Introduction

The evaluation of cardiovascular hemodynamics is crucial for diagnosis and management of patients with valvular anomalies [1], congenital heart disease [2] as well as cardiomyopathies [1]. Time-resolved (CINE) phase-contrast (PC) MRI sequences, in which velocity is encoded in one direction through a 2D plane (2D-1dir) are used in clinical routine, usually during breath-holding, for quantifying blood flow and peak flow velocities [3]. However, placement of the acquisition plane remains challenging, and can lead to the underestimation of peak velocities if misplaced or not orthogonal to the flow of interest. This is a common occurrence in cases involving complex flow and where changes in flow direction occur throughout the cardiac cycle, such as with valvular stenosis, valvular regurgitation, or complex congenital heart disease. These underestimations can be improved by taking into account all flow directions, which is achieved by three-directional encoding of all three principal velocity directions inside a slice of interest (2D-3dir) [4]. Alternatively, 4D flow MRI (3D CINE PC-MRI with three-directional velocity encoding) enables post-hoc time-resolved three-dimensional visualization of blood flow and retrospective quantification at any location in a 3D volume [5]. The usefulness of the technique has been increasingly demonstrated for the assessment of blood flow hemodynamics in diverse cardiovascular territories such as the aorta [6], ventricles [7], atria [8], pulmonary arteries [9], intracranial arteries [10], portal [11] and splanchnic arteries [12].

Several previous works have focused on the comparison of 2D and 4D flow MRI for flow quantification, involving in vitro experiments [13] and in vivo in healthy subjects [13–22] or patients with congenital heart disease [17, 20, 22–24], dilated cardiomyopathy [15], valve stenosis [18], suspected intracardiac shunts [25], and other miscellaneous populations [26]. However, in these latter studies, either 2D-1dir or 2D-3dir measurements were used. In addition, the studied indices were either velocity, peak flow or flow volume. Our aim in this study is to investigate the effect of encoding 1 vs. 3 velocity directions and of acquiring data through a 2D plane vs. a 3D volume, by reporting a systematic and comprehensive comparison of 2D-1dir, 2D-3dir and 4D flow MRI for the quantification of peak systolic velocity, peak flow as well as net flow volume in the ascending (AA) and descending (DA) aorta in healthy subjects. Finally, the impact of eddy current correction, as proposed by Walker et al. [27], was assessed for the estimation of each hemodynamic parameter.

Materials and methods

Population

Sixteen healthy volunteers [age 51 ± 18 (19–78) years; 11 men] with no history of cardiovascular disease were recruited between March and May 2014, with the approval of Northwestern University institutional review board and informed consent. Each subject underwent an MRI exam, which included acquisition of velocity-encoded data in the thoracic aorta using 3 sequences (2D-1dir, 2D-3dir, 4D flow MRI). All studies were performed in the morning between 7:30 and 10am after fasting and the instruction to avoid caffeine. Weight was recorded before the exam and brachial blood pressures monitored in real time and recorded every 10 min during the exam.

MRI acquisition protocol

Aortic PC-MRI data were acquired for each subject on a 1.5T MAGNETOM Aera scanner (Siemens Medical Systems, Erlangen, Germany), without the use of any contrast agent, using the three following techniques:

  1. 1.

    One-directional velocity encoding in the through-plane direction with a 2D CINE PC-MRI sequence (2D-1dir) during breath-holding at end-expiration, in an axial slice placed perpendicular to the mid-AA, at the level of the right pulmonary artery center.

  2. 2.

    Three-directional velocity encoding using a 2D CINE PC-MRI sequence (2D-3dir) during free breathing, at the same location as for 2D-1dir measurements.

  3. 3.

    Three-directional velocity encoding using a 4D flow PC-MRI sequence during free breathing, in a sagittal oblique volume which included the thoracic aorta (number of slices = 24–30).

Acquisitions were not performed in the same order for all subjects in order to avoid potential effects such as a decrease in heart rate and cardiac output with scan time as they lie in the magnet. For all acquisitions, prospective ECG gating with a similar covered fraction (~80 %) of the cardiac cycle, as well as segmented k-space, were used. Acquisition parameters are provided in Table 1. Respiration navigator gating was used for acquisition of 4D flow and 2D-3dir data, while positioning the acceptance window at the end-expiration phase with an efficiency of 80 %, which was previously shown to correct for respiratory motion while providing the best trade-off between scan time and measurement of peak systolic maximal velocity magnitude when compared to other schemes [28]. Of note, our 4D flow protocol is in agreement with the 4D flow consensus statement [29].

Table 1 Acquisition parameters used for each of the 2D-1dir, 2D-3dir PC-MRI and 4D flow sequences
Full size table

MRI data analysis

The data analysis workflow is schematically illustrated in Fig. 1. For each subject, the three acquired datasets (2D-1dir, 2D-3dir, 4D flow MRI) were preprocessed (Fig. 1a) using a tool programmed in Matlab (The Mathworks, USA) which was previously described [30]. Preprocessing included the suppression of noise by masking pixels with a low magnitude signal intensity and also those with a high velocity standard deviation throughout the cardiac cycle [27]. Then, an automated unwrapping filter was applied to correct for velocity aliasing in voxels with a velocity shift above the Venc, as compared to the velocity of spatially and temporally neighboring pixels. Finally, eddy current correction was performed by a least squares plane first-order fit to the velocity images separately for each velocity direction, applied slice by slice to static regions at end-diastole, as identified by a low velocity standard deviation across the cardiac cycle [27], while excluding areas with spatial fold-over [31]. Eddy current correction was achieved by subtracting, for each time frame, separately for each velocity direction, the fitted plane from the original velocity image.

Fig. 1
figure 1

Analysis workflow of 2D one-directional (2D-1dir), three-directional (2D-3dir) and 4D flow phase-contrast (PC) MRI data: preprocessing (a), coregistration of the 2D slice location on 4D flow data (b), as well as 2D segmentation of ascending (AA) and descending (DA) aortic borders (c), used for the estimation of peak through-plane velocity (peak Vzmax), peak velocity magnitude (peak Vmagmax), peak flow (peak Q) and net flow volume (Qnet)

Full size image

The slice location used for 2D-1dir and 2D-3dir measurements was automatically co-registered in the thoracic aortic volume based on DICOM location and orientation information, to extract using EnSight (CEI, Apex, North Carolina, USA), at the same level in the AA and DA and using the same orientation, 2D magnitude and velocity images from 4D flow measurements (Fig. 1b).

The next step consisted of segmenting both the AA and DA lumen borders (Fig. 1c). The segmentation was performed on 2D-1dir and 2D-3dir PC-MRI data using the Segment software version 1.9 R4040 (Medviso AB, Lund, Sweden; http://segment.heiberg.se) [32], which provided an automated time-resolved detection of aortic contours on modulus images, based on the level set method, from a manual initialization of a circular region of interest (ROI) on a single cardiac phase. A custom software [19] was used for segmentation of 2D images extracted from 4D flow data that were lower in contrast than the 2D-1dir or 2D-3dir images, which hindered the use of level set algorithms. The software allowed for manual drawing of ROI and ROI fitting based on cubic B-splines using both the magnitude and velocity images.

AA and DA systolic maximal area was extracted for 2D-1dir, 2D-3dir PC-MRI and 4D flow data. Furthermore, after applying a median filter on boundary pixels to reduce the effect of noise in velocity images, the AA and DA contours were used to estimate, for each cardiac time frame t, the following hemodynamic indices, from each sequence:

  1. 1.

    cross-sectionally averaged over the pixels [Vzavg(t)] and maximal [Vzmax(t)] through-plane velocity;

  2. 2.

    maximal velocity magnitude \(\left[ \text{Vma}{{\text{g}}_{\text{max}}}\left( t \right)=\text{max}\left( \sqrt{\text{Vx}{{\left( t \right)}^{2}}+\text{Vy}{{\left( t \right)}^{2}}+\text{Vz}{{\left( t \right)}^{2}}} \right) \right]\);

  3. 3.

    flow rate [Q(t) = cross-sectional area(t) × Vzavg(t)].

Note that Vzmax and Vmagmax are identical for the 2D-1dir data. Also note that the through-plane Vzmax estimated using the free breathing 2D-3dir data is equivalent to Vzmax that would be obtained using a fourth free breathing through-plane 2D-1dir sequence. The peak systolic maximal through-plane velocity (peak Vzmax, cm/s), peak systolic maximal velocity magnitude (peak Vmagmax, cm/s), peak systolic flow (peak Q, ml/s) and net flow volume (Qnet, ml), as obtained by integrating flow rate throughout systole, were further computed.

To test the effect of eddy current correction on hemodynamic measurements for each sequence, the data analysis workflow was additionally performed without correcting for eddy currents.

Statistical analysis

For global qualitative visualization and comparison, maximal through-plane velocity Vzmax, maximal velocity magnitude Vmagmax and flow rate Q waveforms in the AA and DA provided by each MRI velocity-encoded sequence were interpolated using a 39 ms time step for each healthy volunteer, then averaged over the whole group and reported as mean waveforms with standard deviation bars for each cardiac time frame. Quantitative differences across the three MRI sequences at each cardiac time frame were assessed using one-way Kruskal–Wallis. Normal distribution of aortic indices was tested using the Kolmogorov–Smirnov test. AA and DA peak systolic Vzmax and Vmagmax, peak systolic Q, Qnet volume and systolic area were reported as mean values ± standard deviations. Paired comparison between 2D-1dir and 2D-3dir as well as between 4D flow and 2D-3dir measurements was achieved using Bland–Altman analyses. Mean bias, limits of agreement, along with relative difference expressed in percentage of the mean value, were reported. Linear regression was used to study the relationship between peak Vzmax and Vmagmax as well as peak Q and Qnet, as estimated with and without eddy current correction. Slopes, Pearson correlation coefficients r, as well as biases and limits of agreement were provided. Finally, significant differences between hemodynamic indices obtained using 2D-1dir and 2D-3dir, 4D flow and 2D-3dir, as well as with and without correcting for eddy currents, were assessed using a Wilcoxon signed-rank test. A p value <0.05 was considered as statistically significant. Statistical analyses were performed using Matlab (MathWorks, Natick, MA, USA).

Results

Population description

One female subject (6 % of the subjects) was excluded because of severe aliasing artifacts on PC-MRI data, which could not be fixed using the dedicated software nor manual correction. Mean age of the resulting 15 healthy volunteers was 51 ± 19 [19–78] years; mean weight was 192 ± 34 [130–206] lbs and systolic/diastolic blood pressures were 126 ± 25 [92–167]/76 ± 17 [30–105] mmHg.

Comparison between 2D-1dir, 2D-3dir and 4D flow MRI

Figure 2 illustrates examples of velocity profiles inside the AA and DA ROIs as obtained using 2D-1dir, 2D-3dir PC-MRI and 4D flow during systole in a healthy subject. While systolic velocity profiles look similar between 2D-3dir and 4D flow both in the AA and DA, velocities were lower when using 2D-1dir data at both locations.

Fig. 2
figure 2

Examples of velocity profiles obtained in the AA (a) and DA (b) using 2D-1dir (top row), 2D-3dir (middle row) PC-MRI and 4D flow (bottom row) data throughout the systolic period in a healthy volunteer

Full size image

Figure 3 shows maximal through-plane velocity Vzmax, maximal velocity magnitude Vmagmax and flow rate Q for 2D-1dir, 2D-3dir PC-MRI and 4D flow in the AA and DA over the cardiac cycle averaged over the group of healthy volunteers. While no significant differences were observed for DA Vzmax and Vmagmax as well as for AA and DA Q throughout most of cardiac phases, systolic Vmagmax and Vzmax in the AA were significantly different across the three sequences. Of note, a shorter averaged systolic duration time was observed for 2D-1dir breath-hold data, when compared to 2D-3dir or 4D flow MRI data.

Fig. 3
figure 3

Group-averaged maximal through-plane velocity Vzmax (left), maximal velocity magnitude Vmagmax (middle) and flow rate Q (right) throughout the cardiac cycle as estimated in the AA (a) and DA (b), using 2D-1dir (dashed blue lines), 2D-3dir (dotted grey lines) PC-MRI and 4D flow (solid red lines) data. Standard deviations over the whole group are represented as bars for each cardiac phase. *Indicates p < 0.05 for comparison across the three sequences

Full size image

Results for the quantification of peak Vzmax and Vmagmax, peak Q, Qnet and systolic area in the AA and the DA for all three sequences are summarized in Table 2. We found no significant differences between 4D flow and 2D-3dir indices in the AA nor in the DA. However, both in the AA and the DA, for all parameters, the breath-hold 2D-1dir resulted in significantly reduced measurements when compared to 2D-3dir. These findings were confirmed by Bland–Altman analysis for comparisons between 2D-1dir, 2D-3dir PC-MRI and between 4D flow, 2D-3dir PC-MRI (Figs. 4, 5 respectively). Overall, mean biases, limits of agreement and relative percentage differences obtained for peak Vzmax, calculated when considering only the through-plane component, were lower than those obtained for peak Vmagmax, which takes into account all three velocity components. However, peak Vzmax estimated using 2D-1dir data was still significantly lower than 2D-3dir peak Vzmax (Table 2). Note that, as expected, Vzmax was significantly lower than Vmagmax when compared within the same sequence for the 2D-3dir and 4D flow measurements, both in the AA and DA (p < 0.0005 for all).

Table 2 AA and DA peak Vzmax, Vmagmax and Q, Qnet and systolic area mean values
Full size table
Fig. 4
figure 4

Bland–Altman diagrams for comparison between 2D-1dir and 2D-3dir PC-MRI peak systolic maximal through-plane velocity (peak Vzmax), peak systolic maximal velocity magnitude (peak Vmagmax), peak systolic flow (peak Q) and net flow volume (Qnet), in the AA (a) and DA (b). Solid red line indicates mean bias while dashed red lines indicate limits of agreement. Mean values of the relative difference in percentage of the averaged value are provided

Full size image
Fig. 5
figure 5

Bland–Altman diagrams for comparison between 2D-3dir PC-MRI and 4D flow peak through-plane velocity (peak Vzmax), peak velocity magnitude (peak Vmagmax), peak flow (peak Q) and net flow volume (Qnet), in the AA (a) and DA (b). Solid red line indicates mean bias while dashed red lines indicate limits of agreement. Mean values of the relative difference in percentage of the averaged value are provided

Full size image

Effect of eddy current correction on measurement of hemodynamics indices

No significant differences were observed in peak Vzmax and Vmagmax as well as peak Q and Qnet in neither the AA nor the DA using the three PC-MRI sequences, with or without eddy current correction (p ≥ 0.77). Figure 6 shows the results of linear regression analysis for the comparison of hemodynamic indices with and without eddy current correction. Overall, the obtained correlation coefficients were high (r = 0.97–0.999, p < 0.0001). The lowest correlations (r = 0.98 and r = 0.97 in the AA and DA, respectively, p < 0.0001) were obtained when considering Qnet. Table 3 reports results of the Bland–Altman analysis, separately for each of the three PC-MRI sequences. The lowest biases and narrowest limits of agreement were obtained with the 2D-1dir sequence.

Fig. 6
figure 6

Linear regressions for comparison between peak through-plane velocity (peak Vzmax), peak velocity magnitude (peak Vmagmax), peak flow (peak Q) and net flow volume (Qnet), in the AA (a) and DA (b) estimated with and without performing eddy current correction (ECC), while pooling the three PC-MRI sequences. Slopes and Pearson correlation coefficients (r) are provided

Full size image
Table 3 Bland–Altman analysis for comparison between indices estimated with and without performing eddy current correction
Full size table

Table 4 summarizes findings previously reported in the literature for the comparison of flow measurements between 2D PC-MRI and 4D flow, while focusing on the aorta in healthy subjects. The corresponding findings obtained in our study are summarized at the bottom of the table for comparison.

Table 4 Previous comparisons in the literature between aortic 2D PC-MRI and 4D flow in healthy populations
Full size table

Discussion

This is the first study systematically assessing differences in peak velocity, peak flow and volume estimated in the AA and DA of healthy volunteers, based on three flow-sensitive MRI techniques that are currently used in clinical routine and research settings. Our main findings were that: (1) breath-hold 2D-1dir measurements were significantly lower than the free breathing 2D-3dir and the synthetic free breathing 2D-1dir measurements (i.e. the through-plane ‘Vz’ velocity component extracted from the 2D-3dir data); (2) good agreement was observed for the comparison between the free breathing 2D-3dir and 4D flow sequences; (3) the highest sensitivity to eddy current correction was observed for the integrated flow indices.

Several studies previously focused on the comparison of hemodynamic indices estimated using 4D flow MRI compared to standard 2D velocity-encoded sequences [13–26]. When comparing our findings with those previously reported in the literature, focusing on proximal aorta studies including healthy volunteers (Table 4), we found that, to the best of our knowledge, only one compared 4D flow and 2D-3dir sequences. First, our AA and DA net flow volume mean values obtained using 2D-3dir and 4D flow data are within the ranges reported in this study on 19 healthy volunteers (AA 2D-3dir = 74.4 ± 16.2 ml and 4D flow = 73.9 ± 18.9 ml; DA 2D-3dir = 61.6 ± 13.0 ml and 4D flow = 45.4 ± 11.7 ml) [19]. In further concordance with this previous work [19], we found that the net volumes were lower when calculated from 4D flow than 2D-3dir data, which might be due to the lower 4D flow spatial resolution in one dimension, although the difference was not significant. Finally, the mean absolute bias (2D-3dir − 4D flow Qnet = 8.88 ± 7.02 ml) and median relative difference (2D-3dir − 4D flow Qnet ~12 %) obtained by the authors [19] were close to ours, despite their 2D-3dir temporal resolution twice as high as their 4D flow temporal resolution. Interestingly, in our study, only four volunteers had AA 2D-3dir flow volume and peak lower than 4D flow measurements. This could be explained by the fact that, from what we observed retrospectively on 4D flow data, the 2D slice used for 2D-1dir and 2D-3dir measurements was not placed exactly orthogonal to the AA axis in these subjects (and thus partial volume effects associated with anisotropic voxels common to the 2D acquisitions may play role).

We further compared the two 2D PC-MRI sequences, and found that all 2D-1dir velocity and flow indices were significantly lower than 2D-3dir indices both in the AA and DA. The underestimation of peak velocity using a 2D through-plane velocity-encoded sequence when the acquisition plane is not positioned exactly orthogonal to the direction of flow is widely known [33], although its effect on the flow calculation should be minor. Indeed, in our study, peak through-plane velocity was significantly lower than peak velocity magnitude both estimated using the 2D-3dir data. In addition, 2D-1dir aortic peak velocity was still significantly lower than 2D-3dir peak velocity calculated while taking into account the through-plane component only, albeit to a lesser extent than when compared to 2D-3dir peak velocity magnitude calculated using the three principal components. Since data were acquired at the same location with the same slice orientation, both with prospective ECG gating and with the same range of pixel spacing, slice thickness, temporal resolution and encoding velocity, this discrepancy might be due to the differences in the employed number of segments and, importantly, in physiological conditions between acquisitions, specifically breath-holding vs. free breathing. Indeed, it was previously shown that cardiac output measured in the aorta was significantly lower during breath-holding than during free breathing [34]. This can be related to changes in systolic duration time and more generally heart rate, as previously shown between free breathing and breath-hold MRI acquisitions in patients after acute ST-segment elevation myocardial infarction [35]. In our study, we tried to avoid potential systematic changes in physiological conditions over scan time by acquiring the 3 PC MRI sequences in a different order for all subjects. Furthermore, we acquired breath-hold 2D-1dir data during end-expiration and free breathing data while positioning the respiratory navigator acceptance window at the end-expiration phase, as a recent non-electrocardiographic triggered real-time PC-MRI study further exhibited differences in flow throughout the respiratory cycle and between normal and forced breathing [36]. Of note, mean biases, limits of agreement and relative percentage differences for comparison between 2D-1dir peak through-plane velocity and 2D-3dir velocity magnitude were lower in the DA. This finding suggests that complex flow is more prominent in the AA, in agreement with previous observations [37], and highlights that in-plane velocity components should not be neglected in this region.

When referring to previous literature for the comparison between 2D-1dir and 2D-3dir measurements, we found that most studies compared 4D flow and 2D-1dir, thereby confounding the ability to understand the separate contribution of different number of velocity encoding directions and volumetric coverage. As previously reported (Table 4), we found that flow volume [14–16, 21, 25, 26] and peak flow [14] were lower and that peak velocities were higher [18] when estimated using 4D flow in comparison with 2D measurements. Of note, our net volume measurements were lower than those previously reported both at 3T (2D-1dir = 89 ± 10 ml [18] or 100.6 ± 27.8 ml [16] and 4D flow = 88 ± 10 ml [18] or 95.9 ± 19.1 ml [16] in the AA). This could be due to the differences in the studied populations, spatial resolution and gating of MRI data, but also field strength or method used for eddy current correction. Indeed, it was previously demonstrated that flow indices estimated at 3T were higher than those estimated on a 1.5T scanner [14] that we used in our study.

The highest differences between sequences were obtained when considering the spatially integrated peak flow and especially the both spatially and temporally integrated net flow volume. This might be explained by the fact that these flow indices are very sensitive to the segmentation of aortic borders, which can prove to be challenging in cases of low contrast in the images, as is often found with 4D flow data due to the trade-off between good spatial resolution, large volume coverage and limited scan time. Indeed, a recent study focused on the effects of 4D flow MRI data segmentation methods used to calculate flow [26]. The following strategies were compared: (1) segmentation of the systolic frame only and propagation of the ROI throughout the cardiac cycle, in a single 2D slice; (2) same method 1, but averaged over 6 additional consecutive 2D slices along the aortic axis; (3) segmentation of all phases in a single slice. The authors reported that the best agreement against 2D-1dir measurements was found when using the latter method [26], which we also used in our study.

Finally, we found high correlations and low biases for the comparison of aortic indices calculated using each of the three PC-MRI sequence with and without eddy current correction. We found that ascending aortic 4D flow measurements were the most sensitive to eddy current correction, which was expected since the correction was applied slice by slice and on all three velocity components. Again, up to 10 ml differences can be obtained for net flow volume, whose high sensitivity to correction is known given its double spatially and temporally integrated feature, which confirms the need to correct for eddy currents [38].

The main limitation of our study is the lack of a true gold standard for hemodynamic measurements, which does not exist for in vivo measurements in humans. We could have considered Doppler ultrasound as a reference, although this technique provides only one-directional velocity and is highly operator-dependent. In addition, cine SSFP MRI could have been used as a gold standard for the non-invasive estimation of left ventricular stroke volume, but it was not part of the designed protocol which was focused on velocity-encoded acquisitions. However, the purpose of this study was to provide a head-to-head comparison between velocity-encoded MRI sequences, and the validation of such techniques is beyond the scope of this article. The use of prospective ECG gating could be considered as a second drawback, since it does not provide a full coverage of the cardiac cycle. It was used in our study to enable, during free breathing, the addition of a respiration navigator pulse at end-diastole. Indeed, respiration control is essential for 4D flow MRI to mitigate breathing artifacts that would otherwise cause severe ghosting or blurring, and the combination of navigator gating with retrospective ECG gating is challenging, as it would require to interrupt the data acquisition during late diastole to insert the navigator pulse and thus forward estimation of RR-interval duration. This option was not available in our 4D flow sequence and all data were thus acquired using prospective ECG gating, including for 2D PC-MRI data in order to be consistent with 4D flow acquisition and allow for a more direct comparison between hemodynamic indices to overcome the effect of using either retrospective or prospecting gating. Quantitative comparison of velocity profiles as well as more sophisticated indices such as wall shear stress or vortices obtained using different sequences could also be studied. Another limitation is the use of two different algorithms for 2D PC-MRI and 4D flow image segmentation. However, the aortic systolic area was similar between 2D-3dir and 4D flow data, but different between 2D-3dir and 2D-1dir for which the same segmentation method was applied, probably due to different physiological conditions when acquiring data under free breathing or breath-holding. Finally, it should be underlined that the present study aimed at evaluating the effect of encoding velocity in either 1 or 3 directions and of acquiring data either in a 2D plane or in a 3D volume. To achieve this objective, we tried to match temporal and spatial resolutions between the three MRI techniques. Accordingly, different results might be obtained when using the 2D-1dir PC-MRI sequence used in clinical routine with a better achievable temporal resolution (as low as 10 ms using k-space segmentation factor = 1).

In conclusion, we found a good agreement between the two three-directional velocity-encoded sequences (differences within 10 %). However, aortic flow and velocity indices were significantly lower when measured during breath-hold than during free breathing. Differences obtained using the clinically used 2D-1dir PC-MRI sequence highlight the need to standardize acquisition protocols and to establish hemodynamic reference values specific to a given technique. In particular, care should be taken when reporting the widely used stroke volume.