Clinically-Feasible Non-Contrast Abdominopelvic MRA using 3D Radial Stack-of-Stars k-space Sampling and Compressed Sensing
Marc D Lindley1,2, Daniel Kim2, Kristi Carlston2, Leif Jensen2, Daniel Sommers2, Ganesh Adluru2, Edward VR DiBella2, Christopher J Hanrahan2, and Vivian S Lee2

1Physics, University of Utah, Salt Lake City, UT, United States, 2Radiology, UCAIR, University of Utah, Salt Lake City, UT, United States


Quadruple inversion-recovery (QIR), non-contrast (NC) MRA was developed as an alternative to contrast-enhanced MRA, and it performance was evaluated in patients with peripheral arterial disease. The scan time for QIR with respiratory gating, however, is on the order of 10-15 minutes. We sought to accelerate QIR using a combination of 3D radial stack-of-stars sampling with tiny golden angles and compressed sensing (CS). This study shows that 5.3-fold accelerated QIR with radial k-space sampling and CS produces images that are comparable to those produced by original QIR (e.g., 2.7-fold acceleration using GRAPPA). In 10 human subjects, normalized signal difference and vessel dimensions were not significantly different between original QIR and 5.3-fold accelerated QIR with radial k-space sampling and CS.


Non-contrast MRA (NC-MRA) is emerging as an alternative test for patients with poor renal function. Compared with contrast-enhanced MRA, NC-MRA has the added advantage of repeat scanning as needed. We have developed the quadruple inversion-recovery (QIR) NC-MRA method to image the aortoiliac arteries using multiple inversion recovery preparations and balanced steady-state of free precession readout (1) and evaluated its performance in patients with peripheral arterial disease (2). The scan time of QIR with respiratory gating, however, is on the order of 10-15 minutes, depending on the patient’s breathing pattern and rate. For clinical translation, we sought to accelerate QIR using 3D radial stack-of-stars k-space sampling with compressed sensing (CS) and compare its performance versus original QIR using parallel imaging.


All imaging was performed on a 3T scanner (Tim Trio, Siemens). To determine limits of acceleration factors (R), we performed a numerical simulation experiment, where a fully sampled data set was retrospectively undersampled according to Figure 1. We tested three different acceleration strategies: (i) radial with R = 5.3, (ii) radial with R = 10.7, and (iii) radial with R = 16. We note that QIR requires a consistent inversion time for each kz plane, so we designed R to be multiples (sans small rounding) of previously published R = 2.7 using parallel imaging. For image reconstruction performed offline in Matlab, for each sampling pattern, coil sensitivity maps were self-calibrated from the densely sampled center of k-space (3). CS reconstruction was performed using non-local means (NLM) (4,5), which is a popular denoising filter that is applicable for CS. Image reconstruction was performed with 50 iterations with normalized NLM weight of 0.7 and normalized fidelity weight of 0.7, where normalization is based on maximum value. These weights were determined empirically based on visual inspection of training data. Based on visual inspection of images shown in Figure 1, prospective acceleration was limited to R=5.3 and 10.6.

For prospective imaging, 6 healthy volunteers (3 male, 3 female, mean age = 39+/-13) and 4 patients (2 male, 2 female, mean age = 63+/-8) with diagnosed aortoiliac disease were imaged after obtaining IRB-approved informed consent, with the original QIR with R = 2.7 and accelerated QIR with R=5.3 and 10.6. Other than the acceleration factors, all QIR NC-MRA pulse sequences used identical parameters: spatial resolution 1.3 x 1.3 x 1.7 mm, TR = 1 respiratory cycle, flip angle = 100°, inversion time = 1600 ms, and respiratory gating. Arterial vessel dimensions were measured at 4 locations as shown in Figure 2 (see arrows). Vessel dimensions were compared using analysis of variance (ANOVA), and with Bonferroni correction to compare each pair with original QIR as control. We also measured normalized signal difference [(vessel-background)/vessel] as a surrogate for contrast-to-noise ratio (CNR) at these locations. Note that true CNR is not easily measurable from GRAPPA and CS reconstructed data.


Mean scan time was 14.6 ± 2.9 minutes, 7.2 ± 1.8 minutes, and 3.7 ± 0.7 minutes for GRAPPA and CS R=5.3 and R=10.6 acquisitions, respectively. Representative maximum intensity projection images are shown in Figure 2. Radial QIR with R=5.3 produced images that are comparable to GRAPPA QIR with R = 2.7, whereas radial QIR with R = 10.6 produced noticeably lower image quality. Normalized signal difference results and vessel dimensions results were not significantly different (p > 0.05) at all levels (see Tables 1 and 2), likely due to low sample size (n=10).


We have demonstrated feasibility of 5.3-fold accelerated QIR using 3D radial stack-of-star with CS, enabling 2-fold additional acceleration compared with previously described GRAPPA QIR with R = 2.7. Further study is warranted to determine optimal CS constraints and diagnostic performance in patients with aortic disease.


This work was supported in part by the following grants:

NIH- 5R01DK063183-11

NIH- 5R01HL116895-02

AHA - 14GRNT18350028


1. Atanasova IP, Kim D, Lim RP, Storey P, Kim S, Guo H, Lee VS. Noncontrast MR angiography for comprehensive assessment of abdominopelvic arteries using quadruple inversion-recovery preconditioning and 3D balanced steady-state free precession imaging. J Magn Reson Imaging 2011;33(6):1430-1439. 2. Atanasova IP, Lim RP, Chandarana H, Storey P, Bruno MT, Kim D, Lee VS. Quadruple inversion-recovery b-SSFP MRA of the abdomen: initial clinical validation. Eur J Radiol 2014;83(9):1612-1619.

3. Walsh DO, Gmitro AF, Marcellin MW. Adaptive reconstruction of phased array MR imagery. Magn Reson Med 2000;43(5):682-690.

4. Tristan-Vega A, Garcia-Perez V, Aja-Fernandez S, Westin CF. Efficient and robust nonlocal means denoising of MR data based on salient features matching. Computer methods and programs in biomedicine 2012;105(2):131-144.

5. Buades A, Coll B, Morel J-M. A review of image denoising algorithms, with a new one. Multiscale Modeling & Simulation 2005;4(2):490-530.


Figure 1: A fully sampled data set was undersampled by different stack-of-stars sampling patterns: (i) R = 5.3, (ii) R = 10.7, and (iii) R = 16. Note that tiny golden angles were used for optimal CS performance. White lines in the sampling patterns correspond to acquired samples

Figure 2: Representative maximum intensity projection images of a patient with prior diagnosis of aortoiliac disease with different acceleration rates, (i) GRAPPA with R=2.7, (ii) radial with R=5.3, and (iii) radial with R=10.6. Note that image quality is degraded for R = 10.6, arrows point to locations where vessel diameters were measured.

Table 1: Normalized signal difference between arterial and background signals, where arterial signal is used as control: (i) GRAPPA R=2.7, (ii) radial R=5.3, and (iii) radial R=10.6.

Table 2: Vessel diameter measurements obtained from three acquisitions: (i) GRAPPA R=2.7, (ii) radial R=5.3, and (iii) radial R=10.6.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)