Accelerated motion correction with deep generative diffusion models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Accelerated Magnetic Resonance Imaging (MRI) is crucial for reducing scan times.
• Subject motion during MRI introduces artifacts, degrading image quality and complicating reconstruction.
• Existing motion correction methods often struggle with accelerated data or require external reference signals.

Purpose of the Study:

• To develop a robust method for accelerated MRI image reconstruction.
• To simultaneously correct for subject motion and forward model imperfections.
• To address the ill-posed inverse problem in motion-corrupted MRI data.

Main Methods:

• A Bayesian framework utilizing deep generative diffusion models was employed.
• The method jointly estimates motion-free images and rigid motion parameters.
• Reconstruction is performed on subsampled, motion-corrupted 2D k-space data.

Main Results:

• Successfully reconstructed motion-free images from accelerated 2D Cartesian and non-Cartesian MRI scans.
• Demonstrated effective motion correction without reliance on external reference signals.
• Outperformed existing correction techniques on both simulated and prospectively acquired accelerated data.

Conclusions:

• A flexible framework for retrospective motion correction in accelerated MRI was developed.
• The proposed method leverages deep generative diffusion models for enhanced reconstruction.
• Potential applications extend to correcting other forward model corruptions in MRI.