Motion-robust accelerated MRI using deep learning

Abstract

Restoring motion-corrupted MRI scans to diagnostic quality without requiring a repeat acquisition has the potential to alleviate workflow pressure on MR technologists and improve the patient experience, particularly for vulnerable populations who struggle to remain still during imaging. This thesis investigates intrascan motion estimation by reconstructing and registering fractions of k-space from the MRI scan, called segments. Each segment contains multiple shots, providing enough information for an approximate reconstruction of the anatomy by a neural network. These reconstructions are aligned with groupwise registration to estimate per-segment motion parameters, which enable data-consistent correction of the scan and improve data fidelity and image quality. After initial motion correction, it is proposed to iteratively regroup shots, which may have non-adjacent acquisition times, into new segments for further correction. This increases the temporal resolution of motion estimation and produces higher-fidelity images in scans with severe motion.