We recruited 11 healthy volunteers (mean age 31.6 ± 11.5 years, 63.6% males) without health issues from the Edinburgh area, regardless of age or sex, of whom 10 had complete MRI data across both visits. One subject did not complete their second scan session (specifically, the 4D flow scan) due to a technical issue.

For the patient subgroup, we recruited 11 patients (mean age 60.0 ± 11.1 years) from the Mild Stroke Study 3 (MSS3), a prospective study of stroke presentations of sporadic SVD that recruited patients with lacunar ischaemic stroke or mild cortical stroke from NHS Lothian clinical stroke services. One of these patients exhibited a lot of movement and was unable to tolerate the scanner following the previous sequences (unrelated to this work) and withdrew before the flow scans took place. We subsequently scanned a new subject so as to reach the desired N of 10.

The MSS3 rationale and design has been previously published (Clancy et al., 2021). The volunteer and MSS3 studies were approved by the South East Scotland Research Ethics Committee (REC 14/HV/001 and 18/SS/0044, respectively). Written informed consent was obtained from all participants.

We scanned each healthy volunteer across two sessions within a single visit to assess repeatability, and each SVD patient in one session to assess 2D-4D differences.

During each visit, we imaged participants on the same 3 T MRI scanner (MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany; syngo MR Numaris 4 VE11C) using a 32-channel phased-array head coil. We also acquired a time-of-flight (TOF) sequence to collect a high-resolution volume covering the major cerebral arteries. We acquired three 2D PC-MRI scans the “gold-standard,” well-established (Shi et al., 2018a; Blair et al., 2020, 2021; Clancy et al., 2021) method of in-vivo, non-invasive blood flow quantification (Figure 1) perpendicular to the (i) first segment of the right MCA, (RMCA) (TR/TE: 10.34/6.17 ms, temporal resolution: 20.68 ms, flip angle (FA): 12°, v_enc: 80 cm/s), (ii) the second segment of the anterior cerebral arteries (ACAs) (TR/TE = 10.44/6.24 ms, temporal resolution: 20.88 ms, FA: 12°, v_enc: 70 cm/s) and (iii) venous sinuses [superior sagittal sinus (SSS) and straight sinus (StS)] (TR/TE: 10.94/6.62 ms, temporal resolution: 21.88 ms, FA: 12°, v_enc: 50 cm/s) with retrospective pulse triggering using a pulse oximeter and reconstruction into 32 timeframes. The venous slice was placed across both veins mentioned, as perpendicular to both as possible when anatomy allowed, otherwise the larger vessel (SSS) took precedent. Table 1 shows the sequence parameters.

During each session, following the 2D scans, subjects also underwent a 4D flow scan using a prototype pulse sequence (WIP 785B) covering the Circle of Willis and major cerebral arteries and veins (volume 224 × 180 × 40 mm³ with oblique-axial orientation, TR/TE: 5.48/2.85 ms, temporal resolution: 87.68 ms, FA: 7°, v_enc: 100 cm/s) with retrospective pulse triggering. Data were reconstructed into 25 timeframes per cardiac cycle. The relatively small FA of 7° was chosen due to the short TR of the sequence and thus the need to prevent saturation and increase signal. We used GRAPPA (GeneRalised Autocalibrating Partial Parallel Acquisition) to speed up data acquisition, with undersampling occurring in the in-plane phase-encoding direction (acceleration factor 3). See Table 1 for more information. Both 2D and 4D scans were performed on each subject per session, with two sessions occurring in the same day (and a break of 5–10 min between) for the healthy volunteer subgroup. We recorded heart rate (bpm) during each session so as to assess any differences that may affect measurement of flow and pulsatility (Wahlin et al., 2013).

These 4D parameters were determined based on sequence optimisation on a previous group of healthy volunteers, wherein scan duration, image quality, and 4D-2D conformity were assessed. These parameters provided the optimal balance of these factors.

We calculated mean flow (ml/s) and an adapted Gosling’s (Gosling and King, 1974) pulsatility index (PI = (flow_max − flow_min)/flow_mean) for each vessel examined. By using the lumen-wide volumetric flow rates to calculate PI, rather than calculating PIs in individual pixels and taking the average of these, we reduced the potential effects of noise and obtained an arguably better representation of overall blood flow pulsatility in each given vessel. With the 4D flow data, we compared the sum of mean flow in the combined ICAs and BAs (“inflow”) with the sum of mean flow in the combined ACAs, MCAs, and PCAs (“cerebral flow”) so as to assess conservation of mass and therefore the precision of the sequence.

To determine intra-rater reliability in the healthy volunteers, we processed a subset of data consisting of an artery (RMCA) and a vein (SSS) a second time (approximately 6 months after the first rating) for each subject using the 2D and 4D scans, as described above.

We assessed test-retest repeatability and intra-rater reliability in the healthy volunteer group using paired sample t-tests, intraclass correlation coefficients (ICC) and Bland-Altman analysis. Koo and Li’s grading system (Koo and Li, 2016) was used to interpret ICC scores: <0.5: poor, 0.5–0.75: moderate, 0.75–0.9: good, >0.9: excellent. For both healthy volunteer and patient groups, we assessed conformity between 2D and 4D results using paired sample t-tests, Bland-Altman analysis, and Pearson r correlation. P-values < 0.05 were considered statistically significant. We subsequently carried out a secondary analysis wherein both groups were combined for greater statistical power. Mean flow and PI measurement differences between healthy and patient groups were assessed using paired sample t-tests.

To assess the effect of vessel size on the test-retest repeatability error, we used Pearson’s r to calculate correlations between lumen cross-sectional area and the absolute mean difference in flow parameters between scan sessions.

Of the subjects who completed the scanning session, nearly all lay still with minimal movement during the flow scans (mean 4D flow scan duration: 15:03 ± 1:47 min for healthy volunteer group and 14:18 ± 1:46 min for patient group) and no discernible movement artefacts or blurring were seen on the associated images. However, one patient displayed noticeable movement during the scan.

For both 4D and 2D scans, flow waveforms were generally visually similar in shape between visits, but the absolute flow values were more susceptible to variation. Recorded heart rates showed and a mean (±SD) difference between visits of 3.6 ± 5.0 bpm, reflecting a percentage difference of 5.3 ± 7.5% relative to the visit 1 measurements (inter-visit correlation r = 0.79). We consider these differences acceptably small.

The test-retest repeatability results for PI and mean flow measurements across both 2D and 4D scans are presented in Tables 2, 3, respectively. The 2D scans showed good PI repeatability (median ICC: 0.765, range 0.345–0.937; Table 2), except for the ACAs (0.345 and 0.571 for right and left ACAs, respectively), and moderate-to-good mean flow repeatability overall (median ICC: 0.711, range 0.674–0.964; Table 3). The 4D scans showed good PI repeatability overall (median ICC: 0.772, range: 0.226–0.912; Table 2), with poor ICCs seen in certain vessels (RICA: 0.432, RPCA: 0.526, and LTS: 0.226), and moderate mean flow repeatability (median ICC: 0.571, range: −0.160 to 0.946; Table 3) (Figure 3).

We found a small, non-significant negative correlation between vessel lumen area and 4D flow-derived PI measurement ICC (r = −0.21, p = 0.485) (Figure 4). Conversely, we found a significant positive correlation between lumen area and 4D flow-derived mean flow measurement ICC (r = 0.70, p = 0.008).

The mean difference between inflow and cerebral flow was −0.57 ± 1.03 mL/s at visit 1 and 0.14 ± 1.13 mL/s at visit 2. A negative difference indicates higher measured cerebral flow than inflow and a positive difference indicates the reverse. These values represent mass conservation errors between 1 and 5% of average inflow, which we feel are acceptably low.

For repeated analysis of the relevant 2D data, RMCA PI and mean flow measurements produced ICC scores of 0.991 (excellent) and 0.975 (excellent), and mean (± SD) differences of 0.01 ± 0.03 (p = 0.136) and 0.01 ± 0.12 mL/s (p = 0.823), respectively. SSS PI and mean flow measurements produced ICC scores of 0.999 (excellent) and 0.986 (excellent), and mean differences of 0.00 ± 0.05 (p = 0.299) and 0.12 ± 0.01 mL/s (p = 0.002), respectively.

For repeated analysis of the 4D flow data, RMCA PI and mean flow measurements produced ICC scores of 0.877 (good) and 0.459 (poor), and mean differences of −0.02 ± 0.08 (p = 0.405) and 0.39 ± 0.70 mL/s (p = 0.094), respectively. SSS PI and mean flow measurements produced ICC scores of 0.906 (excellent) and 0.723 (moderate), mean differences of −0.01 ± 0.06 (p = 0.659) and 0.57 ± 0.84 mL/s (p = 0.049), respectively.

On average, PI measured using 4D flow was of lower magnitude than that obtained by the 2D method for four of the five vessels measured (p < 0.05 for RMCA and RACA; Table 4; Figure 5). Moderate to strong PI correlation was seen in all vessels (r = 0.63–0.89, p < 0.05) except the StS (r = 0.15, p = 0.660). Conversely, the 4D method overestimated mean flow compared with the 2D method for four of the five vessels (p < 0.05 for RMCA, RACA, and SSS). Arterial flow measurements showed poor to moderate correlation between methods (r = 0.22–0.60, p > 0.05) while venous flow measurements showed moderate to excellent correlation (r = 0.69–0.93, p < 0.05).

As seen in the healthy subjects, PI measurements were higher using the 2D method (p < 0.05 for RMCA), with moderate to good correlation between methods (r = 0.64–0.88, p < 0.05). Mean flow measurements were higher using the 4D method (p < 0.05 for SSS), with poor to moderate correlation (r = 0.44–0.57, p > 0.05).

Combining data collected from both healthy and patient groups, we found good PI measurement correlation between methods for the RMCA (r = 0.79) and SSS (r = 0.88), but not the StS (r = 0.36) (Table 4 and Figure 5). As before, arterial PI measurements were higher using the 2D method (p < 0.001) while this was not the case for the SSS (p = 0.571) or StS (p = 0.953). Mean flow measurement correlation between methods was moderate for the RMCA (r = 0.56) and StS (r = 0.67), and excellent for the SSS (r = 0.94). The 4D method measured higher for the RMCA (p = 0.006) and SSS (p < 0.001). We saw no noticeable difference in inter-method variation between healthy and patient groups.

PI measurements were higher in patients with SVD compared to healthy volunteers across all three vessels assessed (Figure 6), with both 2D (RMCA: p = 0.097; SSS: p = 0.157; StS: p = 0.045) and 4D (RMCA: p = 0.038; SSS: p = 0.037; StS: p = 0.521) methods. Meanwhile, mean flow measurements were lower in patients across all vessels, with both 2D (RMCA: p = 0.079; SSS: p < 0.001; StS: p = 0.012) and 4D (RMCA: p = 0.003; SSS: p < 0.001; StS: p = 0.276) methods.

The 2D PC-MRI PI measurements showed mostly good test-retest repeatability, while the mean flow measurements showed mostly moderate repeatability—with smaller vessels (ACAs) showing lower agreement in both cases. As previously reported, 2D PC-MRI flow and pulsatility measurements (at 3 T) have excellent repeatability in vessels at the C2-C3 area (Sakhare et al., 2019) and slightly less (while still acceptable) in the main intracranial arteries (De Verdier and Wikström, 2016), the latter aligning with the findings presented here.

The 4D flow measurements showed acceptable PI repeatability and moderate mean flow repeatability. As the 4D scan allowed for the measurement of more cerebral vessels, we were able to assess considerably more data. These results show the potential of 4D flow to evaluate and compare differences in blood flow across a broad range of vessels in a single scan. While the paired t-tests comparing the PI and mean flow results between scanning sessions visits sometimes produced statistically significant p-values, indicating the mean difference between visits was not zero, the small sample size means we should not rely too heavily on these results. It is promising to see that the intracranial veins demonstrated acceptable repeatability in flow rate and (in most cases) PI measurement. In the case of the transverse sinuses, the right vessel was nearly always dominant and easier to measure, while the left, however, was often small and exhibited fairly flat, low flow rates—leading to poor PI repeatability. This highlights the fact that some vessels are much easier to assess within the 4D volume than others and will vary between subjects. While we found little statistical evidence to suggest that poor PI measurement repeatability is related to vessel lumen size, we did confirm that flow measurements show higher repeatability in larger vessels.

Overall, it seems the 4D method is more reliable for PI measurements than it is for flow measurements. We speculate that this is due to the former being mostly dependent on the shape of the velocity waveforms captured within the segmented pixels (that are often largely similar across pixels), while the latter is also dependent on the number of these segmented pixels and the specificity of their selection. Measurements from several (three, for example) adjacent slices, with the mean flow value calculated across these taken, is likely a sensible approach to reduce this variability (Schrauben et al., 2015).

In contrast to the high repeatability of ICA flow and PI measurements using 2D PC-MRI reported previously, our 4D results showed lower repeatability in the ICAs. This is likely due to the limited coverage of the 4D volume, preventing us from capturing the SSS, circle of Willis and ascending straight sections of the ICAs in one acquisition. Limited to just the tortuous carotid siphon, flow plane placement on the ICAs presented a challenge and likely led to the bilateral ICC differences between these vessels.

With repeated analysis by the same observer, the 2D data processing was very reliable in both arterial and venous vessels. This may reflect that the same image slices were processed each time, perhaps reducing scope for variation. The 4D data proved to be less reliable, as the volume had to be resliced before the masks could be drawn, allowing for more variation in flow measurements. In future experiments, and with more experience, a more standardised and established 4D plane selection protocol, as well as semi-automatic segmentation methods such as intensity-based thresholding (Dunas et al., 2019), may help reduce this variation. These will be important for standardising future clinical research to avoid apparent variation in vessel sampling between studies or centres being attributed to true between-sample differences. As with the test-retest repeatability, the PI measurements showed higher intra-rater reliability than the mean flow measurements.

The multicentre reproducibility, test-retest reliability, and inter-rater dependence of blood flow from 4D flow scans in healthy volunteers have previously been shown to be strong (Wen et al., 2019). The results presented here somewhat conform to these previous findings, although not fully as the authors found higher reproducibility of blood flow measurements than found here—possibly due to the use of advanced k-t MR acceleration and specialised flow analysis software (which allowed for semi-automatic placement of ROIs and likely reduced variability).

The 2D results often showed higher arterial PI measurements than their 4D counterparts. This is likely due to the higher spatial and temporal resolutions of the 2D method, potentially allowing pulsatile flow (including the peak) to be more accurately sampled. The longer acquisition time of the 4D flow compared to a single 2D scan may have led to temporal smoothing if flow waveforms were inconsistent between heartbeats, potentially reducing PI measurement due to the reconstructed data representing an average across heartbeats (Wahlin et al., 2013). Mean flow, however, was often measured higher using the 4D method, likely because the lower spatial resolution can lead to partial volume effects and overestimation of flow (Peng et al., 2015). Unfortunately, increasing 4D flow spatial resolution would increase scan time and decrease SNR. Partial volume correction methods have been suggested previously for 2D data (Tang et al., 1995) and more recently for 4D data (Bouillot et al., 2018), which will be considered in future work.

Although the difference in 2D and 4D measurements was statistically significant in some vessels, the small sample size should again be taken into consideration (as a single large difference in measurements could have strong leverage on the mean) when interpreting these results. Overall, the results shown here are promising in that 4D flow appears to perform relatively close to the standard set by its 2D predecessor when it comes to PI measurement, although more apparent differences exist with mean flow measurements.

Several studies have compared 4D flow with 2D PC-MRI and found strong correlations between the techniques when measuring flow in the internal carotid (Schrauben et al., 2015), basilar (Meckel et al., 2013), and cerebral arteries (Meckel et al., 2013; Wahlin et al., 2013; Dunas et al., 2019). However, this is the first time that this many intracranial arteries and veins have been examined using single 4D flow acquisitions across repeat visits, with comparisons between 2D and 4D methods. Importantly, we showed that there exists good agreement between methods when quantifying flow and pulsatility in intracranial veins rarely examined using these techniques. We also showed relatively small differences and high correlations in PI measurement when assessing older SVD patients. This is a promising finding for the future of SVD research.

The variation in flow measurement differences seen between methods was larger than that of a previous, similar study (Wahlin et al., 2013) in which SDs ranged from 0.21 to 0.43 mL/s across a range of cerebral arteries (compared to SDs ranging from 0.24 to 0.72 mL/s here). This may be due, in part, to their use of a novel, in-house-developed software to segment the vessels, utilising composite complex difference images and region-growing—as opposed to the manual segmentation carried out here that may have introduced more variation due to potential lack of consistency.

We also showed that patients with SVD exhibit lower blood flow and increased pulsatility in both arteries and veins compared to healthy volunteers. Interestingly, these differences were more often statistically significant in the 4D data. These results align with previous literature on SVD, as vessel stiffening and reduced flow are commonly seen in patients with the disease (Shi et al., 2016, 2018b; Wardlaw et al., 2019). However, it is important to note that these subject groups were not age-matched [with age having previously been shown to be positively associated with pulsatility (Lefferts et al., 1985)] and there are other confounding factors, such as blood pressure (Aribisala et al., 2014), at play. Regardless, it is encouraging to see these findings in both the 2D and 4D data.

For the purposes of examining small cerebral vessels, 1 mm isotropic voxels are suboptimal [as even the 2D method’s in-plane resolution of 0.8 × 0.8 mm² has been suggested to be too low for vessels <3 mm in diameter (Fukuyama et al., 2017; Isoda and Fukuyama, 2022)], but the protocol optimisation phase of our study indicated that increasing the isotropic resolution decreased SNR considerably whilst providing little in the way of increased flow measurement conformity to 2D results. 2D PC-MR images, on the other hand, can be acquired with high in-plane resolution and thick slices to maintain high SNR—provided the vessels are reasonably straight.

While the acquisition times of the 4D flow scans were substantially longer than the 2D scans (11–17 min vs. 1–3 min) the greater coverage in a single acquisition minimises the scan setup time and does not require a priori identification of specific vessels, which can be extremely challenging due to anatomical variability, potentially leading to repeated scans.

A limitation of the study is the small sample sizes used (11 healthy volunteers and 10 SVD patients). This meant that the results were more sensitive to outliers, allowing for a single pair of mismatched measurements to notably reduce the repeatability of a vessel. Secondly, the relatively long 4D flow scan times meant that only single-v_enc acquisition could be obtained whilst keeping scan times feasible. To capture all possible blood flow velocities, this was set to 100 cm/s and thus will have reduced the sensitivity to slower velocities in the veins and arterial lumen perimeters. It is likely that some of the lower ICC values are attributable to velocity measurement errors caused by this high v_enc, especially so in smaller vessels. In future studies, v_enc could be set to the vessel type of interest (i.e. arteries, veins, cerebrospinal fluid) for higher SNR. Alternatively, a lower v_enc than used here could be used for all vessel types, with increased reliance on potential phase-unwrapping, or perhaps a multi-v_enc approach (Schnell et al., 2017; Nakaza et al., 2022). Meanwhile, the v_enc chosen for the 2D RMCA scans led to wraparound in some cases (due to being too low) and therefore required correction.

We found that smaller vessels such as the PCAs, and tortuous vessels such as the ICAs, provided the greatest challenge for 4D flow assessment, due to limited spatial resolution and challenging flow plane placement caused in part by limited volume coverage of vessels, respectively. A volume that captured the proximal ICAs (pre-siphon) would likely have led to less challenging ICA analysis and thus more repeatable results. Furthermore, although it would have been beneficial to assess repeatability in the patient group, time limitations and the nature of the SVD study prevented this.

In conclusion, the 4D flow MRI protocol resulted in a clinically feasible scan time (approximately 10–15 min), with an isotropic spatial resolution of 1.0 mm, capable of assessing the major vessels of the brain (and even some smaller branches) with good PI repeatability, intra-rater reliability and accuracy in relation to 2D PC-MRI. In fact, for arterial PI measurements, 4D flow often demonstrated better test-retest repeatability than 2D PC-MRI, suggesting we can be confident in using our 4D flow protocol to measure pulsatility across the major cerebral vessels. Mean flow rate measurements of these intracranial vessels exhibit more test-retest variability in both methods, however. We have shown that the sequence is feasible in the evaluation of intracranial veins thus far rarely examined, which is an important finding. Recent and future methodological advances will likely extend 4D flow’s application to the smaller vessels implicated in neurodegenerative pathologies such as cerebral small vessel disease. Our findings confirm and quantify its reliability, and should therefore aid the design of future clinical studies and trials.