IMG-15. A Synergistic AI Framework using Graph and Texture for Non-Invasive Survival Prediction in IDH-Wild-Type Glioblastoma Multiforme

BACKGROUND Glioblastoma multiforme (GBM), specifically the IDH-wild-type subtype, is an aggressive brain tumor with a median survival of 12–18 months. Non-invasive estimation of overall survival (OS) is very much crucial for personalized treatment strategies. However, existing methods often fail to capture the complex spatial and structural heterogeneity of GBM. To address this we present an AI-driven framework that consists graph and texture-based features for robust OS prediction. METHODS This study utilized multiparametric MRI (T1, T2, FLAIR, T1GD) from 528 IDH-wild-type GBM patients, which was divided into a discovery cohort (n=400) and a replication cohort (n=128). Tumor subregions were segmented and encoded as graph structures using intra- and inter-slice connectivity. Concurrently, handcrafted texture features were extracted from radiomic descriptors. These features were combined with clinical variables (age, gender). Feature selection was performed using Recursive Feature Elimination with Support Vector Machines (RFE-SVM). A nested 5-fold cross-validation strategy was used for model training and validation. RESULTS The fused feature model significantly outperformed individual feature sets. For the discovery cohort it achieved cross-validated AUC of 0.85 (95% CI: 0.82 - 0.89), accuracy of 0.78 (95% CI: 0.74 – 0.81), and recall of 0.84 (95% CI: 0.80-0.87). The replication cohort confirmed the model’s robustness with an AUC of 0.80 and accuracy of 0.75. Survival stratification was validated via Kaplan-Meier curves (p < 0.0001) and effect sizes (Cohen’s d = 0.80 for discovery; 0.77 for replication). CONCLUSION The proposed Graph and Texture Network offers a powerful non-invasive tool for OS prediction in GBM. By capturing and handling both spatial structure and radiomic detail, this framework enhances prognostic accuracy and holds promise for clinical translation in precision neuro-oncology.