DeepResp: Deep learning solution for respiration-induced B₀ fluctuation artifacts in multi-slice GRE.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Respiration causes B0 fluctuation, leading to phase errors in k-space and corrupting MRI images.
• Existing artifact correction methods often require sequence modification or additional hardware.

Purpose of the Study:

• To propose and evaluate DeepResp, a deep learning method for reducing respiration-induced artifacts in multi-slice gradient echo (GRE) MRI images.
• To extract respiration-induced phase errors from complex MRI images using deep neural networks.

Main Methods:

• DeepResp utilizes deep neural networks to extract respiration-induced phase errors from complex MRI images.
• The extracted phase errors are applied to k-space data to generate artifact-corrected images.
• Network training involved computer-simulated images generated from artifact-free images and respiration data.

Main Results:

• DeepResp significantly improved image quality in simulated scans, reducing normalized root-mean-square error (NRMSE) and ghost-to-signal ratio (GSR), while increasing structural similarity (SSIM).
• In-vivo evaluations under deep and natural breathing conditions demonstrated substantial artifact reduction, with notable improvements in NRMSE, SSIM, and GSR.
• The method achieved NRMSE reductions from 13.9% to 5.8% and 20.2% to 5.7% GSR in deep breathing scans.

Conclusions:

• DeepResp effectively reduces respiration artifacts in GRE MRI images without requiring sequence modification or extra hardware.
• The deep neural network approach provides interpretable and reliable extraction of respiration-induced phase errors.
• This method holds promise for widespread application in clinical MRI to enhance image quality.