k-Space deep learning for reference-free EPI ghost correction.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Nyquist ghost artifacts in echo planar imaging (EPI) arise from phase mismatches between even and odd echoes.
• Conventional correction methods struggle with high-field MRI due to nonlinear magnetic field variations and often require reference scans.
• Recent advancements reframe ghost correction as a k-space interpolation problem solvable with structured low-rank Hankel matrices and deep convolutional neural networks.

Purpose of the Study:

• To develop a k-space deep learning approach for immediate EPI ghost correction without reference scans.
• To address phase mismatches in both accelerated and non-accelerated EPI acquisitions.
• To enhance image quality and reduce processing time in high-field MRI.

Main Methods:

• Utilized directional redundancy in even and odd k-space data by configuring them into two input channels.
• Exploited multi-coil data redundancies by stacking them into additional input channels.
• Trained a k-space ghost correction network to learn interpolation kernels for estimating missing virtual k-space data, including handling subsampling in accelerated EPI.

Main Results:

• Demonstrated superior image quality compared to existing methods on 3T and 7T in vivo data.
• Achieved significantly faster computation times than conventional techniques.
• Validated the method's effectiveness in both accelerated and non-accelerated EPI acquisitions.

Conclusions:

• The proposed k-space deep learning method offers a robust and fast solution for EPI ghost correction.
• It is compatible with accelerated acquisitions, making it a valuable tool for high-field MRI.
• The approach allows for artifact correction without altering existing MRI acquisition protocols.