Related Experiment Video
Updated: Aug 5, 2026
Live-3D-Cell Immunocytochemistry Assays of Pediatric Diffuse Midline Glioma
Published on: November 11, 2021
Uncertainty-Aware Risk Stratification in Pediatric Low-Grade Glioma Using Multimodal Data
Fadel Batal1, Bhavyasri Vunnava1, Adam Kraya1
1From the Center for Data-Driven Discovery in Biomedicine (D3b) (F.B., B.V., A.K., D.G., A.F., K.R., S.V., D.C., S.R., O.F., A.K., A.B., P.B.S., A.R., A.V., A.N., A.F.K.), Neurosurgery (P.B.S., A.R., A.F.K.), Radiology (A.V.), Children's Hospital of Philadelphia, Philadelphia, PA, USA; Department of Bioengineering (F.B., D.C., A.F.K.), Neurosurgery (P.B.S., A.R., A.F.K.), Perelman School of Medicine, Radiology (A.V., A.N.), Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA and Transitional Year Residency Program, Southeast Health Medical Center, Dothan, AL, USA.
This study developed an AI framework integrating imaging, molecular, and clinical data for better risk stratification in pediatric low-grade glioma (pLGG). The multimodal approach improved accuracy and transparency in predicting patient outcomes.
Area of Science:
- Neuro-oncology
- Medical imaging
- Artificial intelligence
Background:
- Pediatric low-grade glioma (pLGG) exhibits significant heterogeneity, complicating risk stratification.
- Current methods struggle to accurately predict outcomes due to biological and clinical variability.
Purpose of the Study:
- To develop an uncertainty-aware multimodal survival framework for improved risk stratification in pLGG.
- To integrate deep learning features from T2-weighted MRI, molecular subtype, and clinical data.
Main Methods:
- Extracted deep learning features from T2-weighted MRI using a fine-tuned model.
- Developed a clinico-radiomic model (DL-M1) and a clinical-molecular model.
- Created a multimodal model (DL-M2) via late fusion and used bootstrap resampling for uncertainty quantification.
Main Results:
- DL-M1 achieved comparable performance to radiomic pipelines (C-index 0.70-0.73).
- Integrating molecular subtype improved performance in the replication cohort (p=0.016).
- The multimodal model significantly reduced prediction uncertainty (75% reduction).
Conclusions:
- Multimodal survival modeling integrating molecular subtype enhances pLGG risk stratification.
- Deep learning from T2-weighted MRI offers a streamlined alternative to complex radiomics.
- Uncertainty-aware multimodal modeling provides a transparent approach for pLGG risk stratification.