Long scan times limit the application of 4D flow in clinical practice. Even at 7T, considerable scan times are needed for modest spatiotemporal resolutions. This work demonstrates the advantage of Compressed Sensing acceleration of 4D flow MRI at 7T with a novel undersampling technique. Healthy subjects (n=5) were scanned using standard SENSE and a proposed undersampling technique with Compressed Sensing reconstruction. Flow analysis showed minor differences, and image quality improved for Compressed Sensing reconstructions with maintained resolution and reduced scan time. The method enables further increases of acceleration and spatiotemporal resolution, adding more physiological details beyond current resolution limitations.
Intracranial 4D flow MRI can quantify hemodynamics in the Circle of Willis (CoW), which may benefit diagnosis, treatment, and screening of vascular pathologies. Moreover, the SNR increase at 7T compared to 3T allows for improved flow visualization and quantification in intracranial arteries1. However, even with higher SNR, 4D flow acquisitions in a clinically acceptable scan time (<20 min) are still limited to a modest spatiotemporal resolution. Ideally, a spatial resolution of 0.5 mm and a temporal resolution between 50-100 ms is needed to accurately capture local hemodynamics (e.g. flow, pulsatility, and wall shear stress)2-4. To go beyond current scan time limitations, advanced acceleration techniques are needed to reduce the scan time.
Therefore, the purpose of this abstract is to investigate the benefit of Compressed Sensing5,6 (CS) accelerated 4D flow MRI at 7T in terms of image quality and acceleration.
All experiments were performed on a 7T MR scanner (Achieva, Philips, Best, The Netherlands) using a dual-channel transmit/32-channel receive head coil (NOVA Medical, Wilmington, MA, USA). The scanner software was modified to continuously sample the k-space in a pseudo spiral Cartesian fashion. Sampling profiles (ky,kz) were updated after every flow encoding segment (4$$$\times$$$TR), and together with a physiological heart rate variability, random undersampling over time could be achieved7. With this sampling strategy, designed to exploit the total variation (TV) in the CS reconstruction, retrospective binning according to trigger time created a unique variable density sampling pattern per cardiac frame (Figure 1). Furthermore, the acquired data set can be reconstructed with different temporal resolutions.
The study population consisted of healthy subjects (n=5, average age (SD) = 32.2 (7.7), 4 males) with no history of cardiovascular disease. Each subject was scanned with the proposed CS undersampling technique with acceleration factor of 8 (scan time 5:09 min), and with a standard SENSE 3.5 scan (7:42 min). A SENSE factor of 3.5 in phase encoding direction was chosen as the highest acceleration still giving acceptable image quality. For both sequences, scan parameters were: FOV 180x180x20 mm3, voxel size 0.5x0.5x0.5 mm3, TE/TR 3.0/7.0 ms, FA 8 degrees, and VENC 150 cm/s.
The acquired data was reconstructed by ReconFrame (Gyrotools, Zurich, Switzerland) and retrospectively binned in defined cardiac frames. CS reconstruction was performed using Berkeley Advanced Reconstruction Toolbox (BART)8. A sparsifying TV transform in time and a wavelet transform in space where used with regularization parameters of rTV = 0.01, rW = 0.01, and 50 iteration steps. All CS data sets were reconstructed into 6, 9, and 12 cardiac frames, resulting in final acceleration factors of 8, 12, and 16 (called CS8, CS12, CS16), respectively. SENSE data had 6 cardiac frames.
For blood flow pathline visualization, the intracranial arteries were segmented in Mimics Medical (Materialise, Leuven, Belgium)(Figure 2) and imported into GTflow (Gyrotools, Zurich, Switzerland). Flow measurements were performed in GTFlow in 4 perpendicular slices in the left and right Internal Carotid Artery as well as Middle Cerebral Artery.
Bland-Altman plots and flow curves were used to compare flow differences between CS8 and SENSE scans (6 cardiac frames), CS8 and CS12 (9 frames), and CS8 and CS16 (12 frames). The CS sets with fewer cardiac frames were linearly interpolated by a shape-preserving piecewise cubic interpolation to match the number of cardiac frames.
A: Sampling distribution after binning. Continuously sampling the k-space in a pseudo spiral Cartesian fashion results in a variable density distribution by design.
B: Sampling scheme illustration of 3 pseudo spirals trajectories with golden-angle increment.
C: Unique sampling masks of the different cardiac frames after binning.