Article Text

Download PDFPDF

Original research
4D-CT angiography versus 3D-rotational angiography as the imaging modality for computational fluid dynamics of cerebral aneurysms
  1. Nicole M Cancelliere1,
  2. Mehdi Najafi2,
  3. Olivier Brina3,
  4. Pierre Bouillot3,4,
  5. Maria I Vargas3,
  6. Karl-Olof Lovblad3,
  7. Timo Krings1,5,
  8. Vitor M Pereira1,5,
  9. David A Steinman2
  1. 1 Department of Medical Imaging, Toronto Western Hospital, Toronto, Ontario, Canada
  2. 2 Department of Mechanical & Industrial Engineering, University of Toronto, Toronto, Ontario, Canada
  3. 3 Department for Diagnostic and Interventional Neuroradiology, Hôpitaux Universitaires de Geneve, Geneva, Switzerland
  4. 4 École Polytechnique Fédérale de Lausanne (EPFL), EPFL Laboratory for Hydraulic Machines, Lausanne, Switzerland
  5. 5 Department of Neurosurgery, Toronto Western Hospital, Toronto, Ontario, Canada
  1. Correspondence to Professor David A Steinman; steinman{at}


Background and purpose Computational fluid dynamics (CFD) can provide valuable information regarding intracranial hemodynamics. Patient-specific models can be segmented from various imaging modalities, which may influence the geometric output and thus hemodynamic results. This study aims to compare CFD results from aneurysm models segmented from three-dimensional rotational angiography (3D-RA) versus novel four-dimensional CT angiography (4D-CTA).

Methods Fourteen patients with 16 cerebral aneurysms underwent novel 4D-CTA followed by 3D-RA. Endoluminal geometries were segmented from each modality using an identical workflow, blinded to the other modality, to produce 28 'original' models. Each was then minimally edited a second time to match length of branches, producing 28 additional 'matched' models. CFD simulations were performed using estimated flow rates for 'original' models (representing real-world experience) and patient-specific flow rates from 4D-CTA for 'matched' models (to control for influence of modality alone).

Results Overall, geometric and hemodynamic results were consistent between models segmented from 3D-RA and 4D-CTA, with correlations improving after matching to control for operator-introduced variability. Despite smaller 4D-CTA parent artery diameters (3.49±0.97 mm vs 3.78±0.92 mm for 3D-RA; p=0.005) and sac volumes (157 (37–750 mm3) vs 173 (53–770 mm3) for 3D-RA; p=0.0002), sac averages of time-averaged wall shear stress (TAWSS), oscillatory shear (OSI), and high frequency fluctuations (measured by spectral power index, SPI) were well correlated between 3D-RA and 4D-CTA 'matched' control models (TAWSS, R2=0.91; OSI, R2=0.79; SPI, R2=0.90).

Conclusions Our study shows that CFD performed using 4D-CTA models produces reliable geometric and hemodynamic information in the intracranial circulation. 4D-CTA may be considered as a follow-up imaging tool for hemodynamic assessment of cerebral aneurysms.

  • Brain
  • Aneurysm
  • Angiography
  • CT Angiography
  • Blood Flow
  • Computational Fluid Dynamics

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.


Aneurysm rupture may produce devastating consequences to patients, including death, in approximately 50% of cases and significant neurological impairment for half of the survivors.1 The prevalence of asymptomatic cerebral aneurysms has been rising in recent years,2 and thus the search for bio- or imaging markers of aneurysm instability is pertinent in order to guide treatment decisions.

Despite the plethora of studies applying computational fluid dynamics (CFD) to demonstrate the role of hemodynamics in aneurysm disease, translation to clinical practice has been hindered due to inconsistent results and lack of reproducibility between groups in a biased cohort of patients.3–7 Among the limitations, one can cite the variability in estimated inlet boundary conditions due to lack of patient-specific flow rates, which may have a significant impact on hemodynamic results.8 Additionally, segmentation methods and imaging modality sources (ie, magnetic resonance angiography (MRA) vs CT angiography (CTA) vs three-dimensional rotational angiography (3D-RA)) influence geometrical output, also influencing CFD results.4 9–11

In an attempt to improve the current CFD workflow, we propose utilization of novel four-dimensional CT angiography (4D-CTA) imaging for aneurysm hemodynamic analysis. 4D-CTA is a time-resolved volumetric CT imaging sequence that produces angiographic images free from bone with good spatial resolution (0.5 mm3 resolution), from which patient-specific cycle average flow rates can be reliably calculated.12 Although the image resolution is inferior to 3D-RA (the current gold standard for CFD; 0.15–0.35 mm3 resolution), 4D-CTA is advantageous as it is non-invasive and provides accurate patient-specific flow rate information that 3D-RA does not.

The primary purpose of this study is to evaluate whether reliable hemodynamic results can be achieved using 4D-CTA imaging by comparing CFD results from patient-specific aneurysm models segmented and simulated from 4D-CTA versus 3D-RA. As geometric differences may influence CFD results, several geometric parameters were also compared and discussed.

Materials and methods

Patient selection

This study included consecutive patients with intracranial aneurysms undergoing endovascular treatment for a brain aneurysm at the Toronto Western Hospital (Toronto, Canada) that required preoperative imaging for device or therapeutic planning. Patients underwent 4D-CTA within 2 weeks before the endovascular procedure and 3D-RA imaging on the day of surgery (see online supplementary figure 1A in online supplementary file 1). Institutional review board approval was obtained for access to anonymized medical records and images (REB#09–0059). As the purpose of this study was only to compare geometric and hemodynamic properties between the two modalities, cerebral aneurysms of all locations and sizes were considered in the study.

Supplemental material


Study design

Patients’ arterial lumen geometries were segmented and edited from each scan by an experienced technologist using an identical standardized workflow (online supplementary figure 1B). Initially, 4D-CTA models were prepared first, blinded to 3D-RA imaging, followed by 3D-RA model preparation 6 months later, blinded to 4D-CTA imaging and models. After segmentation of these 'original' models, each geometry was then minimally edited to match the length of branches between both imaging modalities for each patient (online supplementary figure 1C), creating 'matched' models. For one case (patient 2), outlet number was also matched. Simulations were performed on the 'original' models using an estimated flow rate (representing the real-world experience), and on the 'matched' models using the calculated flow rate from 4D-CTA (representing controlled experiment to evaluate impact of modality alone) (online supplementary figure 1). Analyzed geometric and hemodynamic parameters were chosen based on those which have been shown to be important for aneurysm rupture risk assessment in recent meta-analyses.13 14

Imaging protocols


CTA images were acquired using a dynamic volumetric acquisition protocol on an Aquilion ONE CT scanner (Toshiba Medical Systems, Nasu, Japan) and an intravenous contrast agent injection. This scanner has 320 detector rows providing a volume with a 16 cm long field-of-view (FOV) in a single gantry rotation and a resolution of 0.43×0.43×0.5 mm. In order to determine the contrast timing arrival delay, a test bolus was performed (20 mL at 6 mL/s; Ultravist 370, Bayer; antecubital vein) while simultaneously imaging a single axial slice every second at the level of the C1 vertebrae. Once the contrast agent delay was determined, a NCCT ‘mask’ image was acquired, followed by a second contrast agent bolus injection (50 mL at 6 mL/s) during a volume-mode continuous scan acquisition (0.5 s revolution time, 80 kV, 100 mA during 20 s). This results in 41 volumetric datasets covering fully the contrast agent wash-in and wash-out phases. The volume of interest included as much of the internal carotid arteries (ICAs) as possible while keeping the circle of Willis up to the fourth level of branching divisions. Using the CT console, bone subtraction was performed on each of the time series datasets using the mask image. The subtracted datasets were subsequently anonymized and exported as DICOM files for flow rate analysis. The effective radiation dose for the full 4D-CTA acquisition was 5 mSv.


Catheter-based 3D-RA was performed according to our standard of practice. Thus, 3D imaging was performed at the beginning of the interventional procedure with the patient under general anesthesia, reducing the likelihood of vasospasm and motion artifact. Images were acquired using a calibrated flat-detector biplane DSA system (Allura Clarity, Philips Healthcare, Best, The Netherlands). 3D-RA images were acquired using a 3 mL/s injection of Omnipaque/iohexol (300 mgI/mL) over a 4.07 s acquisition (30 frames/s), with a 3 s delay (24 mL total injection) to allow for sufficient contrast mixing and fully opacified vessels during the entire length of the acquisition (122 frames total).

Acquisitions were reconstructed into 3D volumetric datasets using the DSA system’s Xtravision software package (Philips). To improve spatial resolution, volumes were reconstructed using a 50% size FOV and a 384 mm3 matrix to include the aneurysm, parent artery, and outlet vessels, producing imaging with an isotropic resolution between 0.15 mm3 and 0.35 mm3. A normal reconstruction kernel was used as previous groups have shown that reconstruction of 3D-RAs with a sharp kernel causes high-frequency noise content that results in insufficient quality for accurate segmentation.15

Vascular geometry surface preparation

Both 4D-CTA and 3D-RA DICOM datasets were segmented using a similar segmentation strategy that uses a morphological gradient-based watershed analysis16 and operator-defined topological cues. 4D and 3D models were segmented independently from one another by the same operator, more than 6 months apart, using only each source's imaging so as to not bias the editing or choice of cues.

Briefly, using an interactive MATLAB script, the operator identified the spatial relationship between regions of interest by selecting cues inside and outside the vessels, producing a 3D surface model in stereolithography format. Extensive upstream segments were conservatively retained to minimize the impact of inlet velocity profiles on aneurysm flow patterns,17 and small side branches were kept to ensure proper distribution of flow, especially near the aneurysms. The surface file was then edited using an open source 3D sculpting-based computer-assisted design program (MeshMixer) in order to remove any tiny or unresolved branches and also to perform smoothing to clear artifacts. The inlet and outlet branches were then clipped and circular flow extensions were added using VMTK. Finally, surface re-meshing with edge-length constraint and adaptive loop subdivision was performed using VMTK to improve surface triangulation in regions with steep curvatures and small branches.

In order to control for ‘real-world’ differences between models introduced by blindly editing the first set of models independent from one another, a second set of models was created to match the length of inlet and outlet branches. For example, 3D-RA model inlet arteries were typically shorter than 4D-CTA models due to catheter placement in the parent vessel, therefore 4D models typically had to be clipped shorter. As noted before, these two groups of geometries will be referred to as 'original' and 'matched' geometries, respectively (online supplementary figure 1B,C).

Geometric evaluation

Quantitative geometric features between 3D-RA and 4D-CTA models were calculated using previously described standardized methods.18 Parameters included parent artery diameter, sac volume, ostium area, and maximum neck width.

Hemodynamic simulation

Using VMTK, refined surface models were used to create high-resolution volume meshes with four layers of boundary elements. The open-source, minimally-dissipative, second-order solver Oasis19 was used for CFD analysis. All simulations were performed on Compute Canada’s Niagara HPC platform, each on one 40-core node. The time‐step size was Δt=0.0951 ms (ie, 10 000 steps/cycle for a period of 0.951 s). Three cycles were run to dampen initial transients. Outlet flow divisions were based on our recently introduced ‘splitting’ method20 and imposed as traction-free boundary conditions at all outlets. Blood dynamic viscosity was 0.037 Poise with density 1.06 g/cm3, giving a kinematic viscosity of 0.035 Stokes.

Regarding inflow conditions, the 'original' 4D-CTA and 3D-RA models were simulated using an estimated inflow rate (n=28). Specifically, cycle-averaged mean inlet velocity was assumed to be 27 cm/s for the anterior (ICA inlet) cases and 19 cm/s for the posterior (vertebral inlet) case, under the assumption that flow rate scales with cross-sectional area.21 These resulting cycle-averaged flow rates were used to scale a representative older adult flow waveform22 damped by 10%, and imposed as Womersley velocity profiles at the inlet. Because these inflow rates depended on inlet diameter, they might be different between the 4D-CTA and 3D-RA models for a given patient. The second set of 'matched' models were simulated using the same assumed waveform shape, but now scaled to the corresponding patient-specific cycle-averaged flow rate calculated from 4D-CTA blood velocity measurements.12 In other words, for each patient, the 'matched' 4D-CTA and 3D-RA models had the same inflow rates.

Statistical analysis

Several geometric parameters were tested for significant differences between the 3D-RA and 4D-CTA models. Following D’Agostino and Pearson omnibus normality tests, parameters having normal distributions were reported as mean and SD and differences were tested using a paired t-test. For non-normal parameters, median and IQR were reported and differences were tested using a Wilcoxon signed rank test. Statistical significance was assumed for p<0.05. For all hemodynamic simulations, time-averaged wall shear stress (TAWSS), oscillatory index (OSI), and spectral power index (SPI) were calculated and then averaged over the sac, and regression plots were used to assess the agreement between 4D-CTA and 3D-RA models for each aneurysm. All statistical analyses were performed using Prism 6 (GraphPad Software, San Diego, California, USA).


We screened 15 patients harbouring 17 unruptured aneurysms. One patient with one aneurysm (aneurysm 17) was excluded from the CFD analysis due to incomplete filling of the aneurysm sac in 3D-RA (online supplementary figure 2). A total of 56 pulsatile simulations were performed to assess 16 aneurysms. 'Original' 4D-CTA and 3D-RA surface models are shown in figure 1. We included the following locations: internal carotid artery (ICA) (n=11), middle cerebral artery (MCA) (n=2), anterior communicating artery (Acom) (n=2), and basilar artery (n=1). Note that inlet and outlet lengths vary between models and some 3D-RA models are slightly dilated (eg, aneurysm 11) and have a different angle of proximal ICA segments (eg, aneurysm 2).

Figure 1

'Original' segmentations of 4D-CTA (pink) lumen models overlaid on 3D-RA (grey) models demonstrating the operator-introduced variability of inlet and outlet lengths and clipping sites, but good agreement overall. Aneurysms by location: internal carotid artery (n=11; aneurysms 2, 5–14), middle cerebral artery (n=2; aneurysms 1 and 4), anterior communicating artery (n=2; aneurysms 3 and 15), and basilar artery (n=1; aneurysm 16).

Geometric analysis

Quantitative analysis of various geometric parameters showed minimal differences between 4D-CTA and 3D-RA models (table 1). There was no significant difference between ostium area or maximum neck width between models. Overall, 4D-CTA lumens appeared slightly but significantly narrower, as reflected by the 8% smaller parent artery diameters and, although the differences were minimal overall, sac volumes were on average 2% smaller. Larger differences were observed in smaller aneurysms, with the most extreme difference in sac volume being observed in a tiny distal aneurysm, aneurysm 4, a 2.5 mm MCA bifurcation aneurysm (4D-CTA: 3.45 mm3 vs 3D-RA: 6.47 mm3), and the largest difference in neck maximum width was aneurysm 3, a 4 mm Acom aneurysm (4D-CTA: 5.31 mm vs 3D-RA: 3.67 mm).

Table 1

4D-CTA vs 3D-RA geometric parameters (mean±SD or median (IQR))

Hemodynamic analysis

The correlation between 'original' 4D-CTA and 3D-RA models was moderate for TAWSS (R2=0.59) and OSI (R2=0.64) and poor for SPI (R2=0.13). However, after matching the inflow rates and length of branches, the correlation between TAWSS, OSI, and SPI improved dramatically (TAWSS: R2=0.91; OSI: R2=0.79; and SPI: R2=0.90). Distribution of TAWSS, OSI, and SPI in the 16 'matched' 4D-CTA and 3D-RA cerebral aneurysms are represented in figure 2.

Figure 2

Linear regression analysis of computational fluid dynamics results for (A) 'original' and (B) 'matched' 4D-CTA and 3D-RA models. Variability in 'original' models may have arisen from differences in inlet clipping location, inlet and outlet length, and consequently inlet flow rate estimation and outflow division. Attempts to reduce this variability and control for modality alone (4D-CTA vs 3D-RA) was made for 'matched' models (B), which were re-edited to match inlet and outlet number and lengths and were simulated using matched patient-specific flow rates, calculated from 4D-CTA. Regression lines are solid black; lines of unit slope are dashed. TAWSS, time-averaged wall shear stress (Pa); OSI, oscillatory shear index; SPI, spectral power index.

As shown in figure 3, cycle-averaged, oscillatory, and fluctuating high frequency WSS patterns were consistent between 4D-CTA and 3D-RA models irrespective of aneurysm size, geometry, and location. Not only were these patterns consistent for large aneurysms (aneurysms 1, 5, 10, and 12), but they were consistent for small aneurysms with overestimated neck planes (aneurysms 3 and 4) as well as those with under-resolved aneurysm blebs (aneurysms 6 and 7). The largest outlier on TAWSS and OSI plots was aneurysm 13, a small (<4 mm) aneurysm located distal to giant aneurysm 12.

Figure 3

Computational fluid dynamics (CFD) results comparison between 16 aneurysms from 'matched' 4D-CTA and 3D-RA models. (A) Time-averaged wall shear stress (TAWSS, in Pa), (B) oscillatory index (OSI), and (C) spectral power index (SPI) are demonstrated between 3D-RA (left) and 4D-CTA (right) aneurysm domes. All CFD simulations were performed using patient-specific flow rates calculated from 4D-CTA.


Our study has shown that aneurysm and arterial lumen geometries segmented from 3D-RA and 4D-CTA are comparable. Parent artery cross-sectional mean diameters were slightly but significantly smaller, and sac volumes were negligibly smaller for 4D-CTA models compared with 3D-RA models, with the exception of small aneurysms which demonstrated larger percent differences. Regarding hemodynamic results, there was a moderate correlation between TAWSS and OSI and a poor correlation for SPI for our 'original' 4D-CTA and 3D-RA models. However, after controlling for inflow rates and branch lengths with the 'matched' models, it improved significantly for TAWSS and SPI. In the end, for these matched models, wall shear stress, oscillatory and fluctuating high frequency patterns were consistent irrespective of aneurysm size, geometry, and location. Our study highlights that, when operator-introduced variability and flow rate are controlled for, CFD results obtained from 4D-CTA are comparable to those obtained from 3D-RA.

The fate of an aneurysm—whether it will grow, rupture or remain stable—is influenced by a multitude of intrinsic and external factors including genetics and the biological environment, which includes the blood flow dynamics. CFD can be used to study how hemodynamics influence the initiation and evolution of cerebral aneurysms.13 14 23 However, overall there is a lack of consistency with CFD methods causing inconsistent results which is hindering their use clinically.8 24 Variability may be derived from different sources including imaging modality quality (3D-RA vs CTA vs MRA),9 10 15 25 segmentation methods (ie, watershed vs threshold-based),4 solver settings,26 27 and/or boundary condition assumptions or estimations made.28 Although 3D-RA imaging is regarded as the gold standard for CFD assessment due to its superior image quality, it too comes with a number of limitations including its invasiveness and lack of patient-specific flow rates.

Aneurysm models from 4D-CTA produce reliable CFD results

The geometric results in both modalities were comparable with the exception of slight differences in parent artery lumen diameter and curvature, possibly due to the catheter in 3D-RA. Also, although not statistically significant, there was a trend towards 4D-CTA models underestimating aneurysm sac volumes and overestimating aneurysm neck size, particularly for smaller aneurysms. One could speculate that the variance was due to differences in spatial resolution (4D-CTA 0.5 mm vs 3D-RA 0.15–0.35 mm); however, our group and others previously performed in vitro experiments using aneurysm models and showed that vessel diameters were consistent between both imaging modalities.10 12 Another possibility is that injection of the large contrast bolus by the automatic injector pump during the 3D-RA acquisition was causing an enlargement of the proximal ICA due to compliance of the vessels. Another hypothesis for this discrepancy could be attributed to the drugs administered to the patient by the anesthetist during the endovascular procedure inducing vasodilatation, since all patients were under general anesthesia as per our institution’s standard of care.

The hemodynamic results appeared to be affected by the imaging modality for our 'original' models. This was likely due to the fact that these models relied on estimated inflow rates based on parent artery inlet diameter. These models were clipped at different locations; 3D-RA more distally than 4D-CTA models due to angiographic catheter placement just proximal to the petrous segment. Therefore, 3D-RA estimated flow rates were on average higher than 4D-CTA models, with aneurysm 15 being the most extreme case with a 53% discrepancy in inlet diameter. Since flow instabilities are highly influenced by inflow rate,29 we were not surprised to see that correlation of SPI between 3D-RA and 4D-CTA ‘original’ models was very poor.

However, after matching flow rates for both modalities and trimming geometries to control for operator-introduced branch differences, correlation between TAWSS, OSI, and SPI improved dramatically for 'matched' models. The improvement was greatest for SPI since inflow rate was now controlled between both models. This highlights the value of obtaining patient-specific flow rates for CFD, particularly to assess sensitive parameters like SPI, and also removing another operator-introduced assumption.

The improvement of TAWSS and SPI correlation for 'matched' models can be attributed in part to the same patient-specific flow rates used for both models. Because 4D-CTA models were 8% smaller in diameter, velocities were therefore higher (ie, by ~16%) than the 3D-RA models which could explain the slopes >1 for WSS and SPI. On the other hand, since flow rates for the 'original' models were estimated under the assumption of a fixed inlet velocity, slopes were expectedly closer to 1. Slight overestimation of TAWSS and SPI for 4D-CTA models should not be overlooked and velocity correction should be considered if 3D-RA diameters are regarded as the gold standard.

Benefits of 4D-CTA as an alternative follow-up imaging protocol for cerebral aneurysms

The results of this study demonstrate the following potential advantages of 4D-CTA imaging as a follow-up imaging protocol for hemodynamic assessment of cerebral aneurysms:

  1. The 4D-CTA imaging protocol is non-invasive, which means it is safer and more cost-effective than catheter angiographic imaging and can be more easily implemented in peripheral hospitals that do not have endovascular capabilities.

  2. It provides patient-specific flow rates for CFD analysis. It is well established that flow rate can have a great impact on various hemodynamic parameters and thus differences in the ways which flow rate is estimated can introduce large variability into CFD results. As demonstrated in this study and by others, the length of proximal anatomy that is kept and where the parent artery is clipped can impact calculation of the inlet diameter and influence the flow rate depending on the estimation method used (ie, power laws).21 28 4D-CTA may remove this bias by providing accurate (in vitro validated) flow rate calculations.12

  3. 4D-CTA provides accurate geometric and hemodynamic results. The automatic subtraction built into our 4D-CTA protocol provides angiographic images free from bone, which allows for improved segmentation compared with traditional CTA methods. The results of our study show that 4D-CTA provides consistent geometric and hemodynamic results compared with 3D-RA. An alternative dose-free follow-up imaging modality is MRI, since phase-contrast MRI could be used to provide patient-specific flow rate (and even waveform) information; however, the spatial resolution of contrast-enhanced MRA is inferior to CTA. Time-of-flight MRA offers further inferior spatial resolution and potential slow-flow artifacts in the aneurysm sac.

​Study limitations

First, our study population was limited to cases in which preoperative imaging was required for endovascular treatment strategy decision-making and vascular measurements for device planning, and therefore may not be representative of the general aneurysm population. Also, 4D-CTA protocols carry with them an ionizing radiation dose. Although less than that of a diagnostic cerebral angiogram,30 the absorbed dose of our 4D-CTA protocol is 275 mGy and the effective dose is 5 mSv.


The results of our study show that CFD performed on 4D-CTA models produces reliable hemodynamic information regarding cerebral aneurysms. 4D-CTA should be considered as a follow-up imaging tool for hemodynamic assessment of cerebral aneurysms since it is non-invasive and reduces variability in CFD results by providing patient-specific flow rates. While 3D-RA may be the nominal gold standard for aneurysm CFD, its invasiveness and pitfalls may render it no closer to the 'truth', or at least not significantly different, than models derived from less invasive 4D-CTA.



  • Twitter @NMCancelliere, @lovblad, @biomedsimlab

  • Contributors All authors contributed to the design of the study, the acquisition and/or analysis of the data, and the drafting and/or editing of the manuscript. All authors approved the final manuscript.

  • Funding This study was supported by grant G-16-00012564 from the Heart and Stroke Foundation of Canada. Computations were performed on Compute Ontario’s Niagara cluster, with priority core-hours provided by Compute Canada.

  • Competing interests None declared.

  • Patient consent for publication Not required.

  • Provenance and peer review Not commissioned; externally peer reviewed.

  • Data availability statement Data are available upon reasonable request.