Multimodal Radiomics with Graph Neural Networks for Predicting Glioma Methylation Subtypes

Abstract

Accurate prediction of glioma methylation subtypes is crucial for treatment planning, yet current methods rely on invasive tissue sampling. This study developed a non-invasive approach integrating multimodal radiomics with graph neural networks (GNNs) to predict methylation status from MRI and PET imaging in 148 glioma patients. The GNN architecture represented tumor regions as interconnected nodes with attention mechanisms emphasizing discriminative features across modalities. Using five-fold cross-validation and external validation (n=46), the multimodal GNN achieved 89.7% accuracy, 91.3% sensitivity, 88.6% specificity, and 0.923 AUC, significantly outperforming conventional approaches (random forest: 76.4%, SVM: 73.9%). Performance remained robust across WHO grades, with multimodal integration providing 7.2% AUC improvement over single-modality approaches. Feature analysis identified textural heterogeneity and metabolic parameters as key predictive biomarkers. External validation confirmed generalizability with 84.8% accuracy. This research demonstrates the potential of graph-based deep learning on multimodal imaging for non-invasive molecular subtyping of gliomas, potentially eliminating the need for invasive procedures in treatment planning.