Federated Image Reconstruction

Unlike traditional centralized deep learning reconstruction (DLR) where raw sensor data is pooled, federated reconstruction keeps k-space data (MRI) or sinogram data (CT) strictly local. Each institution trains a local reconstruction network on its own scanner data, and only model gradients or weights are shared with a central aggregation server. This preserves patient privacy while enabling collaborative optimization of the mapping from under-sampled or noisy measurements to high-fidelity diagnostic images.

• Local data stays local: Raw acquisition data never leaves the hospital firewall
• Collaborative optimization: Institutions jointly improve reconstruction quality without data pooling
• Scanner-specific adaptation: Models learn to handle diverse vendor protocols and acquisition parameters

The dominant aggregation strategy is FedAvg (Federated Averaging), adapted for image reconstruction architectures like U-Nets, variational networks, or unrolled optimization networks. Each client trains locally on its own under-sampled/fully-sampled image pairs, then sends model updates to a central server that computes a weighted average. The global model is redistributed for the next round. This iterative process converges toward a reconstruction model that generalizes across diverse scanner geometries and patient populations.

• Weighted aggregation: Larger institutions with more data contribute proportionally
• Multiple communication rounds: Typically 100-500 rounds for convergence on medical imaging tasks
• Non-IID robustness: Handles heterogeneous data distributions across sites

Even sharing model updates can leak information about training data through gradient inversion attacks. Federated reconstruction systems integrate differential privacy (DP) by clipping gradients and adding calibrated Gaussian noise during local training. The privacy budget (ε, δ) quantifies the trade-off between reconstruction fidelity and privacy protection. Typical medical imaging deployments target ε values between 1-10, ensuring that individual patient scans cannot be reconstructed from shared model updates.

• Gradient clipping: Bounds the influence of any single training example
• Noise injection: Calibrated Gaussian noise masks individual contributions
• Privacy-utility trade-off: Lower ε values provide stronger privacy but may reduce reconstruction sharpness

A core challenge in federated reconstruction is domain shift across institutions with different MRI field strengths (1.5T vs 3T), CT detector configurations, or under-sampling patterns. Federated reconstruction frameworks address this through federated domain adaptation layers that learn site-specific normalization parameters locally, while sharing domain-invariant feature extractors globally. This ensures the global model doesn't overfit to any single institution's acquisition protocol.

• Site-specific batch normalization: Local statistics remain private while global weights are shared
• Federated adversarial training: Domain discriminators help learn scanner-invariant representations
• Multi-vendor generalization: Models trained across Siemens, GE, and Philips scanners simultaneously

Reconstruction networks like deep unrolled networks can have millions of parameters, making gradient transmission a bandwidth bottleneck. Federated reconstruction employs gradient sparsification (transmitting only the top-k gradients by magnitude) and quantization (reducing gradient precision from 32-bit to 8-bit or even 1-bit). These techniques reduce communication overhead by 100-1000x without significantly degrading reconstruction quality, enabling practical deployment across hospital networks with limited uplink bandwidth.

• Top-k sparsification: Transmit only 1-10% of gradient values per round
• Stochastic quantization: Compress gradients to low bit-width representations
• Error feedback: Accumulate compression residuals to maintain convergence accuracy

In multi-institutional settings, faulty or malicious nodes could upload corrupted model updates that degrade reconstruction quality—a critical risk for diagnostic imaging. Federated reconstruction systems implement Byzantine-robust aggregation algorithms like Krum, trimmed mean, or median-based aggregation that detect and exclude anomalous updates. This ensures that even if a subset of participating hospitals provides corrupted gradients, the global reconstruction model remains safe and diagnostically reliable.

• Krum algorithm: Selects the update closest to a majority of other updates
• Coordinate-wise median: Replaces each parameter with the median across clients
• Fault tolerance: Typically robust against 20-33% malicious clients