Decoding glioblastoma recurrence: harnessing radiomics and machine learning for early prediction
Abstract
This study aimed to analyze postoperative periresidual FLAIR hyperintensity and assess radiomic features from periresidual enhancing and non-enhancing subregions for predicting early GBM recurrence (pre-CRT RANO criteria). We also evaluated the incremental prognostic value of multi-subregion radiomic signatures over conventional clinical and MRI features. Adult patients with surgically confirmed GBM were retrospectively enrolled and randomly split into training ( n = 128) and test ( n = 56) cohorts at a 7:3 ratio. Postoperative pre-radiotherapy CE-T1WI and FLAIR images were fused to semi-automatically delineate periresidual enhancing (ROI-1) and non-enhancing hyperintense (ROI-2) regions using 3D Slicer. Radiomic features were extracted from both subregions and their fusion area. LASSO regression was used to select optimal features related to early recurrence from ROI-1, ROI-2, and combined subregion datasets. Five machine learning algorithms (DT, LR, NB, SVM, XGBoost) were applied to build radiomic prediction models. To evaluate the incremental value of radiomic signatures, a clinical/MRI-only baseline model was constructed using conventional clinical and morphological MRI features. Multivariate logistic regression was performed to identify independent clinical and MRI predictors. A quantitative Rad-score was generated based on the optimal combined radiomic features, and an integrated radiomics-clinical model with a predictive nomogram was established by incorporating the Rad-score and independent clinical/MRI variables. Ten-fold cross-validation was used for internal validation. Model performance was comprehensively assessed using ROC curves, calibration curves, and decision curve analysis in terms of discrimination, calibration, and clinical utility. A logistic regression analysis of 184 GBM cases found subventricular zone involvement [HR 4.868 (1.657–14.308), p = 0.004] and partial resection [HR 3.640 (1.341–9.883), p = 0.011] as predictors of early recurrence. The baseline model achieved an AUC of 0.701. Among five models using ROI-1 features, XGBoost performed best with validation AUC of 0.663. For ROI-2, the DT model excelled with 0.713. Combining ROI-1 and ROI-2, XGBoost achieved 0.752. Integrating radiomics with clinical and MRI data, XGBoost reached 0.817. DCA indicated the combined model had the highest clinical effectiveness. DCA showed that a threshold probability of 0.1 to 0.6 yields a net gain. In-depth analysis of the enhanced and non-enhanced regions, within the FLAIR hyperintense surrounding the residual cavity of GBM before pre-CRT using ML, has revealed certain value for the early prediction of GBM recurrence. Furthermore, a combined model incorporating radiomic features, clinical and conventional MRI attributes enhances the predictive accuracy for survival prognosis.