Deep learning-based cross-attention fusion of multimodal MRI for survival prediction and risk stratification in IDH-wildtype glioblastoma: a multicenter study

Abstract

Background Glioblastoma (GBM) exhibits profound molecular and spatial heterogeneity, complicating prognostic evaluations. While multiparametric MRI provides crucial multidimensional biological information, conventional end-to-end deep learning integration strategies, such as early or late fusion, often fail to capture complex nonlinear cross-modal interactions. We aimed to systematically evaluate a cross-attention fusion (CAF) architecture for GBM survival prediction and quantify its incremental prognostic value relative to existing clinical tools. Methods In this multicenter retrospective study, 386 adults with IDH-wildtype, WHO grade 4 GBM were assembled from an institutional cohort (n = 226), the Chinese Glioma Genome Atlas (CGGA, n = 62), and The Cancer Genome Atlas (TCGA, n = 98). Using a unified 3D ResNet-18 backbone, we compared single-modality models, early fusion, late fusion, and CAF on preoperative T1-weighted, contrast-enhanced T1-weighted (T1CE), and T2-weighted MRI, and integrated the resulting deep learning risk score with routine clinical variables through multivariable Cox regression. Performance was assessed using Harrell’s C-index, time-dependent AUC, and decision curve analysis. Results CAF showed numerically higher, more consistent C-index trends than early fusion, late fusion, and single-modality models (pooled C-index 0.629, 95% CI 0.594–0.664), although pairwise differences in time-dependent AUC were not statistically significant. Integrating clinical variables raised the pooled C-index to 0.691 (95% CI 0.660–0.721) in the treatment-era model, with comparable performance across the three cohorts (Local 0.688; CGGA 0.716; TCGA 0.689); a pre-treatment configuration excluding adjuvant therapy yielded a pooled C-index of 0.642. Under leave-one-cohort-out external validation, the combined model retained significant risk stratification in all held-out cohorts (C-index 0.63–0.71; all log-rank P < 0.01), albeit with attenuated discrimination. The deep learning risk score remained independent after multivariable adjustment (HR 1.41 per SD, 95% CI 1.26–1.57; P < 0.001). Kaplan–Meier analysis confirmed significant high- versus low-risk separation in all cohorts, and decision curve analysis showed greater net benefit than clinical-only and deep-learning-only models. Conclusion The CAF-derived risk score offers prognostic information complementary to routine clinical variables, representing a promising noninvasive tool for individualized risk stratification when molecular profiling is incomplete or unavailable; these findings warrant prospective external validation before clinical use.