FedSynthCT-Brain: A Federated Learning Framework for Multi-Institutional Brain MRI-to-CT Synthesis

TL;DR

This work introduces FedSynthCT-Brain, a cross-silo horizontal Federated Learning framework for MRI-to-sCT synthesis in the brain, addressing generalisability and data privacy by training a U-Net model across four centres and testing on an unseen fifth. The study systematically benchmarks architectures and aggregation strategies, finding that FedAvg+FedProx provides the best balance of convergence speed and image quality, achieving a median MAE of 102.0 HU and SSIM of 0.89 on unseen data. Results demonstrate that federated sCTs are clinically plausible across heterogeneous scanners and protocols, with MAE and SSIMs indicating robust soft-tissue accuracy and reasonable bone representation, while highlighting artefacts in specific regions. The approach supports private, multi-institution collaboration for MRI-guided radiotherapy planning and lays groundwork for expanding federated brain sCT to additional clinics and dosimetric validation.

Abstract

The generation of Synthetic Computed Tomography (sCT) images has become a pivotal methodology in modern clinical practice, particularly in the context of Radiotherapy (RT) treatment planning. The use of sCT enables the calculation of doses, pushing towards Magnetic Resonance Imaging (MRI) guided radiotherapy treatments. Deep learning methods for MRI-to-sCT have shown promising results, but their reliance on single-centre training dataset limits generalisation capabilities to diverse clinical settings. Moreover, creating centralised multi-centre datasets may pose privacy concerns. To address the aforementioned issues, we introduced FedSynthCT-Brain, an approach based on the Federated Learning (FL) paradigm for MRI-to-sCT in brain imaging. This is among the first applications of FL for MRI-to-sCT, employing a cross-silo horizontal FL approach that allows multiple centres to collaboratively train a U-Net-based deep learning model. We validated our method using real multicentre data from four European and American centres, simulating heterogeneous scanner types and acquisition modalities, and tested its performance on an independent dataset from a centre outside the federation. In the case of the unseen centre, the federated model achieved a median Mean Absolute Error (MAE) of $102.0$ HU across 23 patients, with an interquartile range of $96.7-110.5$ HU. The median (interquartile range) for the Structural Similarity Index (SSIM) and the Peak Signal to Noise Ratio (PNSR) were $0.89 (0.86-0.89)$ and $26.58 (25.52-27.42)$, respectively. The analysis of the results showed acceptable performances of the federated approach, thus highlighting the potential of FL to enhance MRI-to-sCT to improve generalisability and advancing safe and equitable clinical applications while fostering collaboration and preserving data privacy.

Figures (11)

• Figure 1: Visual comparison of image appearance across datasets from multiple Centres. The FOV of the images from Centre A was limited to the level of the mouth and included partial cut-off of the ears. Images from centre B were acquired using a PET-MRI hybrid scanner. Furthermore the higher CT slice thickness introduced partial volume artefacts in the skull bone reconstruction. Images coming from centre C exhibited higher spatial resolution in the cranial caudal direction, but presented other problems, including partial cropping at the top of the skull and metal artefacts. Centres D and E--derived from SynthRAD2023 challenge--shared similar imaging regions with less bias fields compared to Centre C.
• Figure 2: The figure illustrates the background masking for each institution. By eliminating potential misleading features for the model, the masking approach limits the prediction of the model to areas of interest and facilitates the harmonisation of the model's knowledge within the federation. The masking process was conducted within the individual Centre. The parts of interest were established beforehand, as would be done in a real-world context.
• Figure 3: Federated setup. Each centre (also called client) has its own local data and trains a model on the local dataset. The weights of the local model are exchanged with the server, which, after the aggregation process, returns a federated model to the clients. The server also hosts a dataset that is not used during training, but acts as a benchmark to determine the generalisation ability of the model.
• Figure 4: Proposed training methodology. The pre-processed and augmented 3D (MRI) volume is subdivided into 2D slices, extracted along the three anatomical planes (sagittal, axial, coronal). These slices are then subjected to a random shuffling process, ensuring that the model is agnostic with respect to both the anatomical plane and the imaging sequence. The shuffled batch of images is then provided as input to the model.
• Figure 5: Reconstruction of the sCT. The axial, coronal, and sagittal slice predictions are rearranged to construct three distinct volumes. Subsequently, a median voting approach is adopted using the three volumes in order to enhance the final result.