164 Feature Extraction and Tabular Foundation Modelling for Clinically Integrated Meningioma Grade Prediction

Abstract Introduction Accurate differentiation between low-grade (WHO grade 1) and high-grade (WHO grades 2–3) meningiomas is central to surgical planning, adjuvant radiotherapy decisions, and determination of surveillance intervals. Although radiomics-based machine learning approaches report high AUC values, balanced accuracy is frequently under-reported, limiting interpretability in class-imbalanced datasets. Furthermore, models requiring multi-sequence MRI inputs can impede integration into routine neuroradiology workflows. We therefore evaluated a tabular foundation model for binary grade prediction, with specific emphasis on deployability within standard clinical imaging pathways, including a reduced-input model using post-contrast T1-weighted imaging (T1ce) alone. Methods The BraTS 2023 meningioma dataset (T1, T1ce, T2, FLAIR) with neuroradiologist-reviewed segmentations was analysed (n = 701; 526 low-grade, 175 high-grade). Volumetric, morphological, and radiomic features were extracted from tumour sub-compartments using PyRadiomics, alongside patient age and sex. Low-variance features were removed. TabPFN2.5 was implemented using 5-fold stratified cross-validation with 80:20 train–test splits, and decision thresholds were optimised for balanced accuracy. A second model was trained using T1ce-derived features only to reflect standard surveillance imaging. Performance was assessed using ROC AUC and balanced accuracy. Results The multi-sequence model achieved mean ROC AUC 0.760 ± 0.051 and balanced accuracy 0.705 ± 0.032. The T1ce-only model demonstrated comparable internal performance (BraTS: ROC AUC 0.765; balanced accuracy 0.704). However, performance decreased on an independent clinical cohort (ROC AUC 0.639; balanced accuracy 0.555), indicating domain shift between curated challenge data and real-world imaging heterogeneity. Conclusion TabPFN demonstrates competitive grade discrimination while enabling a streamlined, PACS-compatible workflow using T1ce alone, supporting integration into routine meningioma surveillance and preoperative risk stratification pathways.