Imaging Biomarkers and Machine Learning for Preoperative Prediction of Progression in Glioblastoma

Background Glioblastoma is characterized by marked biological heterogeneity and early clinical deterioration, making progression assessment and risk stratification challenging in routine practice. Preoperative MRI radiomics may provide quantitative imaging biomarkers to support individualized prognostic modeling. We investigated whether a standardized radiomics–machine learning workflow can predict progression in histologically confirmed glioblastoma. Methods We retrospectively analyzed adults with CNS WHO grade 4 glioblastoma treated between 2019 and 2024. Progression was defined using RANO criteria with clinical-radiological follow-up to reduce misclassification in the early post-treatment window. From preoperative contrast-enhanced T1-weighted MRI, 159 radiomic features were extracted using an IBSI-oriented pipeline after semi-automatic 3D segmentation. After preprocessing and correlation filtering, feature selection was performed within a nested leave-one-out cross-validation framework. Nine classifiers were benchmarked, and uncertainty was quantified by bootstrap resampling (1,000 iterations). Performance was assessed by balanced accuracy, AUROC, AUPRC, Brier score, and calibration. Results Eleven patients met eligibility criteria (progression, n = 5; no progression, n = 6). Three GLRLM texture features were consistently selected across validation folds (ShortRunHighGrayLevelEmphasis, ShortRunLowGrayLevelEmphasis, LowGrayLevelRunEmphasis). The Extra Trees classifier achieved the best overall performance (balanced accuracy 0.817, AUROC 0.833, AUPRC 0.710, Brier score 0.188), outperforming alternative models in this cohort, with favorable calibration. Conclusions A compact preoperative MRI radiomics signature combined with machine learning showed promising performance for progression prediction in glioblastoma. This study supports the feasibility of a rigorously standardized, calibration-aware predictive framework and provides a basis for external multicenter validation.