Artifact-free fat-water separation in Dixon MRI using deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Data Analysis

Background:

• Chemical-shift encoded MRI (CSE-MRI) is crucial for body composition and metabolic disorder research.
• UK Biobank uses whole-body Dixon MRI for over 100,000 participants, generating large datasets.
• Current fat-water separation methods are slow, prone to artifacts like fat-water swaps, and lead to data loss or bias.

Purpose of the Study:

• To develop a robust method for accurate fat-water separation in large-scale Dixon MRI datasets.
• To address challenges posed by artifacts, particularly fat-water swaps, in quantitative analysis.
• To improve the efficiency and reliability of body composition analysis in population studies.

Main Methods:

• Formulated fat-water separation as a style transfer problem using a conditional generative adversarial network (cGAN).
• Developed a novel loss function for the generator model in the cGAN.
• Evaluated the model's performance using single (in-phase) and dual (in-phase and opposed-phase) input data.

Main Results:

• The cGAN method accurately predicts artifact-free fat and water volumes from Dixon MRI data.
• The model successfully corrects fat-water swaps, a common artifact in MRI.
• Dual input data (in-phase and opposed-phase) yielded superior separation results compared to single input.

Conclusions:

• The proposed AI method offers a faster and more accurate solution for fat-water separation in large MRI datasets like UK Biobank.
• Eliminates the need for manual inspection or discarding of data affected by fat-water swaps.
• Enhances downstream analysis of body composition and metabolic health in population-based studies.