Peer-reviewed article

LGG-05. IMAGING-BASED DEEP LEARNING FOR EVENT-FREE SURVIVAL PREDICTION IN PATIENTS WITH PEDIATRIC LOW-GRADE GLIOMA

· 1 Jun 2023

Abstract

Abstract Pediatric low-grade gliomas (pLGGs) have heterogeneous outcomes, making individualized clinical decision-making challenging. Imaging-based biomarkers for progression and/or recurrence may better risk-stratify pLGG patients and guide management. We developed and externally validated a deep learning (DL) algorithm to predict event-free survival (EFS) in pLGG using multi-institutional databases. We collected linked, T2-weighted diagnostic Magnetic Resonance Images (MRI) and clinical data for patients with WHO grade 1-2 pLGG (diagnosed from 1991-2017) from the Children’s Brain Tumor Network (CBTN, n=176, median follow-up: 3.7 years, IQR=2-6.8 years) for model development and internal validation, and Boston Children’s Hospital (BCH, n=219, median follow-up: 8.8 years, IQR=4-13.4 years) for external validation. Leveraging a self-supervised learning approach, we trained a DL Unet model for tumor segmentation and used it for extracting image features. We combined 12 DL-imaging features with three clinical variables (age, sex, and surgical resection status (gross total/subtotal/no resection) and trained a Deep Cox Mixtures(DCM) model for survival analysis. We externally validated the performance of three models: clinical, DL-imaging, and combined DL-imaging and clinical. The combined DL-imaging and clinical model outperformed the clinical alone and DL-imaging alone models for prediction of 3-year EFS (time-dependent AUC: 0.85, 0.73, 0.53, respectively) with improved calibration (Integrated Brier Score: 0.08, 0.13, 0.15, respectively). Using the combined DL-imaging and clinical model output, patients were divided into low and high-risk groups based on mean event probability (3-year EFS 53% vs 86%, p=0.02). Variables associated with poor EFS were imaging score, lower age, and lack of gross total resection. In conclusion, deep learning-derived, MRI-based imaging features may improve prediction of event-free survival and risk stratification for patients with pLGG.