KEY POINTS
- The study evaluated patients treated with proton therapy for skull-base chordoma and chondrosarcoma at CNAO. Feature stability was assessed in 52 patients with paired fat-saturated and non-fat-saturated T2-weighted MRI, while brain-radionecrosis modeling used a separate mixed-protocol cohort of 80 patients.
- Automated segmentation of gray matter, white matter and cerebrospinal fluid was followed by extraction of 1,911 first-order and texture radiomics features. Forty preprocessing combinations were tested, including bias-field correction, intensity normalization and Combat feature harmonization.
- Fat saturation had a major effect on quantitative features: median concordance correlation coefficients across preprocessing strategies ranged only 0.21–0.58. Combat provided the most consistent improvement; the simplest optimal workflow was Combat without additional bias-field correction or intensity normalization.
- Even under that optimized workflow, only 30% of features reached CCC ≥0.70 and 10.7% reached CCC ≥0.85. Restricting to the excellent-stability threshold reduced the candidate feature space from 1,911 to 204, eliminating almost 90% of baseline features.
- In the heterogeneous 80-patient cohort, 32 patients (40%) developed grade ≥1 brain radionecrosis during a median follow-up of 23 months—22 grade 1 and 10 grade 2 events. Prediction using all features performed close to chance, with AUC 0.52 (95% CI 0.50–0.54).
- Restricting the model to highly stable features improved heterogeneous-cohort AUC to 0.57 (95% CI 0.55–0.59; p<0.001). In the 35-patient homogeneous subgroup, performance improved most strongly in non-fat-saturated scans—from AUC 0.51 to 0.70 using only excellent-stability features.
- Despite statistically significant improvement, the mixed-protocol model's AUC of 0.57 remains insufficient for clinical prediction. The cohort was small, events were limited, model validation relied on repeated internal cross-validation and no independent external dataset was tested.
CLINICAL TAKEAWAY
This study illustrates a fundamental radiomics problem: a model can learn MRI acquisition protocol as easily as it learns biology. Harmonization and stability filtering improved reproducibility and prediction, but the resulting radionecrosis model is not clinically ready. Protocol robustness should be treated as a prerequisite—not an afterthought—when building MRI radiomics biomarkers.