Federated Motion Correction

Trains neural networks to recognize and correct rigid-body and non-rigid motion artifacts across diverse scanner types without pooling raw k-space data. Each institution contributes gradient updates derived from local motion-corrupted/clean image pairs, enabling the global model to learn from heterogeneous motion distributions—including respiratory, cardiac, and involuntary patient movement—that no single site could capture alone.

Operates directly on raw sensor data or reconstructed images without exposing the underlying patient scans. Techniques include:

• Gradient-only sharing: Only model weight updates leave the institution
• Differential privacy guarantees: Formal bounds on information leakage from updates
• Secure aggregation: Encrypted computation of global parameter averages This ensures compliance with HIPAA and GDPR while enabling multi-institutional model improvement.

Addresses the fundamental challenge of domain shift across MRI and PET scanner manufacturers. By training on motion artifacts from Siemens, GE, Philips, and Canon systems simultaneously, the federated model learns vendor-agnostic correction features. This eliminates the need for per-site fine-tuning and produces a single robust model that generalizes to unseen scanner protocols and field strengths.

Clinical motion patterns are inherently non-identically distributed across sites—pediatric hospitals see different movement profiles than geriatric centers. Federated motion correction frameworks incorporate personalized layers and FedProx proximal terms to handle this statistical heterogeneity, preventing the global model from overfitting to dominant motion patterns while preserving site-specific correction capabilities.

Motion correction models operating on volumetric data produce large gradient updates that strain hospital network bandwidth. Federated implementations employ:

• Gradient sparsification: Transmitting only the top-k significant weight updates
• Quantization: Reducing gradient precision to 8-bit or lower
• Federated matched averaging: Aligning permutation-invariant neural network weights before aggregation These techniques reduce communication overhead by 100-1000x without degrading correction quality.

Federated models learn to perform both retrospective correction on already-acquired images and guide prospective correction during live scanning. The decentralized training framework enables the model to learn from paired examples of motion-corrupted scans and their navigator-corrected counterparts, building a unified understanding of how motion distorts k-space trajectories and how to invert those distortions in real-time.

The decentralized process of learning spatial transformations to align multi-modal or longitudinal scans. Motion correction is a specific, temporally-focused subset of registration, targeting intra-scan subject movement rather than inter-scan anatomical alignment.

• Rigid vs. Deformable: Learns both global 6-degree-of-freedom alignments and complex voxel-level warps.
• Key Benefit: Enables cross-institutional atlas building without exposing raw patient geometry.

A collaborative framework for suppressing various image distortions, including metal artifacts, beam hardening, and motion ghosts. Motion correction is a specialized branch of this broader discipline.

• Shared Goal: Recover the true anatomical signal from corrupted measurements.
• Diverse Data: Training across sites exposes the model to a wider variety of artifact patterns, improving robustness to rare scanner malfunctions.

The privacy-preserving development of neural networks that map raw sensor data (k-space or sinograms) directly to diagnostic images. Motion correction is often integrated as a physics-informed prior within the DLR objective function.

• End-to-End: Jointly optimizes reconstruction and motion compensation in a single pipeline.
• Physics-Based: Incorporates the known MRI or PET acquisition physics to constrain the solution space.

The statistical challenge of training on heterogeneous datasets where motion patterns are not uniformly distributed. A pediatric hospital will have drastically different motion profiles than a geriatric clinic.

• Domain Shift: The global model must generalize across fast, erratic motion (children) and slow, periodic motion (adults).
• Personalization: Often requires local fine-tuning to adapt to site-specific patient demographics and scanner protocols.

A collaborative method for training models to automatically score the diagnostic quality of scans. Motion artifact severity is a primary input feature for these quality control systems.

• Feedback Loop: Quality scores can gate whether a scan is sent for radiologist review or flagged for re-acquisition.
• Standardization: Ensures consistent quality thresholds across a distributed network of imaging centers.

The technique for adapting a global motion correction model to a local hospital's specific scanner vendor (Siemens, GE, Philips) and field strength (1.5T vs. 3T) without sharing local data.

• Vendor Neutrality: Mitigates the performance drop caused by proprietary reconstruction kernels.
• Test-Time Adaptation: Allows the model to self-calibrate on a single new scan without retraining the entire federation.