Enhancement Without Contrast: Stability-Aware Multicenter Machine Learning for Glioma MRI Imaging

Gadolinium-based contrast agents (GBCAs) are vital for glioma imaging yet pose safety, cost, and accessibility issues; predicting contrast enhancement from non-contrast MRI via machine learning (ML) provides a safer, economical alternative, as enhancement indicates tumor aggressiveness and informs treatment planning. However, scanner and population variability hinder robust model selection. To overcome this, a stability-aware framework was developed to identify reproducible ML pipelines for predicting glioma contrast enhancement across multicenter cohorts. A total of 1367 glioma cases from four TCIA datasets (UCSF-PDGM, UPENN-GB, BRATS-Africa, BRATS-TCGA-LGG) were analyzed, using non-contrast T1-weighted images as input and deriving enhancement status from paired post-contrast T1-weighted images; 108 IBSI-standardized radiomics features were extracted via PyRadiomics 3.1, then systematically combined with 48 dimensionality reduction algorithms and 25 classifiers into 1200 pipelines, evaluated through rotational validation (training on three datasets, external testing on the fourth, repeated across rotations) incorporating five-fold cross-validation and a composite score penalizing instability via standard deviation. Cross-validation accuracies spanned 0.91–0.96, with external testing yielding 0.87 (UCSF-PDGM), 0.98 (UPENN-GB), and 0.95 (BRATS-Africa), averaging ~0.93; F1, precision, and recall remained stable (0.87–0.96), while ROC-AUC varied (0.50–0.82) due to cohort heterogeneity, with the MI + ETr pipeline ranking highest for balanced accuracy and stability. This framework enables reliable, generalizable prediction of contrast enhancement from non-contrast glioma MRI, minimizing GBCA dependence and offering a scalable template for reproducible ML in neuro-oncology.