Calibrated and explainable multiparametric MRI radiomics for differentiating tumor positive disease from treatment-related changes in glioblastoma.
Authors
Affiliations (8)
Affiliations (8)
- Division of Neuroimaging and Neurointervention, Department of Radiology, Stanford University, Stanford, CA, United States.
- Department of Electrical and Computer Engineering, National Technical University of Athens (NTUA), Athens, Greece.
- Department of Radiology, Stanford University, Stanford, CA, United States.
- Department of Medicine, National and Kapodistrian University of Athens, Athens, Greece.
- Department of Biological Sciences, Kean University, Union, NJ, United States.
- Internal Medicine-Hematology, University of Patras Medical School, Rion, Greece.
- 1st Department of Neurology, Medical School, National and Kapodistrian University of Athens, Eginition Hospital, Athens, Greece.
- Department of Radiology, University of Miami, Miami, FL, United States.
Abstract
Differentiating tumor from treatment-related changes (TRC) remains a key challenge in post-treatment glioblastoma (GBM) monitoring, as both can exhibit overlapping enhancement and FLAIR abnormalities on standard MRI. Our goal was to build and internally validate an explainable, calibrated, multiparametric MRI radiomics model to distinguish between tumor and treatment effects. We retrospectively included patients from the UCSD-PTGBM cohort who had post-treatment GBM MRI examinations with available reference labels for tumor-positive disease (TP) or TRC. Examinations were eligible when contrast-enhanced T1-weighted imaging, FLAIR, ADC, DSC perfusion imaging, and corresponding tumor segmentation masks were available. After applying the eligibility and exclusion criteria, the final study cohort comprised 168 patients with 224 MRI examinations, 173 TP and 51 TRC cases. Radiomic features were extracted from all four MRI sequences. Feature selection and hyperparameter tuning were performed exclusively on the training data before evaluating the model on the test set. Performance metrics, calibration, decision curve analysis, and SHAP-based interpretability were assessed. The extra trees classifier trained on the top 50 radiomic features demonstrated strong performance on the held-out test set, with an AUC of 0.887, PR-AUC of 0.952, balanced accuracy of 83.7%, sensitivity of 75.0%, and specificity of 92.3%. Platt calibration improved the Brier score from 0.229 to 0.132 and the expected calibration error from 0.275 to 0.084. Decision curve analysis indicated a net benefit across relevant clinical thresholds. SHAP analysis revealed balanced contributions from ADC, DSC, T1CE, and FLAIR modalities. A calibrated and explainable multiparametric MRI radiomics model showed robust internal performance in differentiating tumor from treatment effects in post-treatment GBM. As these findings are hypothesis-generating from an internally validated study, external validation is needed to confirm clinical usefulness and broader applicability.