A physics-informed deep learning model for MRI brain motion correction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Magnetic Resonance Imaging (MRI) is crucial for brain imaging but susceptible to motion artifacts due to long scan times.
• These artifacts can severely degrade image quality, impacting diagnostic accuracy and potentially requiring repeat scans.

Purpose of the Study:

• To develop and evaluate PI-MoCoNet, a novel physics-informed neural network for robust motion artifact correction in high-resolution brain MRI.
• To leverage both spatial and k-space information for artifact removal without explicit motion parameter estimation, enhancing diagnostic reliability.

Main Methods:

• PI-MoCoNet utilizes a dual-network framework: a motion detection network (U-net architecture) and a motion correction network (U-net with Swin Transformer blocks).
• The correction network employs L1 reconstruction loss, LPIPS perceptual loss, and a data consistency loss for k-space fidelity.
• Motion artifacts were simulated, and the method was validated on IXI and MR-ART datasets against baseline models using PSNR, SSIM, and NMSE metrics.

Main Results:

• PI-MoCoNet significantly outperformed baseline methods across all artifact levels on both datasets.
• On the IXI dataset, PI-MoCoNet achieved PSNR improvements from 34.15 dB to 45.95 dB (minor artifacts) and SSIM from 0.87 to 1.00.
• On the MR-ART dataset, PSNR increased from 23.15 dB to 33.01 dB (low artifacts) and SSIM from 0.72 to 0.87.

Conclusions:

• PI-MoCoNet offers a robust, physics-informed solution for mitigating motion artifacts in brain MRI, enhancing image quality and diagnostic reliability.
• The framework's ability to integrate spatial and k-space information makes it clinically applicable, improving patient comfort and reducing the need for repeat scans.