Water-fat separation and parameter mapping in cardiac MRI via deep learning with a convolutional neural network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Water-fat separation is a crucial postprocessing technique in MRI for fat identification and quantification.
• Advancements have complicated water-fat separation due to complex models and diverse data, lacking a unified framework.

Purpose of the Study:

• To assess the feasibility and performance of deep learning (DL) for MRI water-fat separation and parametric mapping using diverse input data.
• To compare DL-based methods against conventional techniques for accuracy and image quality.

Main Methods:

• Retrospective analysis of 90 cardiac MRI examinations (1200 acquisitions) from normal and myocardial infarction subjects.
• U-Net DL model trained with single/multiecho, complex/magnitude inputs, validated against graph cut methods with R2*, off-resonance correction, and multipeak fat spectrum.
• Quantitative (PDFF, R2*, off-resonance) and subjective observer analyses were performed.

Main Results:

• DL accurately visualized myocardial fat and intramyocardial hemorrhage.
• DL-derived quantitative values (R2*, off-resonance, water/fat signals) strongly correlated with conventional methods (R² ≥ 0.97).
• DL parameter maps demonstrated a 14% higher signal-to-noise ratio compared to conventional methods.

Conclusions:

• Deep learning-based water-fat separation is feasible with a broad range of MRI inputs.
• R2* and off-resonance mapping with DL require multiple echoes and complex images for optimal results.
• DL offers quantitative and subjective outcomes comparable to conventional model-based water-fat separation, with enhanced image quality.