FedKBP: Federated dose prediction framework for knowledge-based planning in radiation therapy.

Area of Science:

• Medical Physics
• Artificial Intelligence
• Radiotherapy

Background:

• Dose prediction is crucial for knowledge-based planning (KBP) in radiotherapy.
• Deep learning methods require large datasets, necessitating data sharing.
• Federated learning (FL) offers a privacy-preserving approach for collaborative model training.

Purpose of the Study:

• To evaluate the performance of federated learning (FL) for deep learning-based dose prediction.
• To compare FL against centralized and individual training strategies.
• To assess the impact of data distribution (IID vs. non-IID) on FL performance.

Main Methods:

• Developed the FedKBP framework to train dose prediction models.
• Simulated FL and individual training using 8 sites from the OpenKBP dataset.
• Implemented Independent and Identically Distributed (IID) and non-IID data distributions to test model robustness.

Main Results:

• FL consistently outperformed individual training in speed and accuracy.
• FL achieved performance comparable to centralized training under IID conditions.
• Non-IID data distribution led to performance disparities, with larger sites outperforming smaller ones by up to 19%.

Conclusions:

• Federated learning is a viable, privacy-preserving alternative to centralized training for dose prediction.
• Data heterogeneity in FL (non-IID) necessitates advanced methods beyond simple averaging.
• Collaboration via FL is essential for improving dose prediction accuracy in radiotherapy planning.