STELLA

The project integrates the CAFEIN^® federated learning (FL) platform with the STELLA (Smart Technologies to Extend Lives with Linear Accelerators ) radiotherapy treatment system to enable multi institutional collaboration in radiation oncology without centralizing raw data. Through privacy preserving learning, robust anomaly detection, and standardized validation, the framework improves treatment planning accuracy, streamlines clinical workflows, and supports safer delivery of care while meeting regulatory and security requirements.

Radiotherapy services must deliver precise and timely care, yet many clinics operate with limited data, diverse hardware and software, and strict privacy obligations. Centralizing sensitive records is often infeasible due to regulation and data sovereignty, which slows knowledge sharing and model improvement across centers.

This project proposes a scalable FL framework that connects local PACS and oncology information systems to a secure training and inference environment. Sites keep records on premises while participating in collaborative model development through encrypted updates. The architecture supports common radiotherapy tasks such as image segmentation, dose prediction, plan quality assessment, and workflow optimization. It includes preprocessing pipelines, quality assurance checkpoints, and reference evaluation protocols so that results are comparable across institutions and vendor platforms.

Interoperability is achieved through open standards including DICOM and vendor neutral interfaces, preserving existing STELLA functionality and minimizing disruption to clinical operations. The framework addresses practical constraints in smaller clinics with lightweight clients, resource aware scheduling, and fault tolerant synchronization. Governance features such as role based access control, audit logs, and policy aware deployment help satisfy HIPAA, GDPR, and clinical validation requirements. The result is a collaborative approach that raises care quality, reduces variation, and broadens access to advanced decision support without moving raw data.

Our role in the project

CERN leads the integration of CAFEIN^® with STELLA and provides the privacy and security foundations for cross site learning including secure aggregation, optional differential privacy, access control, and auditable workflows. CERN delivers connectors for PACS and treatment planning systems, orchestrates hierarchical training across participating centers, and supplies MLOps tooling for reliable on premises deployment and updates. CERN also curates radiotherapy specific pipelines for segmentation, dose and outcome modeling, and plan quality analytics, and coordinates benchmarking, external validation, documentation, and training to support ICEC led pilots and broader adoption.