Investigating Prognostic Value of Dynamic Susceptibility Contrast Perfusion MRI-Derived Features for Glioblastoma Survival by Deep Learning

Motivation: State-of-the-art deep learning methods for GBM analysis often overlook hemodynamic information critical to prognosis from dynamic susceptibility contrast perfusion-weighted imaging (DSC-PWI). Goal(s): Find and evaluate DSC-PWI hemodynamic features that enhance survival prediction accuracy for GBM patients beyond 10 months Approach: DSC-PWI hemodynamic features were extracted with HDBNet and combined with clinical and radiomic data in an XGBoost model for predicting survival outcomes. Correlation analysis was conducted between survival outcomes and key clinical and DSC-PWI features. Results: Including DSC-PWI features improved F1 scores, suggesting potential prognostic value for imbalanced survival outcomes in clinical application. Clinical data outperformed both DSC-PWI and radiomic features. Impact: Using our Hierarchical Density-Based Network (HDBNet) to investigate hemodynamic information in dynamic susceptibility contrast perfusion-weighted imaging (DSC-PWI) reveals key features that can enhance GBM prognosis, supporting the importance of including hemodynamic and physiological imaging data in future GBM research