information.In order to solve these problems,we further proposed a methoc combined with nonlinear matrix completion and transductive multi-task feature selection to predict MGMT and IDH1 status jointly.All these two proposed methods were the first time to use systematic and large-scale connectomics features to predict MGMT and IDH1 status,both of which show desirable MGMT and IDH1 status prediction ability.