Overcoming data scarcity through multi-center federated learning for organs-at-risk segmentation in pediatric upper abdominal radiotherapy

Researchers demonstrated that federated learning enables multiple medical centers to collaboratively train pediatric organ segmentation models without sharing sensitive patient data. The approach matched local performance while significantly improving cross-center robustness for CT-based radiotherapy planning, addressing a critical gap in pediatric cancer care where data scarcity has limited model development.

This research addresses a fundamental challenge in medical AI: the inability to train robust models when patient data remains siloed within institutions due to privacy regulations and technical constraints. By implementing federated learning across two European centers, researchers proved that institutions can collectively improve model performance without transferring protected health information. The study evaluated segmentation of 19 organs-at-risk across 310 pediatric CT images, demonstrating that FL models achieved superior cross-center generalization compared to locally-trained alternatives, with measurable performance gains across multiple metrics.

The broader context reveals a persistent problem in medical AI development. Pediatric-specific models consistently underperform because fewer training cases exist compared to adult populations, yet regulations like GDPR prevent convenient data centralization. Federated learning circumvents this bottleneck by distributing model training across institutional boundaries while maintaining local data governance. This approach aligns with growing regulatory expectations around privacy-preserving AI.

For healthcare systems and medical device developers, FL demonstrates a pathway to build better clinical tools without the compliance and logistics burden of traditional data sharing agreements. Radiotherapy centers can now participate in collaborative model improvement while protecting patient confidentiality. The technology reduces deployment friction for AI solutions in specialized medical domains where data fragmentation has historically prevented innovation.

Future developments will likely focus on expanding FL implementations across more centers, exploring additional anatomical regions, and optimizing computational efficiency for resource-constrained healthcare settings. Real-world deployment challenges around infrastructure standardization and quality assurance remain areas requiring continued attention as institutions move beyond research pilots to operational systems.

• →Federated learning enables multi-center model training without requiring institutions to share sensitive patient data across firewalls
• →FL-trained models achieved superior cross-center generalization compared to locally-trained models for pediatric organ segmentation
• →The approach maintains privacy compliance while addressing data scarcity constraints that limit pediatric-specific medical AI development
• →FL models demonstrated robustness to patient orientation variations and reduced false-positive segmentations compared to baselines
• →Real-world implementation across European medical centers validates federated learning's feasibility for clinical radiotherapy workflows