Subregional Radiomics-Based Progression Free Survival Prediction In Glioblastoma On Conventional Magnetic Resonance Imaging

Objectives:To identify intratumoral subregion with distinct heterogeneity based on conventional MRI sequences,to investigate whether radiomics features extracted from subregions are better than those from the whole tumor in the prediction of progression-free survival(PFS)for patients with glioblastoma(GBM),and to investigate whether the prediction performance can be further improved when integrating radiomics features and clinical data(including clinical,molecular and semantic radiological features).Methods:In this retrospectively study,186 cases of GBM from 3 different medical institutions were included,of which 69 were from Nanfang Hospital,30 were from Yanling Hospital,and 87 were from the external public TCGA-GBM dataset.The first two centers were combined to be served as training cohort and the TCIA dataset was used as external independent test cohort.Individual-and population-level clustering was performed on conventional MRI images(including T1-weighted imaging,T2weighed imaging,fluid-attenuated inversion recovery sequence,and enhanced T1 imaging).Each tumor was automatically segmented into several phenotypically consistent subregions.Then,radiomics features extraction was performed on whole tumor and subregions,respectively.Pearson’s correlation analysis,univariate Cox regression analysis,and forward stepwise algorithm were used for features selection.These selected radiomics features of each subregion were then concatenated to create radiomics risk scores(RRS)by multivariate Cox regression analysis.The RRS was independently validated on the external test cohort.In addition,correlations between RRS and basic features were assessed by Spearman correlation analysis.Finally,different prediction performances of the fusion of each RRS and clinical data were computed.Results:Three subregions(denoted as S1,S2,and S3)were identified with distinct MRI radiomics features in both the training and the test cohorts.In both cohorts,the prognostic performance of RRS-S1(C-index:0.8043 and 0.7056,p<0.001)was superior to that of the whole tumor(C-index:0.7336 and 0.6417,p<0.001).After features’ fusion,C-index can be further improved,of which the combination of RRSS1 and the basis features reaches the highest in both cohorts(C-index,0.8283 and 0.7301,for training and test cohorts,respectively).Conclusions:This preliminary study demonstrated significant associations of subregional radiomics features with GBM prognosis,and that subregion S1 had the highest prognostic potential,implying that it is expected to be a novel prognostic biomarker for GBM,which would be useful for risk-stratification and personalized treatment decisions.