Fetal-brain magnetic resonance imaging (MRI) has become an invaluable tool for studying the early development of the brain and can resolve diagnostic ambiguities that may remain after routine ultrasound exams. Unfortunately, high levels of fetal and maternal motion (1) limit fetal MRI to rapid two-dimensional (2D) sequences and frequently introduce dramatic artifacts such as (2) image misorientation relative to the standard sagittal, coronal, axial planes needed for clinical assessment and (3) partial to complete signal loss. These factors lead to the inefficient practice of repeating ~30 s stack-of-slices acquisitions until motion-free images have been obtained. Throughout the session, technologists manually adjust the orientation of scans in response to motion, and about 38% of datasets are typically discarded. Thus, subject motion is the fundamental impediment to reaping the full benefits of MRI for answering clinical and investigational questions in the fetus. The overarching goal of this project is to overcome the challenges posed by motion by exploiting innovations in deep learning, which have enabled image-analysis algorithms with unprecedented speed and reliability. We propose to integrate these into the MRI acquisition pipeline to unlock the potential of fetal MRI. We will develop practical pulse-sequence technology for automated and dynamically motion-corrected fetal neuroimaging without the need for external hardware or calibration. We hypothesize that this will radically improve the quality and success rates of clinical and research studies, while dramatically reducing patient discomfort and cost. We propose as Aim 1 to eradicate (2) the vulnerability of acquisitions to image-brain misorientation with rapid, automated prescription of standard anatomical planes.
In Aim 2, we propose to address (3) motion during the scan with real-time correction of fetal-head motion. An anatomical stack-of-slices acquisition will be interleaved with volumetric navigators. These will be used to measure motion as it happens in the scanner and to adaptively update the slice tilt/position. We propose as Aim 3 to develop a 3D radial sequence and estimate motion between subsets of radial spokes for real-time self-navigation. Adaptively updating the orientation of spokes and selectively re-acquiring corrupted subsets at the end of the scan will enable 3D imaging of the fetal brain (1). Since the applicant has a physics background, the proposed training program at MIT and HMS will focus on deep learning and fetal development/neuroscience during the K99 phase to develop the skills needed for transitioning to independence in the R00 phase. The applicant?s goal is to become a fetal image acquisition and analysis scientist acting as bridge between deep learning, MRI and clinical fetal-imaging applications to shift the boundaries of what is currently possible with state-of-the-art technology. Fulfilling the research aims will promote this, as it will result in a practical framework for automation and motion correction, applicable to a wide variety of fetal neuroimaging sequences.
Subject motion is the fundamental impediment to reaping the full benefits of fetal-brain magnetic resonance imaging, as it frequently produces images with dramatic artifacts. The goal of this project is to exploit innovations in deep learning and integrate them into the acquisition pipeline to overcome the challenges posed by motion in fetal neuroimaging studies. This will be achieved by using fast, automated scan prescription of standard anatomical planes and by adaptively updating the acquisition as motion happens in the scanner, based on sub-second navigator scans interleaved with the imaging sequence.
