Federated Deep Learning Reconstruction (DLR)

Unlike standard federated classification, DLR tackles ill-posed inverse problems where the goal is to recover a high-fidelity image from undersampled or noisy sensor data. Each institution trains a local reconstruction network on its own raw acquisition data—k-space in MRI or sinograms in CT—and shares only model gradients. The global model learns a generalized regularization prior from diverse scanner physics without ever accessing the raw sensor measurements.

• Preserves the Nyquist sampling relationship between sensor domain and image domain
• Learns scanner-specific artifact patterns without centralizing proprietary hardware calibration data
• Enables collaborative optimization of compressed sensing and parallel imaging reconstruction pipelines

The defining characteristic of federated DLR is that raw sensor data never leaves the acquiring institution. Unlike federated segmentation or classification—which operate on reconstructed DICOM images—DLR operates directly on the pre-image domain. This is critical because raw k-space or sinogram data contains patient-identifiable information and proprietary scanner fingerprints.

• Raw data remains behind the hospital firewall; only encrypted gradient updates traverse the network
• Eliminates the risk of model inversion attacks that could reconstruct patient anatomy from shared model parameters
• Compliant with HIPAA and GDPR requirements for protected health information

Medical imaging suffers from domain shift across vendors—a Siemens MRI produces different noise characteristics than a GE scanner. Federated DLR addresses this by training on heterogeneous raw data distributions simultaneously. Each site contributes gradients learned from its specific coil sensitivities, gradient non-linearities, and acquisition protocols.

• The global model learns a vendor-agnostic reconstruction prior that generalizes to unseen scanner hardware
• Mitigates the need for costly scanner-specific fine-tuning at each deployment site
• Enables rare-sequence reconstruction (e.g., arterial spin labeling) by pooling knowledge across institutions

Acquiring fully-sampled ground truth data for supervised DLR training is clinically impractical—it requires prohibitively long scan times. Federated DLR leverages self-supervised learning where each site trains on its own undersampled data, using the acquired measurements as the supervision signal. The network learns to predict unacquired k-space lines from acquired ones.

• No need for paired fully-sampled/undersampled datasets at any single institution
• Each site contributes to a shared physics-informed reconstruction prior
• Compatible with Noise2Noise and Noise2Self training paradigms in the federated setting

DLR models operating on raw sensor data are computationally intensive—a 3D MRI reconstruction network can have tens of millions of parameters. Federated DLR employs gradient compression and sparsification techniques to minimize bandwidth overhead during aggregation rounds. Only the most significant gradient updates are transmitted.

• Uses top-k sparsification to transmit only the largest-magnitude gradient elements
• Applies gradient quantization to reduce precision from 32-bit to 8-bit without convergence loss
• Implements local momentum correction to compensate for compressed gradient errors across rounds

Clinical imaging protocols vary dramatically across institutions—different echo times, repetition times, flip angles, and contrast weightings. Federated DLR must reconcile these protocol differences during collaborative training. The architecture employs protocol-conditioning layers that modulate the reconstruction network based on acquisition metadata.

• Injects acquisition parameter embeddings into the network to handle protocol heterogeneity
• Learns a unified reconstruction manifold that spans T1-weighted, T2-weighted, and FLAIR contrasts
• Enables zero-shot reconstruction of protocols never seen during training at any single site

The overarching framework for collaboratively solving inverse problems in medical imaging. This involves learning a mapping from raw acquisition data (e.g., k-space in MRI, sinograms in CT) to diagnostic images without aggregating the sensor data. It generalizes DLR to non-deep-learning methods but is the direct parent category for neural network-based approaches.

A closely related technique focused on removing stochastic noise rather than solving the full inverse problem. While DLR reconstructs from raw sensor data, federated denoising cleans already-reconstructed images. Key applications include:

• Low-Dose CT: Suppressing quantum noise to maintain diagnostic quality at reduced radiation doses.
• Fast MRI: Removing Rician noise from accelerated acquisitions.
• PET: Correcting for low photon counts in short-duration scans.

A specialized collaborative training paradigm for suppressing deterministic image distortions rather than random noise. This includes learning to remove metal artifacts from orthopedic implants, beam hardening streaks in CT, and motion artifacts from patient movement. DLR models often jointly perform reconstruction and artifact reduction in a single end-to-end network.

The process of adapting a global reconstruction model to the specific data distribution of a local hospital's scanner vendor, field strength, or acquisition protocol. Since raw sensor data distributions vary significantly between GE, Siemens, and Philips scanners, domain adaptation prevents performance degradation without requiring centralized access to the target domain's raw data.

The cryptographic backbone enabling secure DLR. Key techniques include:

• Differential Privacy: Adding calibrated noise to gradient updates to prevent membership inference attacks on the raw sensor data.
• Secure Multi-Party Computation (SMPC): Allowing multiple parties to jointly compute aggregation functions without revealing individual model updates.
• Homomorphic Encryption: Performing mathematical operations directly on encrypted gradients, ensuring the central server never sees unencrypted model updates.

A collaborative method for training models to automatically evaluate the diagnostic quality of reconstructed images across sites. This is critical for DLR because reconstruction quality metrics (e.g., SSIM, PSNR, NMSE) often require reference ground-truth images that cannot be centralized. Federated quality assessment enables consistent standards without sharing the reference data.