MRI motion artifact reduction using a conditional diffusion probabilistic model (MAR-CDPM).

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• High-resolution MRI provides crucial diagnostic information but is limited by motion artifacts in long acquisition sequences.
• Motion artifacts compromise the accuracy of post-processing algorithms in MRI.

Purpose of the Study:

• To develop and evaluate a novel retrospective motion correction method, MAR-CDPM (motion artifact reduction using conditional diffusion probabilistic model).
• To remove motion artifacts from multicenter 3D contrast-enhanced T1 MPRAGE brain datasets with various brain tumor types.

Main Methods:

• Utilized two MRI datasets: one with 3D ceT1 MPRAGE and 2D T2-FLAIR from 230 brain tumor patients, and another with 3D T1W from 148 healthy volunteers.
• Generated in silico motion artifacts in k-space and trained a conditional network (Unet backbone) to reverse the diffusion process, creating MAR-CDPM.
• Evaluated MAR-CDPM against supervised Unet, CycleGAN, and Pix2pix models using quantitative metrics (NMSE, SSIM, PSNR, VIF) and qualitative assessment.

Main Results:

• MAR-CDPM qualitatively outperformed other methods in preserving soft-tissue contrast and brain structures, including tumor boundaries.
• MAR-CDPM achieved superior performance in removing in silico motion artifacts, demonstrated by higher PSNR and VIF.
• The model conditioned on time step and T2-FLAIR showed significant improvements in NMSE, MS-SSIM, SSIM, and MS-GMSD for in silico data.

Conclusions:

• MAR-CDPM effectively removes motion artifacts from 3D ceT1 MPRAGE scans.
• This method is particularly advantageous for imaging elderly patients who may experience involuntary movements during lengthy MRI acquisitions.