Robust water-fat separation based on deep learning model exploring multi-echo nature of mGRE

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Water-fat separation is crucial for MRI diagnostics.
• Existing methods face challenges with speed, accuracy, and generalization.
• Multi-echo gradient-recalled echo (mGRE) sequences offer rich data for improved separation.

Purpose of the Study:

• To develop a deep learning network for fast and accurate water-fat separation using mGRE data.
• To explore correlations between multiple echoes for enhanced separation.
• To evaluate the network's generalization across different echo times, field inhomogeneities, and imaging regions.

Main Methods:

• A multi-echo bidirectional convolutional residual network (MEBCRN) was designed.
• The network features a module for extracting correlations between consecutive echoes.
• A water-fat separation module with multi-layer feature fusion and residual structure was employed.
• The network was trained on in vivo abdomen images and tested on abdomen, knee, and wrist images.

Main Results:

• The MEBCRN accurately separated water and fat images.
• Quantitative metrics and robustness were superior compared to other deep learning methods.
• The network demonstrated effectiveness across different echo times, field inhomogeneities, and imaging regions.

Conclusions:

• The proposed MEBCRN effectively learns echo correlations for accurate water-fat separation.
• The deep learning approach exhibits generalization capabilities for varying echo times and field inhomogeneities.
• The network, trained on abdomen images, shows applicability to other imaging regions.