AUC of 0.887 on the held-out test set. That is the single number that justifies attention from anyone building MRI post-processing pipelines for glioma follow-up. A multi-institutional team — Stanford, NTUA, University of Miami, Eginition Hospital, Kean, and Patras — has internally validated an explainable, calibrated radiomics model that separates tumor progression from treatment-related changes in post-treatment glioblastoma, as reported in Frontiers in Oncology.
The classification layer
The cohort derives from the UCSD-PTGBM dataset: 168 patients, 224 examinations after eligibility filtering, split into 173 tumor-positive (TP) and 51 treatment-related change (TRC) cases. Four MRI sequences enter the feature pipeline — contrast-enhanced T1, FLAIR, ADC, and DSC perfusion. Features are extracted across all four; selection and hyperparameter tuning occur strictly on the training set before a single evaluation on the held-out test set. No leakage.
An Extra Trees classifier on the top 50 radiomic features yields:
- AUC: 0.887
- PR-AUC: 0.952
- Balanced accuracy: 83.7%
- Sensitivity: 75.0%
- Specificity: 92.3%
The sensitivity-specificity gap matters. The model tolerates the clinical priority of minimizing false positives — you do not want to re-irradiate a treatment effect — but it pays for that with a 25% miss rate on true recurrence. In a population where median overall survival sits at roughly 15 months, that trade-off constrains adoption.
Calibration and interpretability
A radiomics model that is accurate but uncalibrated is unusable at the bedside. The uncalibrated Extra Trees output degrades to a Brier score of 0.229 with an expected calibration error of 0.275 — numbers that would make any clinician distrust a probability reading. Platt scaling corrects this. After calibration, the Brier score drops to 0.132; ECE falls to 0.084. Decision curve analysis indicates net benefit across the relevant clinical thresholds.
SHAP analysis decomposes the model's reasoning across modalities. Contributions from ADC, DSC, T1CE, and FLAIR are balanced — no single sequence dominates. For software developers, this is the signal: the pipeline does not collapse into a single-channel shortcut. Each modality carries independent discriminatory weight.
What constrains deployment
The work is hypothesis-generating, internally validated only. External validation on independent cohorts remains the gating requirement before any clinical integration. Multi-vendor scanner variability, protocol drift across institutions, and differences in DSC acquisition will degrade performance. Until the model is tested across those axes, the 0.887 AUC is a ceiling estimate, not a deployable floor.
For radiomics engineers: the feature stack, the train-locked selection protocol, and the calibration step are the components worth replicating. For clinicians: the specificity-heavy operating point is the one to monitor. For everyone: do not confuse a strong internal result with a deployment-ready system.