Survival analysis of diffuse glioma patients using transformer-based enhanced radiomics features

Diffuse glioma prognosis is traditionally addressed using the Cox proportional hazards model, assuming linear risk relationships. To overcome this limitation, we used a transformer-based approach to model complex nonlinear relationships among risk factors. Two open MRI datasets were used: the University of California San Francisco Preoperative Diffuse Glioma MRI (UCSF-PDGM) ( n = 494, train/internal validation cohort) and Burdenko’s Glioblastoma Progression Dataset (BGPD) ( n = 167, validation cohort). Radiomics features were extracted from four MRI modalities: T1-weighted, T1-weighted with contrast, T2-weighted, and T2-weighted fluid-attenuated inversion recovery images. These features were enhanced through embedding within the transformer framework. Cox negative partial log-likelihood loss was used to generate risk scores based on the enhanced radiomics features. In the UCSF-PDGM internal validation cohort, the model recorded a concordance index (C-index) of 0.686 (95% confidence interval [CI], 0.628–0.742), whereas in the BGPD external validation cohort, the C-index was 0.587 (95% CI, 0.516–0.656). Compared to the DeepSurv and Cox Elastic-Net models, our model achieved numerically higher C-indices in both cohorts. Additionally, exploratory attention analysis identified correlations between attention weights and model-predicted risk scores. Our findings suggest that Transformer-enhanced radiomics may support survival risk stratification in diffuse glioma, but external discrimination remains modest and further validation is required.