A privacy-preserving federated learning framework for generalizable CBCT to synthetic CT translation in head and neck

Cone-beam computed tomography (CBCT) has become a widely adopted modality for image-guided radiotherapy (IGRT). However, CBCT is characterized by increased noise, limited soft-tissue contrast, and artifacts. These issues result in unreliable Hounsfield unit (HU) values, which limits electron density estimation for direct dose calculation. These issues have been addressed by deriving synthetic CT (sCT) from CBCT, particularly by adopting deep learning (DL) methods. However, existing DL approaches are hindered by institutional heterogeneity, scanner-dependent variations, and data privacy regulations that prevented multi-center data sharing.

To overcome these challenges, we propose a cross-silo federated learning (FL) approach for CBCT-to-sCT synthesis in the head and neck region. This approach extends the original FedSynthCT framework to a different image modality and anatomical region. A conditional generative adversarial network (cGAN) was trained using data from three European medical centers within the SynthRAD2025 public challenge dataset while maintaining data privacy at each institution. A combination of the FedAvg and FedProx aggregation strategies, alongside a standardized preprocessing pipeline, was adopted to federate the DL model.

The federated model effectively generalized across participating centers, as evidenced by the mean absolute error (MAE) ranging from $64.38\pm 13.63$ to $85.90\pm 7.10$  HU, the structural similarity index (SSIM) ranging from $0.88\pm 0.02$ to $0.92\pm 0.04$ , and the peak signal-to-noise ratio (PSNR) ranging from $32.86\pm 0.94$ to $34.91\pm 1.04$  dB. Notably, performance on an external validation dataset of 60 patients yielded comparable metrics: a MAE of $75.22\pm 11.81$  HU, an SSIM of $0.90\pm 0.03$ and a PSNR of $33.52\pm 2.06$ , confirming robust cross-center generalization despite differences in imaging protocols and scanner types, without additional training. Furthermore, a visual analysis of the results revealed that the obtained metrics were influenced by registration errors.

Our findings demonstrated the technical feasibility of FL for CBCT-to-sCT synthesis task while preserving data privacy, offering a collaborative solution for developing generalizable models across institutions without requiring data sharing or center-specific models.