Deep Learning-Based Water-Fat Separation from Dual-Echo Chemical Shift-Encoded Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Conventional water-fat separation methods in MRI are computationally intensive and susceptible to water/fat swaps.
• These limitations hinder efficient and accurate medical image analysis, particularly in pediatric cases.

Purpose of the Study:

• To develop and validate a deep learning-based dual-echo water-fat separation method.
• To address the challenges of long computational times, water/fat swaps, and metal artifacts in MRI.

Main Methods:

• A densely connected hierarchical convolutional network was designed for water-fat separation.
• The model utilized dual-echo MRI images and echo times as input, with projected power method results as references.
• Training and validation involved 68 pediatric dual echo scans (19382 images), with 8-fold cross-validation and out-of-distribution testing on ankle, foot, and arm data.

Main Results:

• Computational time for volumetric datasets was reduced from approximately 10 minutes to under one minute.
• High fidelity was achieved, indicated by a correlation coefficient of 0.9969, l1 error of 0.0381, SSIM of 0.9740, and pSNR of 58.6876.
• Water/fat swaps were mitigated, and metal artifacts were substantially reduced, even without metal implants in the training data. The method also accurately predicted water/fat images from non-contrast-enhanced scans.

Conclusions:

• The proposed deep learning method offers a fast, robust, and accurate solution for dual-echo water-fat separation in MRI.
• The technique demonstrates significant potential for improving clinical workflow efficiency and image quality, with added benefits of artifact reduction.
• This approach shows promise for broader applications, including non-contrast-enhanced imaging and metal artifact compensation.