Author: Jingyun Chen, Yading Yuan 👨🔬
Affiliation: Columbia University Irving Medical Center, Department of Radiation Oncology 🌍
Purpose: To evaluate centralized and decentralized strategies for federated head and neck tumor segmentation on PET/CT.
Methods: We utilized training data from the HEad and neCK TumOR segmentation (HECKTOR) 2021 challenge, which comprises PET and CT imaging datasets. Our study involved 4-site federated learning (FL) experiments across four HECKTOR-contributing sites, with cases from a fifth site reserved for external testing. The performance evaluation was carried out under two baseline conditions: the Pool Model (PM), utilizing centralized training data, and Individual Model (IM), trained locally at each site. Additionally, we assessed five state-of-the-art FL methods: three centralized approaches (FedAvg, FedProx, and FedPIDAvg) and two decentralized approaches (BrainTorrent (BT) and ProxyFL). We evaluated model performance using three key metrics: averaged Dice Similarity Coefficient (DSC), 95% Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD). All experiments were conducted using four NVIDIA GPUs.
Results: On average across the internal sites, all FL methods consistently outperformed the IM baseline, achieving higher DSC and lower HD95 and ASSD. Among the centralized FL methods, the more advanced FedProx and FedPIDAvg approaches demonstrated superior performance over the original FedAvg, achieving DSC of 77.91% and 77.64% respectively, which were close to the PM baseline (78.52%). Similarly, on the external testing site, all FL methods outperformed the IM baseline. Notably, the centralized FedPIDAvg method and the decentralized BT method both delivered performances comparable to the PM baseline, with DSC of 73.48%, 73.32%, 74.80% respectively.
Conclusion: FL has the potential to achieve comparable performance to centralized training. While FedAvg remains the most widely used FL method, more advanced approaches like FedPIDAvg have demonstrated superior performance. Decentralized FL methods offer the advantage of not relying on a central server, but their performance can vary significantly. This highlights the need for the development of more sophisticated and robust decentralized FL methods.