Zero-Shot Unsupervised Motion Estimation for Motion-Corrected Cardiac T1 Mapping.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Research

Background:

• Cardiac quantitative MRI (qMRI) is vital for diagnosing conditions like myocardial fibrosis.
• Cardiac motion during MRI acquisition necessitates synchronization, leading to prolonged scan times.
• Motion artifacts are a significant challenge in cardiac qMRI, impacting diagnostic accuracy.

Purpose of the Study:

• To develop a novel deep learning-based image registration method for cardiac qMRI.
• To enable non-rigid motion correction for continuously acquired data over multiple cardiac cycles.
• To reduce scan times in cardiac qMRI while maintaining or improving image quality.

Main Methods:

• A zero-shot, U-Net-based deep learning architecture for non-rigid motion estimation.
• Utilizes the physical qMRI signal model and exploits motion smoothness for accurate estimation.
• Robust to undersampling artifacts, enabling motion estimation from sparsely sampled k-space data.

Main Results:

• Achieved a 61.64% improvement in T1 accuracy on numerical simulations.
• Demonstrated a 45.13% improvement in T1 map sharpness on in-vivo data.
• Improved temporal image alignment of motion-corrected dynamics by an average of 11.78%.

Conclusions:

• The method provides accurate non-rigid motion correction for undersampled cardiac qMRI data.
• Enables faster data acquisition by handling continuously acquired data.
• The scan-specific optimization allows easy adaptation to various cardiac qMRI techniques without large training datasets.