Prediction Of Postoperative Overall Survival By MRI Radiomics In Partial Resection Of High-Grade Gliomas

Objective:To investigate the value of radiomic data extracted from MRI of the high-grade gliomas(HGGs)in predicting overall survival(OS)after surgery and chemoradiotherapy.Compare prediction performance of different radiomic models,and try to build a prognostic prediction model based on Radiomic Score(RS).Methods:Outcome information from 60 patients receiving surgery and chemoradiotherapy was evaluated retrospectively.MRI suggested tumor residual.The patient's clinical characteristics included age,gender,tumor location,KPS,etc.We determined OS by telephone follow-up.The patients were randomly divided into training and validation cohort at a ratio of 2:1.MR images were achieved from axial T1-weighted post-contrast,T2-weighted and FLAIR images within 1 week after surgery and 1 month after chemoradiotherapy.We used MATLAB R2018a for images co-registration,VOI segmentation and radiomic features extraction.Radiomic features were extracted from tumor residual enhancement region(RER)and peritumoral edema region(PER)on both pre-and post-chemoradiotherapy.We also calculated the delta radiomic features,which represented changes between radiomic features from one time to another.Radiomic features included first-order statistics features,volume and shape features,texture features and wavelet features.The features were filtered by Intraclass Correlation Coefficient(ICC)and least absolute shrinkage and selection operator(LASSO)method.The most reproducible and valuable radiomic features were used to build a multi-factor Cox proportional hazard model.The prediction performances of different models were compared by C-index.RS was calculated by a formula according to the features and coefficients in optimal model.We used RS and clinical characteristics to construct prediction nomograms.The patients were divided into low-and high-risk groups by RS median,and Kaplan-Meier survival analyses were performed to estimate survival differences between the two groups under the same or different clinical risk factors.Results:The delta radiomic features derived RER had the best OS prediction performance.We found 4 radiomic signatures(M3D,Variance,Energy,Correlation)were significantly associated with OS.RS=2.462 × M3D+1.823 × T1CE_HHH Variance+1.781 × T2_HLL_Energy-2.588 × T2_HHH_Corrlation.The Cox regression models were constructed by RS and clinical data.In training cohort,the OS prediction performance of RS was higher than clinical data(C-index:0.858 VS 0.825),which was better by combining them together(C-index:0.908).The verification cohort had the similar outcome.The nomogram showed a significant improvement over the clinical data in terms of OS prediction performance(C-index:0.905 VS 0.825).The Cox model showed RS was an independent risk factor for OS in both training cohort(HR=1.534,P=4.95×10-5)and verification cohort(HR=1.775,P=0.001).The prediction performance of nomogram combined RS and clinical data was higher than using alone(C-index:0.905 VS 0.825).The calibration curve showed good agreement between nomogram-evaluation and actual OS.Kaplan-Meier survival curve showed OS was significantly lower in high-risk than low-risk patients within the same or different clinical risk factors(P<0.05).Conclusion:MRI radiomic can predict overall survival of HGGs after surgery and chemoradiotherapy.The radiomic signatures and RS showed higher prediction performance consistently.