Pilot Tone Guided 3D Head Motion Correction Enhanced By Pretrained 2D U-Net Regularization
Abstract
Purpose
This work addresses the problem that, even after MRI motion correction (MOCO), images can still have reduced SNR and residual motion artifacts from undersampled gaps. The proposed solution integrates a pretrained 2D U-Net denoising model into an existing 3D Pilot Tone (PT)–based head motion compensation pipeline to improve image quality.
Methods
A PT-guided 3D rigid-body MOCO framework was used to estimate 6 motion parameters (3 translations and 3 rotations) by searching for linear weights that map PT signals to motion, refined by minimizing gradient entropy of the reconstructed image. Using the converged motion parameters, a motion-compensated reconstruction was generated. Then, the 3D volume was sliced into 2D images and processed with a publicly available pretrained 2D U-Net denoiser. The full reconstruction model could be depicted as: x'=argminx||T(t)F(r)Sx-y||2+λ||x-xunet||2 where y is the k-space data, T(t) generates the phase ramps from the translation, F(r) is the inverse Fourier operator with coordinates rotated by r , and S represents the coil sensitivities. The approach was evaluated on 4 volunteers and 16 patients. Volunteers were instructed to move, while patients were asked to remain still, though some exhibited mild involuntary head motion. Images were acquired on a 3T system using a 3D MPRAGE sequence.
Results
We achieved substantial improvements in image quality across various motion levels from 4 volunteers, including cases with large motion ranges (rotations ≤ ±23° and translations ≤ ±20 mm). PT-based MOCO improved image quality over uncorrected images, and adding U-Net regularization further increased SNR and reduced artifacts, producing images close to a motion-free references. In patient data, DL-only reconstruction appeared overly smoothed; PT-based MOCO sharpened anatomy and the combined PT+Unet method further reduced noise and improved overall clarity.
Conclusion
Combining PT-based 3D motion compensation with 2D U-Net denoising improved MRI quality beyond either motion correction alone or DL-only reconstruction.