SISMIK for Brain MRI: Deep-Learning-Based Motion Estimation and Model-Based Motion Correction in k-Space.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Magnetic Resonance Imaging (MRI) is susceptible to patient motion, degrading image quality.
• Existing motion correction methods are often limited and not universally applicable.
• In-plane rigid-body motion in 2D Spin-Echo brain scans is a common challenge in clinical practice.

Purpose of the Study:

• To develop a retrospective method for estimating and correcting in-plane rigid-body motion in MRI.
• To leverage deep neural networks for motion parameter estimation directly in k-space.
• To restore degraded MRI images using a model-based approach, avoiding image artifacts.

Main Methods:

• A deep neural network was trained using supervised learning on extensive motion simulations.
• The method estimates motion parameters from k-space data, enabling correction without a motion-free reference.
• A model-based image restoration technique was employed to reconstruct motion-corrected images.

Main Results:

• The proposed method demonstrated good generalization performance on both simulated and in-vivo data.
• Accurate estimation of motion parameters, even at high spatial frequencies, was achieved.
• Qualitative and quantitative evaluations confirmed the effectiveness of the motion estimation and image reconstruction.

Conclusions:

• The developed deep learning approach offers a robust solution for motion correction in 2D Spin-Echo brain MRI.
• The method's ability to work retrospectively and without a reference is a significant advancement.
• The provided Python implementation facilitates further research and application of this technique.