Diffeomorphic Motion-Compensated (DMoCo) Cardiac MRI (RF_FR_407)

Background: Cardiac Magnetic Resonance (CMR) has shifted towards motion compensation and 3D volumes. Resolving isotropic 3D MRI in a clinically useful acquisition time requires a high acceleration factor of up to 40x. We propose DMoCo, a novel motion-compensated reconstruction method for ungated and free-breathing free-running acquisitions. This technique models each volume in the time series as a deformation of a static template to reconstruct 3D images of the heart with an isotropic resolution. The deformation at each motion state is expressed as a diffeomorphism. We assume the tissue velocities at each phase to be represented by a low-rank model. The displacement at each motion state is expressed as the integral of the velocity field from the template to the motion state. Because the model is more constrained, it yields improved results than the MoCo-SToRM approach, which independently models the deformation at each motion state. The approach enables the imaging of cardiac function from free-breathing data from six minutes of acquisition time.

Methods:
Data were acquired in four healthy patients. A free-breathing and ungated balanced steady-state free precession (bSSFP) research sequence application was added to the standard CMR workflow, and patients were scanned on a 1.5T MRI scanner using a 32-channel cardiac array. This data acquisition scheme uses a 3D spiral phyllotaxis trajectory. The view order consisted of a superior-inferior (SI) k-space navigator, acquired in every 22 spokes. The temporal resolution of each group of 22 spokes is approximately 70 ms. The number of radial readouts acquired per spoke was 5749, which corresponds to a fixed acquisition time of 5:58 minutes. The main acquisition parameters were: TR/TE of 3.14/1.6 ms, radio frequency excitation angle of 60◦, resolution of 1.5 mm3, field of view (FOV) of 240 mm2, and readout bandwidth of 840 Hz/pixel. Subjects were given a Gadolinium (Gd) contrast agent intravenously. The standard cardiac workflow was then performed, and the respiratory bellows were removed. Next, 15-25 minutes after contrast was administered, the whole-heart bSSFP scan was acquired. The data was then reconstructed using DMoCo and XD GRASP. The results of the image were compared against those of a clinical 2D cine.

Results: The data were successfully reconstructed using DMoCo and XD GRASP.  (2) Shows 2D CINE versus our matching slice extracted from the 3D reconstructions, DMoCo, and XD GRASP.  Overall, our reconstruction method gives better image quality than XD GRASP.  The diastolic image for XD GRASP shows a biased motion in the left atrium two-chamber view.  In addition, XD GRASP fails to caption the full contraction of the right ventricle (RV) (2).  Our method solves for the deformation fields and is motion-compensated.

Conclusion: We have proposed a motion-compensated reconstruction that jointly reconstructs and registers the isotropic whole-heart volumes. This greatly reduces the total protocol time from approximately 25 minutes to six minutes. To our knowledge, this is the first example of applying a deep-learning reconstruction to 3D free-running MRI.