Accelerating multi-echo chemical shift encoded water-fat MRI using model-guided deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accelerating chemical shift encoded (CSE) water-fat imaging is crucial for reducing scan times and improving patient comfort.
• Traditional methods for water-fat separation often struggle with artifacts and noise when data is undersampled.
• Deep learning approaches offer potential for enhanced reconstruction but require careful integration with biophysical models.

Purpose of the Study:

• To develop and validate a novel model-guided deep learning water-fat separation (MGDL-WF) framework.
• To accelerate CSE water-fat imaging by enabling reconstruction from undersampled k-space data.
• To improve the accuracy and quality of water and fat images obtained with accelerated acquisition.

Main Methods:

• The MGDL-WF framework combines a multi-peak fat model with a modified residual U-net deep learning network.
• A data consistency layer ensures that reconstructed images adhere to measured k-space data.
• Gauss-Newton iteration is employed for efficient gradient updating within the network.

Main Results:

• MGDL-WF demonstrated significant improvements in peak signal-to-noise ratio (PSNR) compared to compressed sensing (CS-WF) and 2-step methods across various acceleration rates (R=4, 6, 8).
• With Cartesian sampling, PSNR increased by 4.75-6.11 dB at R=8.
• With radial sampling, PSNR increased by 1.68-6.53 dB, and radial sampling outperformed Cartesian sampling at R=8 by 2.07 dB.

Conclusions:

• The MGDL-WF framework effectively leverages water and fat image features from undersampled data for superior accelerated CSE water-fat imaging.
• Radial sampling, when used with MGDL-WF, offers enhanced image quality compared to Cartesian sampling within similar scan times.
• This approach holds promise for faster and higher-quality MRI scans.