Foundation model-enhanced multimodal survival prediction in glioblastoma across multi-institutional cohorts

Abstract MRI foundation models (FMs) have shown potential for improving performance of neuroimaging tasks, but their value specifically in Glioblastoma overall survival risk prediction remains unclear. In this study, we explored foundation model initialization for preoperative risk prediction using publicly available structural MRIs from 1007 patients across three glioblastoma cohorts: UPENN-GBM, UCSF-PDGM and TCGA-GBM. The primary analysis finetuned a Swin vision transformer, initialized with BrainSegFounder weights, and compared performance to matched random initialization, and a radiomics baseline. We additionally evaluated recently released FMs, including BrainMVP, BrainIAC, and TRIAD, within the same adaptation strategy, and explored multimodal integration of diffusion tensor imaging-based risk predictions, age, extent of surgical resection, and MGMT methylation status into predictions. In 10-fold UPENN-GBM cross-validation, BrainSegFounder improved mean C-index compared to matched training from scratch and radiomics (0.667 ± 0.053 vs 0.649 ± 0.040 and 0.612 ± 0.044) and achieved time-dependent AUROCs between 0.766 ± 0.089 and 0.799 ± 0.073 across survival horizons under one year. All tested FMs improved over matched random initialization. Within complete-case subsets, multimodal integration improved performance (C-Index 0.691 ± 0.053). Leave-one-cohort-out validation showed performance in external settings consistent with the broader GBM survival prediction literature. These findings suggest incremental value of FM initialization for GBM risk stratification and supports continued benchmarking, adaptation, and validation.