The Multi-parametric Mri-based High Risk Subregion Identification And Survival Stratification For Glioblastoma Patients

Considering the heterogeneity of Glioblastoma,that the texture characteristics of Magnetic resonance(MR)images in different subregions of glioblastoma(GBM)are different,the hidden heterogeneous information related to prognosis is also different.Therefore,it is the key to stratification of prognosis by tumor spatial structure segmentation and the analysis of risk characteristics in different subregions based on radiomics.In order to obtain the survival-relevant high-risk subregions,which has the most relevance to short-survival patients,and realize accurate stratification of patients' survival.In this study,104 patients with GBM from The Cancer Genome Atlas(TCGA)were analyzed.Firstly,The gross tumor region(GTR)of 104 GBM patients' s normalized MR imageswere automatically segmented into intratumoral heterogeneity subregions.Then,extracting radiomics characteristics from each subregion to build the multi-instance bag.Finally,the multiple instance learning(MIL)model is established and validated using the subregion characteristics.The main work of this thesis includes:(1)Establishment of intratumoral heterogeneous segmentation model for GTR:Rather than the traditional segmentation of the whole tumor,this paper focuses on the GBM intratumoral segmentation model,which can realize the automatic segmentation of GTR according to the heterogeneous spatial distribution information in the tumor.Firstly,GTR was depicted based on standardized MR images with FLAIR,T1-weighted(T1),contrast-enhanced(T1C)and T2-weighted(T2)sequences.The signal intensity of voxels from 104 GTRs were pooled as global intensity vector,and K-means clustering was performed on it to find the optimal global clusters.Finally,Subregions were generated by assigning back voxels belonged to each global cluster,and 294 subregions of 104 patients were obtained.The segmentation results showed that the number of subregions with different GTR,which could represent the heterogeneity pattern.(2)The GBM patients' survival stratification and prediction of high risk area based on MIL methodThe prognosis was used as an indicator of tumor risk characteristics,and the mi-SVM method was validated using radiomics features,realizing the location the ‘high risk' region and the prognosis prediction.In this paper,200 subregions were used to train the MIL model at the subregional level,and 94 subregions were used to test the model.The results showed that a)the accuracy,sensitivity and specificity of overall survival(OS)stratification were 87.88%(29/33),85.71%(12/14)and 89.47%(17/19),respectively.b)In the training group and the validation group,41 high-risk subregions were correctly predicted from patients with short-term survival,in which the median overlap rate of non-enhancing component was 60%,which was the main component of the 41 high-risk subregions.Therefore,as a data-driven method,MIL is more effective in the survival stratification of GBM patients.Moreover,It is suggested that non-enhanced area on MR images was the most important in high-risk subregions.In addition,The MIL approach provides a new perspective on the clinical challenges of GBM with coarse-grained labeling.