Non-iterative and uncertainty-aware MRI-based liver fat estimation using an unsupervised deep learning method.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Liver proton density fat fraction (PDFF) is a crucial biomarker for various diseases.
• Deep learning (DL) methods for PDFF estimation offer speed but lack interpretability and generalizability.
• Clinical adoption of DL-based PDFF estimation is hindered by these limitations.

Purpose of the Study:

• To introduce an interpretable and generalizable AI-based method for PDFF estimation.
• To develop a technique that quantifies the uncertainty associated with PDFF measurements.
• To overcome the limitations of current DL approaches in clinical practice.

Main Methods:

• An Artificial Intelligence-based Decomposition of water and fat with Echo Asymmetry and Least-squares (AI-DEAL) method was developed.
• AI-DEAL performs one-shot MRI water-fat separation by calculating R2* and off-resonance fields.
• A weighted least squares approach computes water-only/fat-only signals and their covariance matrix for PDFF and uncertainty derivation.

Main Results:

• AI-DEAL demonstrated low PDFF biases (0.25% and -0.12%) in in vivo liver ROIs, outperforming state-of-the-art DL techniques.
• The method showed minimal bias (-3.43% and -0.22%) in fat-water and numerical phantoms, even with added noise.
• Estimated uncertainties correlated well with observed errors and ROI variations, indicating reliability.

Conclusions:

• AI-DEAL provides accurate and reliable PDFF estimation with uncertainty quantification.
• The method shows superior generalizability and interpretability compared to existing DL models.
• AI-DEAL holds significant potential for enhancing the clinical utility of MRI-based liver fat quantification.