Improving Pediatric Glioma Segmentation on Multi-Parametric MRI Via Federated Learning Using Adult Patient Data: A Benchmark Study
Abstract
Purpose
To evaluate whether a federated learning (FL) scheme that leverages adult glioma patient data improves multi-parametric MRI (mp-MRI) based pediatric glioma segmentation.
Methods
Pediatric and adult cohorts included 99 and 1251 retrospective patients, respectively. For each patient, the mp-MRI protocol comprised four sequences: T1, post-contrast T1 (T1+C), T2, and T2-FLAIR. Tumor annotations were reformulated into three subregions: enhancing tumor (ET), tumor core (TC), and whole tumor (WT). A baseline pediatric segmentation model (MP) was developed using a 3D U-Net to simultaneously segment ET/TC/WT, with 69/10/20 pediatric patients used for training/validation/testing. To improve glioma segmentation by MP, two additional paradigms were studied: (1) a centralized 3D U-Net trained on pooled pediatric+adult data (MP+A), while preserving pediatric samples’ original train/validation/test roles; and (2) a two-client FL scheme (MFL), with pediatric-only and adult-only 3D U-Net models trained on separate clients and a coordinating server aggregating updates via FedAvg-style optimization for up to 20 communication rounds. Performance was evaluated exclusively on the pediatric test set using Dice, F1-score, and pixel-wise precision and recall (sensitivity).
Results
MP achieved acceptable pediatric segmentation (ET/TC/WT Dice: 0.719/0.783/0.820; F1: 0.732/0.815/0.836). MP+A modestly improved WT performance (Dice=0.883; F1=0.892) but degraded ET and TC segmentation (ET: Dice=0.615, F1=0.612; TC: Dice=0.737, F1=0.758). Notably, the high precision (0.906) yet low recall (0.651) in TC suggests that MP+A became overly conservative, likely due to pediatric-adult domain interference. In contrast, MFL improved segmentation across all subregions (ET: Dice=0.724, F1=0.743; TC: Dice=0.805, F1=0.825; WT: Dice=0.865, F1=0.873).
Conclusion
Federated learning effectively leveraged adult data to improve pediatric glioma segmentation and outperformed centralized data pooling, which may dilute pediatric-specific characteristics. This framework provides a practical strategy for neuro-oncology imaging studies under limited data availability.