Unpaired MR Motion Artifact Deep Learning Using Outlier-Rejecting Bootstrap Aggregation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Deep learning methods for MRI motion artifact correction show promise but typically require paired artifact-free and artifact-corrupted images.
• Generating paired data is challenging for specific clinical applications like transient severe motion (TSM) in enhanced MRI.

Purpose of the Study:

• To develop a novel unpaired deep learning scheme for MRI motion artifact correction that eliminates the need for matched image pairs.
• To address limitations of existing supervised methods in clinical scenarios with difficult-to-model motion.

Main Methods:

• A novel unpaired deep learning scheme involving k-space random subsampling, neural network reconstruction from downsampled data, and an aggregation step via averaging.
• The method probabilistically removes outliers in k-space and reconstructs high-resolution images, reducing motion artifacts.

Main Results:

• Successful artifact correction for both simulated and real motion, including TSM in Gd-EOB-DTPA-enhanced MR.
• Effective performance on both single and multi-coil data, with and without k-space raw data.
• Outperformed existing state-of-the-art deep learning methods in artifact reduction.

Conclusions:

• The proposed unpaired deep learning method offers a viable solution for MRI motion artifact correction in challenging clinical situations.
• This approach broadens the applicability of deep learning in MRI by removing the constraint of paired training data.
• The method demonstrates robust performance and potential for widespread clinical adoption.