Federated prediction for scalable and privacy-preserved knowledge-based planning in radiotherapy

Deep learning has potential to improve the efficiency and consistency of radiation therapy planning, but clinical adoption is hindered by the limited model generalizability due to data scarcity and heterogeneity among institutions. Although aggregating data from different institutions could alleviate this problem, data sharing is a practical challenge due to concerns about patient data privacy and other technical obstacles.

This work aims to address this dilemma by developing FedKBP+, a comprehensive federated learning (FL) platform for predictive tasks in real-world applications in radiotherapy treatment planning.

We implemented a unified communication stack based on Google Remote Procedure Call (gRPC) to support communication between participants whether located on the same workstation or distributed across multiple workstations. In addition to supporting the centralized FL strategies commonly available in existing open-source frameworks, FedKBP+ also provides a fully decentralized FL model where participants directly exchange model weights to each other through Peer-to-Peer communication. We evaluated FedKBP+ on three predictive tasks using scale-attention network (SA-Net) as the predictive model.

Using 340 cases (training: 200; validation: 40; testing: 100) from the OpenKBP Challenge, a 3D dose prediction model trained with FedAvg algorithm outperformed the model trained on the local data, and achieved predictive accuracy comparable to that of a centrally trained model using pooled data in both independent and identically distributed (IID) and non-IID settings. We further evaluated the performance of FedKBP+ against NVFlare on the task of brain tumor segmentation using 227 cases from eight sites in the 2021 BraTS challenge dataset (training: 152; validation: 27; testing: 48). FedKBP+ surpassed the NVFlare framework in both accuracy and training efficiency. FedAvg with FedKBP+ achieved a DSC of 92.38% in 5.92 hours versus NVFlare's 90.75% in 7.78 hours; FedProx with FedKBP+ reached 91.98% in 6.03 hours against NVFlare's 90.37% in 7.80 hours. Finally, in a robustness study of FedKBP+ for organ segmentation using 384 cases from five sites in the PanSeg dataset (training: 269; validation: 39; testing: 76), Gossip Contrastive Mutual Learning - a novel decentralized FL algorithm - demonstrated strong resilience to site failure without significant accuracy loss even when up to 40% of sites randomly dropped out at each round of federated training (p = 0.9097).

Our results demonstrate that FedKBP+ is highly effective, efficient and robust, showing great potential as a federated learning platform for radiation therapy.