A network-assisted joint image and motion estimation approach for robust 3D MRI motion correction across severity...

Area of Science:

• Medical Imaging
• Neuroimaging
• Computational Neuroscience

Background:

• Motion artifacts are a significant challenge in 3D Magnetic Resonance Imaging (MRI), potentially compromising image quality and diagnostic accuracy.
• Existing motion correction techniques often involve trade-offs between correction accuracy and computational time.

Purpose of the Study:

• To develop and evaluate a novel method, UNet+JE, for 3D motion correction in MRI using a combination of neural networks and physical modeling.
• To assess the performance of UNet+JE across various levels of motion corruption in both simulated and in vivo data.

Main Methods:

• The UNet+JE method integrates a neural network (UNet_mag) with a physics-informed algorithm for joint estimation of motion parameters and motion-compensated images.
• The method was trained on datasets with varying motion corruption severity and compared against UNet_mag and a benchmark joint estimation (JE) method.
• Performance was evaluated on T1w 3D MPRAGE scans from 40 participants with simulated motion and 10 participants with in vivo motion.

Main Results:

• UNet+JE demonstrated superior motion correction compared to UNet_mag across all metrics for both simulated and in vivo data (p < 10^-2).
• UNet_mag showed residual artifacts and blurring, with higher susceptibility to data distribution shifts than UNet+JE.
• UNet+JE achieved comparable image correction quality to JE (p > 0.05) but significantly reduced runtimes by a median factor of 2.00-3.80 (simulated) and 4.05 (in vivo).

Conclusions:

• UNet+JE effectively combines the robustness of joint estimation with the speed of neural networks for high-quality 3D MRI motion correction.
• The method provides accurate motion compensation across a wide range of corruption levels while significantly decreasing processing time.