Clinical Scorecard: Estimation of Liver Proton Density Fat Fraction Using a Multi-Decoder Neural Network for Water-Fat Separation with Fewer Echoes

At a Glance

Key Highlights

• NAFLD diagnosis and grading rely on accurate PDFF estimation, traditionally requiring 6-echo MRI acquisitions.
• Graph cut-based water-fat separation is gold standard but computationally intensive and requires longer scan times with multiple breath-holds.
• The proposed multi-decoder U-Net neural network (MDWF-Net) enables accurate PDFF estimation using only 3-echo MRI acquisitions, reducing scan time and breath-hold requirements.

Guideline-Based Recommendations

Diagnosis

• Use chemical shift-encoded MRI to non-invasively estimate liver PDFF for NAFLD assessment.
• Acquire multi-echo gradient echo sequences with at least 6 echoes for standard R2* and off-resonance correction.
• Consider reduced echo acquisitions with advanced neural network methods to shorten scan time.

Management

• Optimize MRI pulse sequence parameters (e.g., TR, flip angle) to minimize T1 bias and maximize slice coverage per breath-hold.
• Apply multi-task deep learning models to jointly estimate water-only, fat-only images, R2*, and off-resonance maps.

Monitoring & Follow-up

• Monitor PDFF changes longitudinally using consistent MRI protocols and validated water-fat separation techniques.
• Validate reduced echo protocols against standard 6-echo methods to ensure accuracy.

Risks

• Long scan times and multiple breath-holds may reduce patient compliance and image quality.
• Shortening repetition time (TR) can introduce T1-weighting bias if flip angles are not appropriately low.

Patient & Prescribing Data

134 volunteers including healthy and fatty-liver subjects; additional healthy subjects scanned prospectively

MDWF-Net enables accurate liver PDFF quantification with 3-echo MRI, reducing scan time from 120 to 54 seconds and breath-hold demands without compromising accuracy compared to 6-echo graph cut method.

Clinical Best Practices

• Exclude confounding liver conditions and significant alcohol use when assessing NAFLD via MRI.
• Use low flip angles during MRI acquisition to reduce T1 bias when shortening TR.
• Employ multi-task neural networks to improve water-fat separation and confounder estimation from fewer echoes.
• Validate new imaging protocols against established reference methods before clinical adoption.

References

• Non-alcoholic fatty liver disease and PDFF validation
• Graph cut water-fat separation method
• Multi-task U-Net architectures for signal processing
• ISMRM water-fat Toolbox