FedSynthCT-Brain: A federated learning framework for multi-institutional brain MRI-to-CT synthesis

Area of Science:

• Medical Imaging
• Radiotherapy
• Machine Learning

Background:

• Synthetic Computed Tomography (sCT) is crucial for radiotherapy dose calculations and improving MRI-PET attenuation correction.
• Current deep learning MRI-to-sCT methods struggle with generalization due to single-center datasets and raise privacy concerns for multi-center data aggregation.

Purpose of the Study:

• To introduce FedSynthCT-Brain, a federated learning approach for MRI-to-sCT image generation in brain imaging.
• To address the limitations of single-center training and privacy concerns in developing robust MRI-to-sCT models.

Main Methods:

• Implemented a cross-silo horizontal federated learning (FL) strategy to train a U-Net-based deep learning model collaboratively across multiple institutions.
• Validated the FL model using real multi-center data with heterogeneous scanner types and acquisition modalities.

Main Results:

• The federated model demonstrated acceptable performance on an independent, unseen center, achieving a median MAE of 102.0 HU.
• Quantitative metrics including SSIM (median 0.89) and PSNR (median 26.58) indicate the model's effectiveness in generating accurate sCT images.

Conclusions:

• Federated learning shows significant potential for enhancing MRI-to-sCT generalizability and advancing safe, equitable clinical applications.
• This approach fosters multi-institutional collaboration while maintaining strict data privacy for improved radiotherapy treatments.