Four-dimensional flow-sensitive MRI of the thoracic aorta: 12- versus 32-channel coil arrays




To evaluate the performance of four-dimensional (4D) flow-sensitive MRI in the thoracic aorta using 12- and 32-channel coils and parallel imaging.

Materials and Methods:

4D flow-sensitive MRI was performed in the thoracic aorta of 11 healthy volunteers at 3 Tesla (T) using different coils and parallel imaging (GRAPPA) accelerations (R): (i) 12-channel coil, R = 2; (ii) 12-channel coil, R = 3; (iii) 32-channel coil, R = 3. The quantitative analysis included SNR, residual velocity divergence and length and curvature of traces (streamlines and pathlines) as used for 3D flow visualization. In addition, semi-quantitative image grading was performed to assess quality of phase-contrast angiography and 3D flow visualization.


Parallel imaging with an acceleration factor R = 3 allowed to save 19.5 ± 5% measurement time compared with R = 2 (14.2 ± 2.4 min). Acquisition using 12 channels with R = 2 and 32 channels with R = 3 produced data with significantly (P < 0.05) higher quality compared with 12 channels and R = 3. There was no significant difference between 12 channels with R = 2 and 32 channels with R = 3 but for the depiction of supra-aortic branches where the 32-channel coil proved superior.


Using 32-channel coils is beneficial for 4D flow-sensitive MRI of the thoracic aorta and can allow for a reduction of total scan time while maintaining overall image quality. J. Magn. Reson. Imaging 2012;35:190-195. © 2011 Wiley Periodicals, Inc.

IN A CONTEXT where functional evaluation of the cardiovascular system is gaining increased importance, four-dimensional (4D) flow-sensitive MRI is getting more widely used (1–8). The technique is characterized by its high-dimensionality (depiction in three dimensions and over time of anatomy and three-directional blood flow velocities) that allows comprehensive blood flow and vessel wall analysis within complete arterial segments. However, the high-dimensionality of 4D-flow-sensitive MRI is limiting the clinical application of the technique by making the image acquisition time-consuming. In this context, recent developments in MRI acceleration techniques such as parallel imaging (9), non-cartesian (10, 11), compressed or sparse sensing are promising. Nevertheless, increased acceleration factors are known to potentially degrade image quality (12). Additionally, the use of a high number of surface coils results in high SNR near to the coil but lower SNR with increased distance from the coil (13, 14). This aspect is further amplified by the geometry or “g”-factor, which depends on the acceleration factor and on the coil sensitivity of the multi-coil array. In this context, the benefits of multi-channel coil arrays and parallel imaging can be limited when imaging structures centered in the body such as the thoracic aorta.

The aim of this study was to evaluate the performance of 4D flow-sensitive MRI in the thoracic aorta with 12- and 32-channel coils using GRAPPA acceleration (9) with acceleration factors R = 2 (12-channel coil) and R = 3 (12- and 32-channel coils).


Acquisition and Preprocessing

The 4D flow-sensitive MRI measurements were performed in the thoracic aorta in 11 healthy subjects (mean age, 28.3 years; range, 24–62 years; all males, bodyweight 53–97 kg). Data were acquired on a 3 Tesla (T) MR system (Magnetom TRIO Tim, Siemens Medical Solutions, Erlangen, Germany) using a standard 12-channel body coil (weight, 0.95 kg) as well as a 32-channel body coil (weight, 1.2 kg, Invivo Corp., Gainesville, FL). All measurements were performed using a RF-spoiled gradient echo sequence with prospective ECG gating and respiratory gating (6). Data were acquired using a thin oblique–sagittal slab covering the thoracic aorta. The sequence made use of parallel imaging based on the GRAPPA technique in one dimension with 24 auto-calibration lines. For every volunteer, three 4D flow acquisitions were performed with different coils and acceleration factors R: (i) 12-channel coil, R = 2, (ii) 12-channel coil, R = 3, (iii) 32-channel coil, R = 3.

Except for the acceleration factor (R), the three 4D flow acquisitions of each volunteer had identical measurement parameters as summarized in Table 1. The two acquisitions with the 12-channel coil were executed consecutively without repositioning. Between the 12- and 32-channel acquisitions the coils needed to be exchanged and, hence, the subjects were repositioned. The acquisition sequences (12-, 32-channel coils and R = 2/3) were randomized to avoid bias due to their succession (e.g., effect of repositioning).

Table 1. 4D Flow-Sensitive MRI Acquisition Parameters
 4D Flow-sensitive MRI
  • a

    One volunteer was imaged with α = 10 °

  • b

    Assuming a navigator efficiency of 50% and a heart-rate of 60 bpm.

Voxel size [mm3]2.68–2.98 × 1.30–1.46 × 2.70–3.00
Volume dimensions72–96× 192 × 20
Temporal resolution [ms]43.2
venc [cm/s]200
TE / TR [ms]2.73–2.80 / 5.4
Bandwidth [Hz/pixel]450
α [°]7a
Number of calibration lines24
Acquisition time R=2 [min]11.1 – 17.8 (mean = 14.2)b
Acquisition time R=3 [min]9.4 – 14.6 (mean = 11.4)b
Navigator efficiency [%]33–78 (mean = 53)

Data preprocessing, quantification, and visualization were performed using in-house tools based on Matlab (MathWorks, USA). Data were corrected for Maxwell terms and eddy currents (second-order correction) (15).

To avoid introducing any observer bias during quantification, all datasets were fully automatically processed. While both acquisitions based on the 12-channel coil had identical field of view (within one subject), the 32-channel coil measurements did not have the exact same field of view center. To avoid introducing bias due to different imaging regions, the 12- and 32-channel data were registered for every subject and only the voxels belonging to both fields of view were used for further processing.

The image background was suppressed by removing the 30% lowest intensity voxels on the temporal average of magnitude data. Vessels area (aorta, supra-aortic branches, and part of the pulmonary trunk) were selected based on the 5% highest velocity voxels at peak-systole (after background suppression). This segmentation based on intensity distribution was motivated by the fact that all datasets had similar spatial coverage (identical within one volunteer) and allowed to avoid subjective observer-based segmentation differences that could influence the results.

The study was approved by the local ethics committee and written informed consent was obtained from all participants.

Quantitative Comparison

Signal-to-Noise Ratio

The signal-to-noise ratio (SNR) was estimated based on the last two end-diastolic timeframes, which were considered as repeated measurements (13):

equation image(1)

where S represents the magnitude image, equation image the 3D spatial coordinates, vessel the area corresponding to vessels (aorta, supra-aortic branches, and pulmonary trunk) and NT the total number of timeframes.

Although the two measurements used for SNR calculation do not correspond to exactly the same instant in the cardiac period, it was assumed that physiological changes appearing at end-diastole are small compared with the effects due to noise. To account for the spatially varying signal of multi-channel coils, only voxels within vessels (aorta, supra-aortic arteries and pulmonary trunk) were considered for the SNR analysis. This is further justified by the fact that these regions correspond to the regions of interest for flow-sensitive MRI and thus directly relate to the velocity-to-noise ratio (16).

Velocity Divergence

It is commonly accepted that blood presents a mostly noncompressible behavior (17). Consequently, the conservation of mass equation for a fluid with constant density requires (18):

equation image(2)

where ∇ is the divergence operator and equation image the local velocity

In practice, velocity divergence measured from flow-sensitive MRI does not fully vanish due to the presence of measurement noise. It has been suggested to use [2] to improve the velocity-to-noise ratio in flow-sensitive MRI (3, 19). Here, the residual divergence of the measured velocity field was used as an estimator of the presence of noise in the velocity data.

The divergence was calculated based on numerical differentiation of the measured three-directional velocity field. The absolute value of the divergence was then averaged over the vessel area and time to derive the residual divergence:

equation image(3)

Streamlines and Pathlines Analysis

The 3D streamlines were calculated at peak-systole within the volume defined by the segmented vessels. The peak-systole was defined based on the peak velocity within the vessel area. Seed points on a regular grid with equidistant spacing within the entire measurement volume were used as initialization points for the streamline calculation.

Pathlines were calculated using a 3D visualization software (Ensight, CEI, Apex, NC). Virtual massless particles were emitted from a plane in the ascending aorta and integrated over time in the measured 4D flow velocity field. The length and cumulated absolute curvature of the resulting traces were then calculated.

The overall length of traces was calculated as the median value between all traces. For each trace, the total curvature was calculated based on the integration of the curvature along the line. The overall curvature was based on the median value over all traces.

Image Grading

Image quality was independently evaluated by four radiologists based on phase-contrast MR angiograms (PCMRA), streamlines and pathlines visualizations; all derived from the 4D flow-sensitive datasets. PCMRA visualizations combine magnitude and velocity information of flow-sensitive images to depict vessel boundaries (20). The PCMRA images were based on a sagittal maximum intensity projection with normalized brightness and contrast. The streamlines and pathlines images were presented as 3D visualization with fixed view angle (as in Fig. 1).

Figure 1.

The 4D flow-sensitive MRI using 12-channel coil & R = 2, 12-channel coil & R = 3 and 32-channel coil & R = 3: magnitude and head-foot velocity (sagittal plane), phase-contrast MR angiography (PCMRA, maximum intensity projection) and pathlines (3D visualization).

All readers were asked to grade on a scale from 1-poor to 4-excellent (half points allowed) the appearance of the PCMRA and streamline images for every dataset and several criteria. For PCMRA, the criteria were: (i) overall contrast of lumen, (ii) prominence of supra-aortic branches, and (iii) presence of background noise/artifacts. For traces (streamlines/pathlines), the criteria were: (i) overall subjective quality, (ii) presence of traces in the supra-aortic branches, and (iii) presence of noisy or aberrant traces. The image order was randomized and the readers were blinded to the imaging parameters and each other's results.

Statistical Analysis

For the quantitative analysis, the distribution of the continuous parameters was assumed to be Gaussian and the statistical significance was evaluated based on two-sided t-tests between scans with different imaging parameters. For image grading, due to the discrete nature of the values, a Mann-Whitney U-test was performed between each imaging setting. The statistical relevance of all image quality parameters was evaluated based on a P < 0.05 significance level. The inter-observer agreement was evaluated by means of the Fleiss' Kappa statistical analysis (21).


The average adaptive respiratory gating scan efficiency was 53% (range, 33–78%) and the average scan times were 14.2 min and 11.4 min for R = 2 and R = 3, respectively. Sample images for one volunteer are depicted in Figure 1 where it is apparent that the spatially variable sensitivity of the 32-channel coil induced much higher signal at the body surface than in the center of the body.

The results of the quantitative analysis for SNR, velocity divergence (an indicator of inconsistencies in the velocity field), and streamlines/pathlines length and curvature are shown in Figure 2. As expected, 4D flow-sensitive MRI with the 12-channel coil & R = 3 resulted in significantly reduced SNR and increased velocity divergence compared with 12 channels & R = 2 and 32 channels & R = 3. The SNR and velocity divergence in images based on the 32-channel coil & R = 3 were similar to the ones of the 12-channel coil & R = 2. Streamlines length using the 12-channel coil & R = 2 and the 32-channel coil & R = 3 showed similar levels and were larger compared with the 12-channel coil and R = 3 (in 8 of 11 volunteers, nonsignificant). Pathlines length presented no difference between modalities. Streamlines and pathlines curvatures using the 12-channel coil & R = 3 were larger compared with 12 channels & R = 2 (nonsignificant) and 32 channels & R = 3 (significant).

Figure 2.

Quantitative analysis: SNR, divergence, and streamlines/pathlines length and curvature using a 12-channel coil and R = 2, a 12-channel coil and R = 3 and a 32-channel coil and R = 3. The error bars are given for ± the standard deviation between volunteers. The horizontal brackets indicate a significant difference (P < 0.05) between two modalities. [Color figure can be viewed in the online issue, which is available at]

Figure 3 provides the results of the image grading of the PCMRA, streamlines and pathlines images. Image grading of the overall lumen contrast from PCMRA was significantly lower for the 12-channel coil & R = 3 compared with the 12-channel coil & R = 2 and the 32-channel coil & R = 3. Average grading of lumen contrast using 12 channels & R = 2 and 32 channels & R = 3 was almost identical. The 32-channel coil & R = 3 produced significantly better depiction of supra-aortic arteries compared with both other modalities. The presence of background noise and artifacts received similar scores for all three modalities. The streamlines and pathlines scores were the lowest for 12 channels & R = 3 in all categories. The traces scores for 32 channels & R = 3 were greater or equal to 12 channels & R = 2 in all categories (nonsignificant). The overall inter-observer agreement was fair as reflected by a Fleiss' Kappa value of 0.22 (P < 0.05).

Figure 3.

Image grading from four radiologists for PCMRA, streamlines and pathlines images (from 1-poor to 4-excellent). The horizontal brackets indicate a significant difference (P < 0.05) between two modalities. [Color figure can be viewed in the online issue, which is available at]


The 4D flow-sensitive MRI using 12-channel and 32-channel coils and one-dimensional GRAPPA with acceleration factors R = 2 and R = 3 was successfully performed in the thoracic aorta of 11 volunteers. Based on systematic and automatic data processing, it was possible to objectively compare the data from different modalities using direct quantification and image grading based on four independent observers. Compared with R = 2, an acceleration factor R = 3 allowed to save 19.5 ± 5% measurement time. Based on an acceleration factor R = 3, changing from 12 to 32 channels, allowed significant gain in image quality: higher SNR, lower residual divergence and reduced streamlines and pathlines curvature (quantitative analysis, Fig. 2) as well as increased lumen contrast, better depiction of supra-aortic branches, higher quality of streamlines and pathlines and reduced noisy traces (image grading, Fig. 3). Generally, using 32 channels and R = 3 produced 4D flow-sensitive data of similar quality compared with using 12 channels and R = 2. There were no significant differences between these two modalities but for the depiction of supra-aortic branches where the 32-channel coil outperformed the 12-channel coil.

The improved depiction of the supra-aortic branches using the 32-channel coil might be due to the spatially variable sensitivity of the 32-channel coil and the smaller distance between the supra-aortic branches and the body coil.

Because the oblique-sagittal slab used for the 4D flow acquisition of the thoracic aorta was very thin (dimension = 20), parallel imaging in the third dimension (slab dimension ∼ number of autocalibrating lines) was not applied. Further improvements could be obtained by extending the acceleration in additional dimension (spatial, temporal, or velocity encoding). These developments in parallel imaging techniques should reinforce the value of multi-channel arrays.

The quantitative SNR estimation was based on the last two timeframes of the cardiac cycle, thus assuming that blood flow did not change much during this time window, which may limit the quality of the SNR estimation. It was assumed that the end-diastolic dynamics impacted the SNR estimation in a similar manner for the three imaging settings compared in this study and hence did not introduce any bias in the comparison.

In addition to the image grading, the quality of streamlines and pathlines was indirectly estimated based on their length and curvature. This assumed that error in the measured velocity field would be more likely to (i) increase the tortuosity of the lines and (ii) results in shorter lines inside the aortic lumen. While intrinsic low-pass filtering due to higher acceleration factors could have had a counter-balancing effect, the comparison R = 3 versus R = 2 for the 12-channel coil shows that this was not the case: visualization of data acquired with R = 3 resulted in shorter streamlines and higher curvature of streamlines.

The gain of using 32-channel coils for imaging of the aorta might be limited by the distance from the aorta to the coil. While this study showed coherent results over a range of bodyweights (57–97 kg), the advantage of multi-channel coil arrays might be reduced with increased bodyweight or with female subjects.

In conclusion, 4D flow-sensitive MRI with a 32-channel coil resulted in improved image quality (for an identical acceleration factor R = 3) or faster scan times with conserved image quality (if increasing R from 2 to 3) compared with a 12-channel coil. Using 32-channel coil arrays is thus beneficial for 4D flow-sensitive imaging of the thoracic aorta. While these results are based on the 1D GRAPPA technique, developments in fast imaging techniques can be expected to further support the use of such coils for 4D flow imaging.


We thank Ms. Lizhen Cao and Mr. Kaiyuan Zhang for their participation in the image quality evaluation.