Deep-Reorder: Refined Motion Correction for 3D Brain MRI via Iterative Data Regrouping

Accurate and temporally high-resolution estimation of subject motion is paramount for effective motion-correction in MRI. This work proposes a novel approach for estimating subject motion within a 3D MRI scan with high temporal resolution.

Starting from an incoherent sampling scheme, images are reconstructed from temporally ordered k-space fractions via deep learning, enabling motion estimation via registration between each image. Assuming the bulk of the motion is removed using these initial estimates, the remaining misalignment between k-space fractions is expected to be minimally correlated in time. Based on this insight, an iterative strategy is proposed in which shots are arbitrarily regrouped into new k-space fractions, regardless of their original temporal ordering, for subsequent motion estimation and correction. This effectively increases the temporal resolution of the approach, enabling individual correction of all shots.

When applied to 3D T1w TFE MRI of the brain, the proposed approach "Deep-Reorder" estimated a unique motion state for all shots, producing significantly better estimations of retrospective motion compared to estimation without iterative data regrouping. On prospectively motion-corrupted scans with intentional head movement by volunteers, the proposed approach consistently improved image quality, producing less noisy images with reduced motion artifacts.

In conclusion, the proposed approach, which is free of external tracking devices or navigators, effectively estimated 3D motion and corrected motion artifacts with increased temporal resolution.