Cardiac-phase-resolved 4D MR angiography (MRA) is a promising technique for evaluating patients with cardiovascular disorders¹. The reconstructed 4D dataset enables retrospective interrogation of vascular anatomies in arbitrary 2D orientation and in different cardiac phases. Respiratory motion compensation is required in these applications because the time of acquiring the 4D dataset is usually too long for breath-holds. However, the majority of existing respiratory motion compensation methods are based on gating in which only data acquired during certain respiratory state is used in image reconstruction. Therefore, conventional 4D MRA scans usually have relative low scan efficiency of 20%-40% and suffer from extended yet unpredictable scan time. To address this issue, we proposed a technical strategy of respiratory motion correction for 4D MRA acquisition. In the proposed strategy, respiratory motion is modeled as non-rigid deformable motion rather than translational or affine transformation², which is usually invalid across the large field-of-view of 4D MRA. Meanwhile, image registration technique is used to quantitatively estimate the pixel-wise motion vector fields. We expect the proposed method could remove the motion artifacts when k-space data acquired in more than one respiratory state is used for reconstruction and thus potentially improve the scan efficiency of 4D MRA acquisitions.

(1) Data Acquisition and Binning: The data is acquired using ROtating Cartesian K-space (ROCK) method (Fig.1) where kykz views of 3D Cartesian grid were reordered using quasi-spiral pattern with successive arm rotates by golden ratio of 2π. The ROCK pattern allows the acquired data to be retrospectively binned into different respiratory states based on a motion surrogate while the k-space integrity in each bin is maintained, as shown in Fig.1. Each under-sampled k-space bin is reconstructed into 3D images using ESPIRiT³ algorithm. In our example, the self-gating signal derived from k-space centerline was used as the respiratory motion surrogate.

(2) Motion Vector Field Estimation: Quantitative motion field estimation is performed between a chosen "reference bin" and other "correction bins" using 3D non-rigid image registration, generating a voxel-wise 3D motion vector field. Motion vectors are then processed using k-means clustering and represented by fewer approximations for reduced computational cost in the following step as shown in Fig.2. In our example, the open source Elastix toolbox was used for image registration and N=50 was chosen for the motion vector clustering.

(3) K-Space Correction: Linear k-space phase corrections are performed based on each motion vector on the "correction bin", which are then combined with the “reference bin” respectively. This process results in N combined k-space, each will be reconstructed into 3D images candidates using ESPIRiT. The final respiratory motion-free images are generated by pixel-based image fusion of the N motion corrected image candidate, where the selection is made based on the respective estimated motion vector.

(4) Validation: The proposed method was applied on a ferumoxytol-enhanced free-breathing 4D MRA dataset acquired on a 1 year old pediatric patient under general anesthesia with mechanical ventilation. The original scan takes 6 minutes to a matrix size of 500x240x120x6 in 1mm³ isotropic resolution. Only data acquired in the first 4 minutes are included in this study to generate more under-sampling and demonstrate the potential increase in scan efficiency. K-space data was retrospectively binned into 2 respiratory states and the under-sampling rate of each bin was ~12X.

The motion correction algorithm is finished in less than 2 hours for a single cardiac phases using a standard PC. The motion corrected image (Fig.3c) has higher visual SNR than the ones reconstructed from single respiratory state (Fig.3a) because nearly double the data is used in reconstruction. When compared with the images reconstructed by directly combine k-space from 2 respiratory states without any correction (Fig.3b), diaphragm and vascular structures are better defined in the motion corrected images. Also note the vascular structures in the head, chest and abdomen are uniformly sharper than the uncorrected ones, although the direction and magnitude of their respective motion is different. This suggest the proposed technique is capable of correcting non-rigid respiratory motion across large field-of-view.Our preliminary result demonstrated that the proposed method could correct for complex non-rigid respiratory motion in 4D MRA applications. The scan efficiency could potentially be increased since more k-space data is used in the reconstruction.