Fat-water MRI separation using deep complex convolution network.

Area of Science:

• Medical Imaging
• Deep Learning
• Magnetic Resonance Imaging (MRI)

Background:

• Deep complex convolutional networks (DCCNs) process complex-valued MRI signals directly.
• Fat-water separation is crucial for quantitative MRI analysis.
• Current methods often split complex MRI signals into magnitude and phase components.

Purpose of the Study:

• To investigate the performance of DCCNs versus real-valued U-Nets for fat-water separation.
• To compare DCCNs and U-Nets within a physics-informed, subject-specific ad-hoc reconstruction framework.
• To evaluate the accuracy of DCCNs and U-Nets against a reference approach.

Main Methods:

• Utilized a comprehensive dataset of 33 multi-echo MRI scans (abdomen, thigh, knee, phantoms) from the 2012 ISMRM fat-water separation workshop.
• Included five additional multi-echo MRIs from MAFLD patients.
• Employed a physics-informed, subject-specific ad-hoc reconstruction method for fat-water separation.

Main Results:

• DCCNs produced fat-water maps with significantly better normalized RMS error and structural similarity index (SSIM) than real-valued U-Nets.
• DCCNs achieved an average SSIM of 0.847 ± 0.069 for fat maps and 0.861 ± 0.078 for water maps.
• The average liver proton density fat fraction (PDFF) from DCCNs showed a high correlation (R=0.847) with the reference approach, outperforming U-Nets.

Conclusions:

• DCCNs demonstrate superior performance in fat-water separation compared to real-valued U-Nets.
• The direct processing of complex-valued MRI signals by DCCNs leads to improved accuracy in quantitative MRI parameters like PDFF.
• DCCNs represent a promising advancement for quantitative fat-water separation in MRI.