Unbiased and reproducible liver MRI-PDFF estimation using a scan protocol-informed deep learning method.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Quantitative MRI

Background:

• Proton density fat fraction (PDFF) is crucial for assessing hepatic steatosis.
• Current deep learning (DL) methods for PDFF estimation lack robustness across different MRI scanners and echo times (TEs).

Purpose of the Study:

• To develop and validate a precise and robust DL-based method for PDFF estimation from chemical shift encoded (CSE) MR images.
• To ensure the method's reliability across diverse MR scanners and acquisition TEs.

Main Methods:

• A two-stage deep learning framework, Variable echo times neural network (VET-Net), was developed.
• VET-Net estimates nonlinear variables of the CSE-MR signal model and uses a vector with TEs as auxiliary input for PDFF calculation.
• Validation was performed on a multi-site, multi-vendor phantom dataset and a single-site liver CSE-MRI dataset.

Main Results:

• VET-Net achieved high reproducibility coefficients (RDCs) of 1.71% and 1.04% in liver regions across different TEs.
• The method demonstrated a small PDFF bias of -0.55% on a multi-site phantom dataset.
• Excluding the auxiliary TE input negatively impacted reproducibility and bias.

Conclusions:

• VET-Net provides unbiased and precise PDFF estimations, outperforming conventional DL approaches.
• The method is robust across different MR hardware vendors and acquisition TEs.
• VET-Net can be utilized to expand MRI-based liver fat quantification for hepatic steatosis assessment.