Liver fat quantification at 0.55 T enabled by locally low-rank enforced deep learning reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Non-alcoholic fatty liver disease (NAFLD) diagnosis requires accurate MRI proton density fat fraction (MRI-PDFF) assessment.
• Low-field MRI offers accessibility but suffers from low signal-to-noise ratio (SNR), hindering precise fat quantification.
• Novel reconstruction techniques are crucial to overcome low-field MRI limitations.

Purpose of the Study:

• To develop and validate a locally low-rank deep learning-based (LLR-DL) reconstruction method.
• To enhance SNR and enable precise MRI-PDFF quantification at low-field (0.55T) MRI.
• To improve consistency of low-field MRI results with higher-field (1.5T) assessments.

Main Methods:

• A novel LLR-DL reconstruction was implemented, alternating between regularized SENSE and a U-Net neural network operating on complex-valued data.
• The U-Net processed spectral projections on local patches across echoes, using the output as a prior for subsequent iterations.
• The final echoes were processed using a multi-echo Dixon algorithm, with imaging performed at 0.55T on phantoms and volunteers.

Main Results:

• LLR-DL significantly improved image quality, increasing peak SNR by 32.7% and structural similarity index by 25% compared to conventional methods.
• Excellent MRI-PDFF repeatability was achieved: 2.33% in phantoms and 0.79% in vivo.
• Narrow cross-field limits of agreement were observed, below 1.67% in phantoms and 1.75% in vivo.

Conclusions:

• The developed LLR-DL reconstruction effectively enhances SNR at low-field MRI.
• Precise and repeatable MRI-PDFF quantification is achievable at 0.55T using LLR-DL.
• This method shows promise for consistent and reliable fatty liver assessment across different field strengths.