Deep learning-based sub-pixel level motion estimation and correction in the context of the bSSFP-based radial fMRI

Motivation: Head motion is unavoidable during fMRI and degrades brain imaging. The radial reading of the k-space partially mitigates the issue, but not when sub-pixel motion occurs. Goal(s): Using the neural network to estimate and correct sub-pixel motion in the fMRI study from the radial-kspace. Approach: We used supervised-learning to create a neural network to estimate sub-pixel motion from the radial-kspace. We next corrected the motion artifact using predicted motion parameters and Fourier transform properties. Results: We found that the motion correction decreased the amount of in-plane motion, which is an indication that the suggested technique decreased the amount of sub-pixel motion. Impact: In this study, the translational and rotational motion at the sub-pixel level was estimated and corrected using the kspace. It may result in more accurate detection of neural activity and potentially improve fMRI research.