The Survival Prediction In Glioma With Radiomics Based On Diffusion Kurtosis Imaging

Objective:To explore the value of Radiomics based on diffusion kurtosis imaging(DKI)for predicting overall survival in glioma.The radiomics signature was developed for risk stratification,and the clinical-genetic model was constructed using clinical features and molecular markers to predict the prognosis of patients with glioma,and the combined model combining radiomics signature and clinical-genetic model was constructed to further improve the predictive performance.Methods:This study retrospectively analyzed pathologically confirmed 58 patients with glioma in our hospital from November 2012 to February 2019 according to the inclusion and exclusion criteria.The DICOM images of conventional magnetic resonance imaging and DKI,clinical data,follow-up time,and molecular marker results of the research subjects were collected.The original DKI images were processed on GE Advanced Workstation 4.4to obtain mean kurtosis(MK)and mean diffusivity(MD)images.Regions of interest about tumor were manually segmented using ITK-SNAP software and radiomics features were extracted using the open-source software FAE.A total of 851 features were extracted from each patient’s MK and MD images,respectively.The radiomics features most associated with overall survival(OS)were selected using z-score,least absolute shrinkage and selection operator(LASSO)Cox regression method and Pearson correlation analysis.Three sets of radiomics signatures were established for the linear fitting features according to their respective coefficients,and then the patients were divided into high-risk set and low-risk set based on the optimal cut-off value for radiomics signatures.Stratified analysis was performed using Kaplan-Meier survival analysis,the area under roc curve for predicting 1-,2-,and 3-year survival and the C-index for radiomics signatures were calculated,and the best radiomics signature was obtained.Cox regression analysis was used to explore the prognostic value of clinical features and molecular markers,construct a clinical-genetic model,and perform risk stratification in subgroups of risk factors using the best radiomics signature.Finally,the best radiomics signature and clinical-genetic factors were selected to develop a combined model which could predict 1,2,and 3-year OS probabilities of patients with glioma by the calibration curves and C-index.The combined model was visualized by the nomogram.The clinical decision curve analysis(DCA)was performed to compare the clinical value of the radiomics model,clinic-genetic model and combined model.Furthermore,the generalization of the combined model was explained using internal 5-fold cross-validation.Results:This study selected 5,5,and 6 radiomics features most associated with OS from MK,MD,and combined-sequences images,respectively.Kaplan-Meier survival analysis showed that all three radiomics signatures could divide patients with glioma into high-risk set and low-risk set,and the difference between survival curves was statistically significant(P<0.001).The radiomics signature Radioscore＿com based on the combined-sequence had better predictive performance than single sequence,and had the highest AUC for predicting the 1-,2-,and 3-year survival of glioma,which are 0.72,0.82,and 0.84,respectively,and the highest C-index for predicting OS was 0.733(95%CI:0.688-0.779).Univariate and multivariate Cox regression analysis showed that pathological grade,treatment and O~6-methylguanine DNA methyltransferase(MGMT)are independent risk factors for OS(P<0.05).The C-index of the clinical-genetic Cox regression model was 0.761(95%CI:0.714-0.808).Kaplan-Meier survival analysis showed that the difference between survival curves was statistically significant according to the cut-off value under different risk factor stratification(P<0.001),and the result of the non-MGMT promoter methylation set was not statistically significant.The combined model combining the best radiomics signature and clinical-genetic model had the best performance in predicting OS,with a C-index of 0.832(95%CI:0.797-0.866),which was superior to the clinical-genetic model.The calibration curves and DCA showed that the combined model has good accuracy and clinical practicability.The internal 5-fold cross-validation showed an average C-index of 0.804 for the validation set,indicating the stability of the model.Conclusion:The radiomics signature based on quantitative features from DKI images is an independent risk factor for OS in patients with glioma,and the patients could be stratified.The combined model combining radiomics signature,pathological grade,treatment and MGMT improves the predictive performance compared with the clinical-genetic model,demonstrating the incremental value of Radiomics in predicting OS in glioma patients,which is expected to provide a basis for risk stratification and improve the clinical management of glioma patients.