Federated learning for enhanced dose-volume parameter prediction with decentralized data.

Area of Science:

• Radiation Oncology
• Machine Learning
• Medical Data Privacy

Background:

• Knowledge-based planning in radiation oncology is limited by data scarcity and challenges in sharing medical data.
• These limitations hinder the widespread adoption of advanced planning techniques.

Purpose of the Study:

• To assess the feasibility of federated learning (FL) to overcome data sharing limitations in radiation oncology.
• To develop a privacy-preserving method for training a centralized model using distributed datasets.

Main Methods:

• A gradient-boosting model predicted bladder and rectum dose-volume metrics (V30Gy, V35Gy, V40Gy) using 273 prostate cancer plans.
• The Federated Averaging algorithm aggregated model weights from 10 simulated clinic subsets.
• Model robustness was tested with varying numbers of sites and imbalanced data distributions.

Main Results:

• The FL model achieved a mean absolute error (MAE) of 4.7% ± 2.9%, significantly lower than individual models (6.5% ± 4.9%) and comparable to a centralized model (4.4% ± 2.8%).
• The FL model demonstrated robustness across different numbers of subsets (5-30) and performed well on imbalanced datasets.
• The FL approach outperformed 36.7% of individual models for bladder and rectum metrics.

Conclusions:

• Federated learning offers a viable solution for knowledge-based planning in radiation oncology, enhancing prediction accuracy without centralizing sensitive patient data.
• FL models maintain high performance even with data scarcity at local sites, proving more robust than individually trained models.