Transformer-based decision support with CNN and ViT for prognostic risk stratification following adjuvant chemotherapy in high-grade glioma: a multicenter study.

Background

High-grade gliomas (HGGs) exhibit marked intratumoral heterogeneity, which is associated with variable survival outcomes among patients receiving standard adjuvant chemotherapy. Conventional methods relying on molecular biomarkers are constrained by technical limitations and tumor heterogeneity. This study aimed to develop a multi-channel 2.5D deep learning framework that integrates Convolutional Neural Network (CNN) and Vision Transformer (ViT) features to noninvasively stratify HGG patients based on expected survival outcomes following postoperative adjuvant temozolomide (TMZ) chemotherapy and to explore the underlying molecular mechanisms.

Methods

Multicenter MRI data (T1-CE/T2-FLAIR; n=507) were analyzed using habitat clustering to define tumor subregions. A 2.5D multi-channel strategy was employed to extract features via ResNeXt101_32 × 8d (CNN) and ViT, which were then integrated using a Transformer fusion model. Transcriptomic analyses, including differential expression, gene set enrichment analysis (GSEA), and protein-protein interaction (PPI) network analysis, were performed to biologically validate the imaging-based predictions. Model performance was assessed by the area under the curve (AUC), accuracy, sensitivity, specificity, and survival analysis (Kaplan-Meier/log-rank test).

Results

The fusion model demonstrated superior performance, consistently outperforming both CNN and ViT models across training (AUC = 0.950, 95% CI: 0.917-0.982), internal-test (0.828, 0.720-0.937), and external-validation (0.810, 0.710-0.911) cohorts. Transcriptomic validation confirmed significant enrichment of cell cycle and DNA damage response pathways in the low-risk patient group, with CHEK1, CDK1, and FEN1 identified as key regulatory genes.

Conclusion

This approach demonstrates robust, non-invasive performance in stratifying HGG patients into prognostically distinct subgroups with differential survival trajectories following post-radiotherapy TMZ chemotherapy. The association with DNA damage response pathways provides biological interpretability, supporting prognostic risk stratification and providing a reference for postoperative surveillance planning.