MRI radiomics reveals peritumoural distance associations with IDH and MGMT status in gliomas

To develop a non‑invasive multiparametric MRI radiomics model for predicting IDH mutation and MGMT promoter methylation in adult diffuse gliomas, systematically evaluating different peritumoural regions (3 mm, 5 mm, 10 mm) and sequence combinations contrast-enhanced T1-weighted imaging (T1WI-CE), T2-weighted fluid-attenuated inversion recovery (T2-FLAIR), and arterial spin labeling (ASL) with multi‑centre external validation. Data from two centres were analysed (UCSF: IDH n = 295, MGMT n = 266; AHHU eligible cohorts: IDH n = 34, MGMT n = 25). The UCSF cohort was stratified into training (80%) and internal test (20%) sets. After model development, the complete pipeline was repeated on all UCSF cases to fit each final model. Feature-complete AHHU subsets were used for the reported external results (IDH n = 24; MGMT n = 22). After training-only imputation and Z-score scaling, radiomic features were screened using ICC, Mann-Whitney U testing, Spearman correlation filtering, a hybrid mRMR procedure, and L1-penalised logistic regression. Five-fold cross-validation within the training set selected the regularisation parameter from 10 logarithmically spaced candidate C values. The locked development models were evaluated on the internal test set, and the full-UCSF refits were evaluated on the external validation set (AHHU). In exploratory external-cohort comparisons, the 5 mm peritumoural T2-FLAIR + ASL model showed the highest observed AUC for IDH prediction (0.977, 95% CI 0.900-1.000), exceeding the 10 mm ASL model ( p < 0.001) and the 3 mm T1WI-CE + T2-FLAIR model ( p = 0.0229). For MGMT prediction, the 3 mm T1WI-CE + T2-FLAIR + ASL model yielded an AUC of 0.800 (95% CI 0.500-1.000). In paired-bootstrap comparisons, it exceeded the tumour-core T2-FLAIR model ( p = 0.018) but not the 5 mm T2-FLAIR model ( p = 0.110). These between-model external comparisons were hypothesis-generating and were not used to refit any model. Peritumoural radiomics showed promising performance for IDH prediction. The 5 mm region had the highest observed performance for IDH in this cohort, whereas the apparent advantage of the 3 mm model for MGMT remains exploratory because of the small external sample and lack of consistent superiority across comparators. Independent validation in larger cohorts is required.