Multi-Decoder Neural Network Enables Accurate Liver PDFF Estimation with Fewer Echoes

Overview

This study introduces MDWF-Net, a multi-decoder U-Net architecture that jointly estimates water-only, fat-only images, R2*, and off-resonance maps from 3-echo MRI data. The method achieves liver proton density fat fraction (PDFF) quantification accuracy comparable to the standard 6-echo graph cut technique, enabling significant scan time reduction.

Background

Non-alcoholic fatty liver disease (NAFLD) is closely linked to hepatic fat content, with proton density fat fraction (PDFF) serving as a validated biomarker. Conventional PDFF estimation uses chemical shift-encoded MRI with at least 6 echoes to address confounders like R2* decay and off-resonance effects, but this results in long scan times and multiple breath-holds. Deep learning approaches, particularly convolutional neural networks, have recently been explored to improve water-fat separation and reduce acquisition time, though prior methods have not fully validated performance with fewer echoes or jointly estimated confounding parameters.

Data Highlights

Key Findings

• MDWF-Net jointly estimates water-only, fat-only images, R2*, and off-resonance maps from multi-echo GRE MRI data.
• The network achieves PDFF quantification accuracy comparable to the gold-standard 6-echo graph cut method using only 3 echoes.
• Reduction from 6 to 3 echoes enables scan time reduction from approximately 120 seconds to 54 seconds, decreasing patient breath-hold requirements.
• Multi-task U-Net architecture with multiple decoders allows dedicated estimation of outputs with different characteristics, improving overall accuracy.
• Low flip angle and optimized TR in the 3-echo protocol minimize T1-weighting bias despite shorter acquisition times.

Clinical Implications

The MDWF-Net approach facilitates accurate liver fat quantification with significantly shorter MRI acquisition times, improving patient comfort by reducing breath-hold demands. This method may enhance clinical workflow efficiency and broaden the applicability of PDFF measurement in routine NAFLD assessment without compromising diagnostic accuracy.

Conclusion

MDWF-Net demonstrates that multi-task deep learning can effectively reduce the number of echoes required for accurate liver PDFF estimation, maintaining performance comparable to established methods while substantially shortening scan times. This advancement supports more efficient and patient-friendly liver fat quantification in clinical practice.

References

1. Reeder et al. 2012 -- Proton density fat-fraction: a standardized MR-based biomarker of tissue fat concentration
2. Yu et al. 2011 -- Quantitative chemical shift-encoded MRI for liver fat quantification
3. Graph Cut Algorithm Reference 2014 -- Robust water-fat separation using graph cuts
4. ISMRM Water-Fat Toolbox 2020 -- Open-source implementation of water-fat separation algorithms
5. Multi-task U-Net Architectures 2019 -- Applications in signal processing