Robust water-fat separation for multi-echo gradient-recalled echo sequence using convolutional neural network.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Image Analysis
• Deep Learning in Radiology

Background:

• Bipolar multi-echo gradient-recalled echo (GRE) sequences are susceptible to artifacts.
• Accurate water-fat separation is crucial for quantitative MRI analysis.
• Existing methods may struggle with artifacts inherent in bipolar GRE sequences.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for accurate water and fat signal separation.
• To address artifacts present in bipolar multi-echo GRE sequences.
• To improve the robustness and accuracy of water-fat separation in MRI.

Main Methods:

• Designed and trained a CNN using multi-echo images and artifact-free water-fat separated references.
• Generated artifact-free references using iterative decomposition of water and fat with echo asymmetry and least-squares estimation (IDEAL).
• Employed a data augmentation technique with synthetic field inhomogeneity to improve CNN robustness and prevent overfitting.

Main Results:

• The proposed CNN achieved accurate water-fat separation in vivo.
• The CNN demonstrated generalization across different anatomical regions (knee, head, ankle) despite being trained on knee images.
• Data augmentation effectively prevented overfitting and enhanced robustness to magnetic field inhomogeneities.

Conclusions:

• The developed CNN successfully separates water-fat images from bipolar multi-echo GRE acquisitions, including those with bipolar gradient artifacts.
• The method offers a robust solution for water-fat separation in challenging MRI sequences.
• The approach shows potential for widespread application in quantitative MRI.