Federated learning overcomes data scarcity for pediatric OAR segmentation

Researchers from Utrecht and Heidelberg demonstrated that federated learning (FL) can overcome data scarcity for developing pediatric-specific organs-at-risk (OAR) segmentation models in upper abdominal radiotherapy. Deep learning models trained on adult data often fail in pediatric patients due to anatomical differences, but pediatric data is scarce and fragmented across institutions. Using CT images from pediatric patients with renal tumors or abdominal neuroblastoma, the team implemented an nnU-Net-based framework to segment 19 OARs. FL enabled privacy-preserving collaborative training across two European medical centers via secure weight exchange on cloud storage, bypassing institutional firewalls. Performance was assessed using the Dice similarity coefficient. The study shows FL's feasibility for robust pediatric auto-contouring without sharing sensitive data.