Multiparametric MRI radiomics and clinical features for preoperative prediction of high-grade meningioma: a two-centre retrospective study

Background Preoperative identification of high-grade meningioma remains challenging using clinical and conventional imaging features. We compared MRI radiomic and clinical models and assessed whether their integration improved external discrimination. Materials and methods This retrospective two-centre study included 284 patients, with 227 used for training and 57 for independent external testing. Radiomics from contrast-enhanced T1-weighted imaging (CE-T1WI), T2-weighted imaging (T2WI), and fluid-attenuated inversion recovery (FLAIR) underwent training-cohort screening, correlation filtering, and LASSO selection. The primary comparison assessed external area under the receiver operating characteristic curve (AUC) for joint-clinical versus clinical logistic regression (LR) models. Secondary comparisons assessed adding clinical predictors to joint imaging and combining sequences relative to T2WI. Bootstrap-based normal-approximation comparisons with paired resampling and Holm correction evaluated these three contrasts. Calibration, decision curves, and K-nearest neighbour (KNN) models were exploratory. Results The external cohort contained 48 low-grade and 9 high-grade tumours. All four imaging-only LR models had higher external bootstrap-mean AUC point estimates than the clinical model (0.663; 95% CI, 0.467–0.840). CE-T1WI yielded 0.863 (0.755–0.949), and FLAIR yielded 0.838 (0.724–0.932). T2WI and joint imaging yielded 0.719 (0.555–0.862) and 0.807 (0.640–0.941), respectively. The joint-clinical model had an external AUC of 0.754 (0.557–0.920). Imaging-only Brier scores (0.113–0.127) were lower than those of clinical (0.246) and joint-clinical (0.209) models. Both clinical and joint-clinical predictions overestimated risk in the external cohort. The primary joint-clinical versus clinical bootstrap-mean AUC difference was 0.091 (95% CI, −0.126 to 0.309; Holm-adjusted P = 0.410). None of the three planned contrasts remained significant after Holm correction. Conclusion MRI radiomics showed promising external discrimination and lower probability error than the clinical model in this cohort. These findings support further evaluation of radiomics for preoperative grade prediction, while the incremental benefit of clinical-radiomic integration remains unconfirmed.