Federated Image Harmonization

Instead of sharing raw pixels, local nodes learn to project their scanner-specific images into a common latent representation. This is achieved by minimizing a distribution divergence metric—such as Maximum Mean Discrepancy (MMD) or CORAL loss—between the local feature embeddings and a global reference distribution. The global model aggregates only the statistical moments of these embeddings, ensuring that a segmentation model trained on this space generalizes across Siemens, GE, and Philips scanners without ever seeing the source DICOM data.

A local domain discriminator network is trained adversarially to distinguish between feature maps generated from the local scanner and a global style reference. Simultaneously, the primary encoder learns to fool the discriminator, effectively stripping scanner-specific "style" from the anatomical "content." Only the gradient updates from the encoder are shared with the global server, preventing the discriminator from ever accessing raw images from other hospitals.

A lightweight protocol that normalizes image intensity distributions without deep generative models. Local nodes compute histogram landmarks (e.g., percentiles of tissue-specific intensities) and share only these aggregate statistics. The server computes a global harmonization curve, which local sites apply via a monotonic intensity transformation. This is particularly effective for MRI bias field correction and CT Hounsfield Unit standardization across sites.

An extension of classic histogram matching using secure multi-party computation (SMPC) or differential privacy. Local nodes encrypt their cumulative distribution functions (CDFs) before transmission. The server performs homomorphic operations to compute a target histogram that represents the union of all sites without decrypting individual site distributions. This prevents membership inference attacks that could reconstruct patient demographics from intensity distributions.

Advanced architectures that harmonize across entirely different modalities (e.g., mapping CT to pseudo-MR or T1 to T2-weighted contrasts) in a federated manner. Using CycleGAN-based federated translation, local generators learn bidirectional mappings constrained by cycle-consistency losses. The discriminators remain local, while the generators share weight updates. This enables multi-modal analysis without requiring any institution to possess all modality pairs.

A dynamic framework where the harmonization function adapts as new scanner models or acquisition protocols are introduced. Using elastic weight consolidation (EWC) or synaptic intelligence, the global model retains previously learned harmonization mappings while integrating new domains. This prevents catastrophic forgetting when a hospital upgrades from a 1.5T to a 3T MRI scanner, ensuring backward compatibility with legacy data distributions.

Managing statistical and system heterogeneity in decentralized clinical datasets where data distributions are not independent and identically distributed. In image harmonization, scanner-induced domain shift is a primary source of non-IIDness. Mitigation strategies include:

• FedProx: Adding a proximal term to local objectives to stabilize convergence under heterogeneity
• SCAFFOLD: Using control variates to correct for client drift during local training
• Federated normalization layers: Maintaining site-specific batch normalization statistics while sharing convolutional weights
• Clustered federated learning: Grouping institutions with similar scanner profiles before harmonization

The collaborative optimization of inverse problem solvers for medical imaging modalities, learning to map sensor data to diagnostic images without aggregating raw acquisition data. This technique is closely related to harmonization as it addresses vendor-specific reconstruction algorithms that introduce systematic biases. Applications include:

• Federated deep learning reconstruction (DLR) for MRI and CT
• Physics-informed neural networks that incorporate scanner geometry into the federated objective
• Cross-vendor generalization: Training on Siemens and GE scanners to produce vendor-agnostic reconstructions
• Federated unrolled optimization: Sharing learned regularizers while keeping raw k-space data local

The collaborative training of neural networks to enhance the spatial resolution of medical images across institutions, enabling high-quality visualization without sharing original high-resolution patient data. Super-resolution serves as a harmonization pre-processing step by normalizing voxel sizes and slice thicknesses across sites. Key considerations:

• Multi-scale architectures: Learning resolution-independent feature representations
• Federated perceptual loss: Using distributed feature extractors to preserve diagnostic texture
• Arbitrary-scale super-resolution: Harmonizing images to a common isotropic resolution regardless of acquisition parameters
• Cross-modality super-resolution: Harmonizing resolution between CT and MRI for multi-modal fusion

A collaborative method for training models to automatically evaluate the diagnostic quality of medical scans across sites, ensuring consistent standards without centralizing quality control data. Quality assessment is a prerequisite for harmonization, identifying scans that require correction before downstream analysis. Components include:

• Federated no-reference quality metrics: Learning to score images without ground-truth references
• Scanner calibration drift detection: Identifying when a local scanner's output distribution shifts over time
• Harmonization quality verification: Validating that harmonized images meet diagnostic standards
• Multi-task quality-harmonization networks: Jointly learning to assess and correct image quality

A collaborative approach to training models that suppress metal artifacts, beam hardening, or other imaging distortions, leveraging diverse scanner data without sharing the artifact-ridden images. Artifact reduction is a critical harmonization sub-task because artifacts are often vendor-specific and confound downstream analysis. Techniques include:

• Federated metal artifact reduction (MAR) for CT scans with implants
• Motion artifact correction using distributed temporal data
• Ring artifact suppression specific to third-generation CT scanners
• Federated GAN-based artifact disentanglement: Separating artifact patterns from anatomical content in a shared latent space