Multimodal Deep Radiomic and Genomic Modeling for Post-Treatment Survival Prediction In Glioma
Abstract
Purpose
Accurate survival prediction following radiation therapy for glioma remains challenging due to heterogeneous tumor biology and treatment response. This study proposes a multimodal framework integrating deep radiomic features from longitudinal MRI, tumor growth dynamics, clinical variables, treatment dose metrics, and genomic alterations to predict overall survival in post-treatment glioma patients.
Methods
Patients with post-radiation MRI, segmentation masks, survival outcomes, and clinical/genomic data were retrospectively analyzed. For each imaging time point, tumor regions were automatically cropped and resampled to a fixed resolution and processed using a pretrained 3D ResNet-18 (MedicalNet) to extract high-dimensional deep radiomic features. Longitudinal tumor growth rate was computed from serial tumor volumes. Clinical variables included age, sex, race, tumor grade, and biologically effective dose (BED). Genomic markers included IDH1/IDH2 mutations, 1p/19q codeletion, ATRX mutation, MGMT methylation, EGFR amplification, PTEN mutation, CDKN2A/B deletion, TP53 alteration, TERT promoter mutation, BRAF V600E mutation, H3-3A mutation, and chromosome 7 gain/10 loss. Imaging, clinical, dose, and genomic features were concatenated into a unified feature vector per patient. Survival modeling was performed using ElasticNet regression, with risk defined as the negative predicted survival time. Model performance was evaluated using three-fold cross-validation on the training cohort and tested on an independent hold-out set using concordance index (C-index).
Results
The proposed multimodal model successfully integrated imaging-derived deep features with longitudinal growth, clinical, and genomic data. The proposed framework achieved a mean 3-fold cross-validation C-index of 0.66, with a test cohort C-index of 0.69. Kaplan–Meier analysis demonstrated clear separation between low- and high-risk groups, indicating effective risk stratification based on predicted survival risk.
Conclusion
This framework demonstrates the feasibility and potential benefit of combining deep radiomics, tumor growth modeling, and genomic biomarkers for post-radiation survival prediction in glioma. The approach is extensible to adaptive therapy and personalized risk-guided treatment strategies.