Deep learning predicts overall survival in high-grade meningioma using MRI and clinical data.

BACKGROUND: High-grademeningiomas (WHO II-III) recurfrequentlyand show wide variation in patient survival,yet clinicians still lack practical ways to estimate prognosis beforesurgery.We developed a deep learning survival model integrating T1-weighted MRI with clinical covariates to predict overall survival in high-grade meningioma patients. METHODS: Patients withhistopathologicallyconfirmed WHO grade II or III meningioma treated between 2010 and 2026wereidentifiedfrom a single academic institution. T1-weighted MRI, age, sex, and tumor grade were obtained for each patient. ADeepSurv-style 3D convolutional neural network using an EfficientNet-B0 backbone was fused with a clinical multilayer perceptron to predict survival risk scores.We evaluated model performance withc-indicesobtained from a stratified 5-foldcross-validationscheme. RESULTS: 55 patients were included (29 males [52.7%], 26 females [47.3%]; mean age 62.6 ± 16.0 years). Forty patients (72.7%) had WHO grade II and 15 (27.3%) had WHO grade III meningiomas. The mean follow-up was 67.2 ± 57.0 months, with 26 deaths (47.3%). The model achieved a mean cross-validated c-index of 0.731 ± 0.112 across five folds (range: 0.571-0.895). Training Cox loss ranged from 0.473 to 1.113 across folds, with validation loss ranging from 0.528 to 1.073. Kaplan-Meier analysis of out-of-fold predicted risk scoresdemonstratedseparation between high- and low-risk groups.The distribution of WHO grade was similar between model-predicted high- and low-risk groups (71.4% vs 74.1% Grade II), suggesting the model extracts prognostic information beyond histological grade alone.No tumor segmentation, skull stripping, or data augmentation was performed. CONCLUSIONS: A 3D CNN integrating T1 MRI with clinical covariates predicted overall survival in high-grade meningioma with a mean cross-validated c-index of 0.731.Our results suggest thatMRI-basedmodels can meaningfully contribute to preoperative survival estimationandwarrantvalidation in larger cohorts.