Federated Artifact Reduction

Trains models to suppress artifacts across diverse scanner vendors and field strengths without centralizing data. The global model learns a generalized artifact representation by aggregating updates from sites with Siemens, GE, and Philips scanners.

• Mitigates domain shift caused by proprietary reconstruction kernels
• Learns vendor-agnostic feature representations
• Eliminates the need for a centralized 'universal' phantom dataset

Enables collaborative training on CT scans containing orthopedic implants or dental fillings without sharing the streak-ridden images. Local nodes train on their own patient populations with specific implant types.

• Handles non-IID data distributions of implant materials (titanium, stainless steel, cobalt-chrome)
• Global model learns to inpaint corrupted sinogram regions
• Avoids centralizing images that could reveal patient identity through unique surgical hardware

Corrects cupping artifacts and dark bands caused by polychromatic X-ray beam attenuation. Institutions collaboratively optimize correction algorithms using their own raw projection data.

• Learns a material decomposition function without sharing raw detector measurements
• Adapts to site-specific tube voltages (kVp) and filtration settings
• Reduces the need for proprietary vendor-specific calibration data

Trains robust models to correct respiratory, cardiac, and involuntary patient motion artifacts in MRI and PET. The federated approach captures a wide range of motion patterns from diverse clinical populations.

• Learns from real-world motion-corrupted scans rather than simulated data
• Handles pediatric, geriatric, and non-compliant patient populations
• Preserves privacy while learning from scans with identifiable motion patterns

Collaboratively optimizes neural networks that reconstruct diagnostic images directly from raw k-space or sinogram data. Sites contribute gradient updates without sharing the raw sensor measurements.

• Enables accelerated acquisition protocols by learning undersampling patterns
• Maintains diagnostic quality while reducing scan time
• Protects raw data that could be reverse-engineered into identifiable images

Establishes a unified image quality benchmark across participating institutions without centralizing quality control data. The federated model learns to harmonize signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR).

• Detects site-specific degradation patterns (e.g., aging detector modules)
• Enables automated quality assurance alerts at the edge
• Facilitates multi-site clinical trials with consistent imaging endpoints

The collaborative training of models to remove noise from medical images without centralizing patient data. This technique is critical for low-dose CT and fast-acquisition MRI protocols.

• Learns noise patterns from diverse scanner vendors
• Preserves diagnostic texture while suppressing artifacts
• Enables safer imaging protocols through collaborative optimization

A direct complement to artifact reduction, focusing specifically on stochastic noise rather than structured distortions like metal artifacts.

The process of adapting a global imaging model to a local hospital's specific data distribution without sharing the target domain data. This addresses scanner-induced domain shift.

• Mitigates vendor-specific biases (Siemens vs. GE vs. Philips)
• Adapts to local patient demographics
• Reduces performance degradation from protocol variations

Essential for ensuring artifact reduction models generalize across institutions with heterogeneous imaging equipment.

A technique for learning a common feature space across heterogeneous imaging scanners and protocols in a decentralized manner. Unlike domain adaptation, harmonization aims to standardize image appearance.

• Creates vendor-agnostic representations
• Reduces inter-site variability in radiomic features
• Enables fair comparison of quantitative biomarkers

Harmonization often precedes artifact reduction to normalize input distributions before collaborative training.

The collaborative optimization of inverse problem solvers that map raw sensor data to diagnostic images. This includes learning to reconstruct from under-sampled k-space or sparse-view sinograms.

• Jointly learns reconstruction priors across institutions
• Addresses both artifacts and resolution simultaneously
• Reduces scan time while maintaining diagnostic quality

Artifact reduction is often embedded within the reconstruction objective rather than applied as a post-processing step.

The decentralized training of algorithms to compensate for patient motion artifacts in MRI, PET, or CT scans. Motion remains a leading cause of non-diagnostic image quality.

• Learns from diverse motion patterns across patient populations
• Addresses both rigid and non-rigid motion
• Critical for pediatric and emergency imaging

Motion artifacts represent a distinct artifact category requiring specialized correction strategies beyond metal or beam-hardening reduction.

A collaborative method for training models to automatically evaluate the diagnostic quality of medical scans across sites. This provides objective quality metrics without centralizing quality control data.

• Detects non-diagnostic scans before AI inference
• Standardizes quality thresholds across institutions
• Enables automated rescan decisions

Quality assessment models often serve as gatekeepers, determining whether artifact reduction is necessary before downstream analysis.