MotionDPS: Motion-Compensated 3D Brain MRI Reconstruction.

Area of Science:

• Medical Imaging
• Biophysics
• Computational Neuroscience

Background:

• Magnetic Resonance Imaging (MRI) is prone to artifacts from patient motion due to lengthy acquisition times and sequential k-space data collection.
• Motion-induced phase inconsistencies lead to blurring, ghosting, and geometric distortions, degrading diagnostic image quality.
• Retrospective motion compensation is difficult, especially in accelerated MRI, due to the ill-posed nature of joint reconstruction and motion estimation.

Purpose of the Study:

• To develop a unified Bayesian framework for motion-compensated 3D MRI.
• To jointly estimate anatomical images, rigid-body motion parameters, and coil sensitivity maps from motion-corrupted k-space data.
• To enable unsupervised reconstruction without requiring paired motion-free training data.

Main Methods:

• Integration of pretrained 3D complex-valued score-based diffusion models as anatomical image priors within a physics-based forward model.
• Inference via alternating diffusion posterior image updates and proximal optimization for motion and coil sensitivity estimation.
• A fully unsupervised reconstruction approach.

Main Results:

• The proposed framework successfully estimates anatomical images, motion parameters, and coil sensitivity maps directly from motion-corrupted k-space data.
• Experiments on simulated and real-motion brain MRI datasets show improved image quality and motion robustness.
• The method outperforms state-of-the-art classical and learning-based motion correction techniques, especially under severe motion and high acceleration.

Conclusions:

• The unified Bayesian framework offers a robust solution for motion compensation in 3D MRI.
• The integration of diffusion models as priors enhances reconstruction quality and motion robustness.
• This unsupervised approach is particularly valuable for accelerated MRI acquisitions with significant patient motion.