A Privacy-Preserving Federated Learning Framework for Generalizable CBCT to Synthetic CT Translation in Head and Neck

Technical Summary: A Privacy-Preserving Federated Learning Framework for Generalizable CBCT to Synthetic CT Translation in Head and Neck

Problem Statement
Cone-beam computed tomography (CBCT) is a standard modality for image-guided radiotherapy (IGRT) in head and neck treatments, offering volumetric imaging within the treatment suite. However, CBCT images suffer from increased noise, limited soft-tissue contrast, and reconstruction artifacts, leading to unreliable Hounsfield Unit (HU) values that preclude direct dose calculation. While deep learning (DL) methods have successfully generated synthetic CTs (sCT) from CBCT to address this, existing approaches face significant barriers to clinical scalability. These include institutional heterogeneity, scanner-dependent variations, and diverse imaging protocols, which often necessitate single-site model training. Furthermore, strict data privacy regulations (e.g., GDPR, HIPAA) prevent the centralization of medical imaging data required to train robust, generalizable models across multiple institutions.

Methodology
This study proposes a cross-silo horizontal federated learning (FL) framework to synthesize sCT from CBCT in the head and neck region, extending the previously developed FedSynthCT-Brain framework to a new imaging modality.

• Data and Federation Setup: The study utilized the public SynthRAD2025 challenge dataset, involving three participating centers (B, C, and E) for federated training and one center (A) reserved exclusively for external validation. The dataset included paired CBCT and CT scans with binary segmentation masks. Notable heterogeneity existed among centers, particularly Center E, which used a different scanner (Varian TrueBeam OBI vs. Elekta XVI) and exhibited distinct acquisition parameters (e.g., exposure times, resolution).
• Preprocessing: A standardized, decentralized preprocessing pipeline was applied locally at each client to ensure consistent spatial resolution (256 × 256 × 256) and intensity harmonization. This involved LUT-based normalization for CBCTs (range [-800, 2000]) and clipping for CTs (range [-1000, 1500]) to mitigate metal implant artifacts. An additional deformable registration step was applied to test data to minimize evaluation bias caused by misalignment.
• Deep Learning Architecture: The core model is a conditional Generative Adversarial Network (cGAN) based on the Pix2Pix architecture.
  • Generator: A 2D U-Net with eight encoder-decoder blocks and skip connections, incorporating dropout to prevent overfitting.
  • Discriminator: A 2D PatchGAN designed to assess local realism.
  • Training Strategy: The model employed a Random-Multi2D sampling approach, presenting randomized axial, sagittal, and coronal slices to enhance generalizability. Data augmentation (flipping, rotation, intensity shifts) was applied locally.
  • Loss Functions: The generator was optimized using a combination of adversarial loss (sigmoid cross-entropy) and pixel loss (Masked Mean Absolute Error, L1), weighted by $\lambda_{pixel} = 100$.
• Aggregation Strategy: The framework utilized a hybrid aggregation strategy combining FedAvg (Federated Averaging) on the server side to weight client contributions by dataset size, and FedProx on the client side. FedProx introduced a proximal term ($\mu = 3$) to the local objective function to penalize large deviations from global weights, thereby mitigating update divergence caused by non-IID data distributions.

Key Results
The federated model demonstrated robust generalization across participating centers and to the external validation set (Center A) without requiring additional training or fine-tuning.

• Quantitative Performance:
  • External Validation (Center A): MAE of 75.22 ± 11.81 HU, SSIM of 0.904 ± 0.034, and PSNR of 33.52 ± 2.06 dB.
  • Training Centers:
    • Center B: MAE 85.90 ± 7.10 HU, SSIM 0.882 ± 0.022.
    • Center C: MAE 64.38 ± 13.63 HU, SSIM 0.922 ± 0.039.
    • Center E: MAE 74.22 ± 13.23 HU, SSIM 0.915 ± 0.013.
• Qualitative Analysis: Visual inspection confirmed that generated sCTs were anatomically consistent with ground-truth CTs. However, the analysis revealed that the primary source of residual errors and metric degradation was not the generative model itself, but rather registration errors and anatomical variations between the input CBCT and the ground-truth CT (e.g., changes in the oral cavity).
• Generalization: Despite Center E using a completely different scanner and protocol, the model maintained performance comparable to other centers, confirming the framework's ability to handle scanner-dependent variations.

Significance and Claims
The paper claims to demonstrate the technical feasibility of using federated learning for CBCT-to-sCT synthesis in the head and neck region while strictly preserving data privacy. The study establishes that:

1. Collaborative model training across institutions is possible without sharing raw patient data, adhering to privacy regulations.
2. A federated approach can yield models that generalize effectively across diverse imaging protocols and scanner types, overcoming the limitations of single-site training.
3. The proposed framework extends the FedSynthCT paradigm from MRI to CBCT, addressing a critical need for robust, center-agnostic models in adaptive radiotherapy.

The authors modestly note that while the model performs well, its clinical integration for dose calculation was not evaluated in this study due to the lack of treatment planning data in the SynthRAD2025 dataset. Additionally, they acknowledge that errors in regions with extreme HU values (air pockets, dense bone) persist, likely due to underrepresentation in training data and registration inaccuracies, suggesting future work in targeted training strategies.