Uncertainty-aware physics-driven deep learning network for free-breathing liver fat and R₂ * quantification using...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Quantitative MRI

Background:

• Accurate liver proton-density fat fraction (PDFF) and R2* quantification are crucial for diagnosing and monitoring liver diseases.
• Traditional MRI methods for these quantifications can be time-consuming and susceptible to artifacts, especially with free-breathing techniques.
• Deep learning offers potential for accelerated and artifact-reduced image reconstruction and quantification.

Purpose of the Study:

• To develop a deep learning-based method for rapid liver PDFF and R2* quantification using self-gated free-breathing stack-of-radial MRI.
• To incorporate built-in uncertainty estimation into the quantification process.
• To suppress artifacts and improve the accuracy of quantitative maps.

Main Methods:

• Developed an uncertainty-aware physics-driven deep learning network (UP-Net).
• UP-Net incorporates phase augmentation, a generative adversarial network architecture, and an MRI physics loss term.
• Trained and tested on data from 105 subjects, with uncertainty scores calibrated to predict quantification errors.

Main Results:

• UP-Net achieved superior image quality (SSIM >0.87) and accuracy (NRMSE <0.18) compared to compressed sensing (CS).
• Achieved low mean differences for liver PDFF (-0.36%) and R2* (-0.37 s-1) compared to CS + graph-cut (GC) reference methods.
• Demonstrated significantly faster computation (79 ms/slice vs. 3.2 min/slice for CS+GC) and accurate error prediction using uncertainty maps.

Conclusions:

• UP-Net enables rapid and accurate calculation of liver PDFF and R2* maps from free-breathing radial MRI.
• The generated pixel-wise uncertainty maps effectively predict quantification errors.
• This method holds promise for efficient and reliable liver disease assessment.