Federated learning-driven intelligent framework for multi-center radiotherapy dose distribution prediction oriented...

Area of Science:

• Medical Physics
• Artificial Intelligence
• Radiotherapy

Background:

• Radiotherapy dose distribution prediction is complex and limited by data silos and planner variability.
• Current deep learning models struggle with generalizability due to single-center datasets and privacy restrictions.

Purpose of the Study:

• To develop a federated learning framework for collaborative, privacy-preserving AI model training across multiple institutions for radiotherapy dose prediction.
• To address challenges of non-independent and non-identically distributed data in multi-center federated learning.

Main Methods:

• Implemented a multi-scale attention U-Net for 3D dose prediction.
• Utilized an adaptive weighted federated aggregation strategy to balance data volume and local model quality.
• Integrated gradient-clipped differential privacy and secure aggregation for enhanced privacy protection.

Main Results:

• Achieved a mean Gamma pass rate of 96.8% across four clinical centers, nearing the centralized training performance (97.5%).
• Outperformed single-center models and standard federated averaging, demonstrating superior generalizability.
• Ablation studies and robustness analyses confirmed the effectiveness of adaptive weighting, attention modules, and fault tolerance.

Conclusions:

• The proposed federated learning framework provides a practical, privacy-preserving solution for AI-driven radiotherapy dose prediction.
• This approach effectively breaks down institutional data silos, facilitating broader AI research and application in radiation oncology.