Abstract
To develop and validate a calibrated, explainable radiomics pipeline for classifying progression status in postoperative glioblastoma (GBM) patients using multiparametric MRI acquired at longitudinal follow-up scans.
We retrospectively analyzed the MU-Glioma-Post dataset, which includes serial post-treatment MRI (T1, T1-CE, T2, FLAIR). Volumes were corrected with N4, resampled at 1 mm³ resolution, co-registered to SRI24, skull-stripped, and segmented using nnU-Net into categories such as enhancing tissue, non-enhancing tumor core, pericavitary FLAIR hyperintensity, and resection cavity. Expert neuroradiologists refined these segmentations to ensure precise delineation for the extraction of radiomic features. Radiomics features were extracted on Pyradiomics, Laplacian-of-Gaussian (σ = 0.5–3.0 mm), and 3D wavelet sub-bands. Stability-aware ranking (variance/correlation filters, L1-logistic, permutation importance) was performed before classifier training. A LightGBM model was optimized with patient-aware 5-fold cross-validation and Optuna tuning, then Platt-calibrated on out-of-fold scores. Performance was assessed on a patient-held-out test set using AUC, confusion matrix, and Brier score, while SHAP provided cohort-level explanations.
The LightGBM model trained on 256 radiomics features achieved an AUC of 0.80 on the held-out patient test set, with a confusion matrix indicating sixteen false positives and six false negatives. Calibration improved the Brier score from 0.093 to 0.088. Global explanations showed that the model mainly relied on coarse-scale wavelet/LoG GLCM/GLSZM textures across all modalities, with little dependence on shape features. Risk-increasing attributions were often due to enhancing rim on T1-CE and pericavitary FLAIR textures.
A calibrated, explainable radiomics model trained on longitudinal post-operative MRI provides strong discrimination and well-calibrated probabilities for GBM progression prediction, supporting threshold-based decision-making in surveillance. Radiomic signatures emphasizing coarse-scale heterogeneity and enhancing texture patterns were most strongly associated with earlier progression. External, multi-centre validation and evaluation of diffusion/perfusion and delta-radiomics are warranted to establish generalizability.
•An explainable radiomics model classifies glioblastoma progression status after surgery.•Stability-aware feature selection improves robustness and reproducibility.•LightGBM classifier achieves strong performance with AUC = 0.80.•Probability calibration enhances reliability for clinical decision support.•Coarse-scale wavelet and LoG textures drive model predictions.