Automated Hemodynamic Assessment for Cranial 4D Flow MRI

Abstract

Cranial 4D flow MRI post-processing typically involves manual user interaction which is time-consuming and associated with poor repeatability. The primary goal of this study is to develop a robust quantitative velocity tool (QVT) that utilizes threshold-based segmentation techniques to improve segmentation quality over prior approaches based on centerline processing schemes (CPS) that utilize k-means clustering segmentation. This tool also includes an interactive 3D display designed for simplified vessel selection and fully automated hemodynamic visualization and quantification.

The performances of QVT and CPS were compared in vitro in a flow phantom and in vivo in 10 healthy participants. Vessel segmentations were compared with ground-truth computed tomography in vitro (29 locations) and manual segmentation in vivo (13 locations) using linear regression. Additionally, QVT and CPS MRI flow rates were compared to perivascular ultrasound flow in vitro using linear regression. To assess internal consistency of flow measures in vivo, conservation of flow was assessed at vessel junctions using linear regression and consistency of flow along vessel segments was analyzed by fitting a Gaussian distribution to a histogram of normalized flow values. Post-processing times were compared between the QVT and CPS using paired t-tests.

Vessel areas segmented in vitro (CPS: slope=0.71, r=0.95 and QVT: slope=1.03, r=0.95) and in vivo (CPS: slope=0.61, r=0.96 and QVT: slope=0.93, r=0.96) were strongly correlated with ground-truth area measurements. However, CPS (using k-means segmentation) consistently underestimated vessel areas. Strong correlations were observed between QVT and ultrasound flow (slope=0.98, r=0.96) as well as flow at junctions (slope=1.05, r=0.98). Mean and standard deviation of flow along vessel segments was 9.33e-16 ± 3.05%. Additionally, the QVT demonstrated excellent interobserver agreement and significantly reduced post-processing by nearly 10 minutes (p<0.001). By completely automating post-processing and providing an easy-to-use 3D visualization interface for interactive vessel selection and hemodynamic quantification, the QVT offers an efficient, robust, and repeatable means to analyze cranial 4D flow MRI. This software is freely available at: : https://github.com/uwmri/QVT.

1. Introduction

Time-resolved, 3D phase contrast imaging with three-directional velocity encoding, often referred to as 4D flow MRI, allows for the non-invasive acquisition of dynamic, volumetric velocity vector fields. This approach can be used to visualize angiograms and blood flow patterns via streamlines, as well as retrospectively analyze hemodynamics in any vessel within the 3D imaging volume[1]. Among other applications, 4D flow MRI has shown promise in evaluating, diagnosing, and characterizing a wide range of cerebrovascular disease processes including aneurysms[2, 3], arteriovenous malformations[4, 5], dural arteriovenous fistulas[6], Alzheimer’s Disease[7], venous drainage[8], and functional challenges[9].

Despite these and other promising results and applications, efficient and repeatable post-processing methods for cranial 4D flow MRI remain a challenge. Typical processing steps require manual or semi-automated vessel segmentation and manual placement of double-oblique tangent planes for hemodynamic analysis, approaches that may limit reproducibility and are impractical when analyzing numerous vessels across a large number of datasets. Commercial 4D flow MRI visualization and analysis packages have recently become available but are targeted toward chest applications and are less suitable for the complexity of the brain vasculature. Furthermore, cranial 4D flow analyses often require the simultaneous analysis of numerous, small vessels, which places a higher demand on segmentation accuracy and can increase post-processing times.

Recently, a centerline processing scheme (CPS) analysis platform for cranial 4D flow MRI analysis was introduced, which automatically segments the cranial vasculature and positions equidistant tangent planes along all vessel centerlines[10]. This approach greatly reduced cranial 4D flow MRI post-processing times and increased robustness over manual analysis, enabling streamlined and reproducible processing in larger patient cohorts[11]. However, limitations still exist in selecting vessel segments from the 2D, single-view CPS interface, particularly for assessing smaller vessels such as the anterior and posterior cerebral arteries. Furthermore, the CPS utilizes k-means clustering segmentation which was recently shown to underestimate flow values compared to 2D phase contrast MRI due to overly conservative vessel segmentation[12].

The primary goal of our work was to develop and validate a tool for rapid, robust, and repeatable cranial 4D flow MRI analysis while providing users with an intuitive and user-friendly platform to visualize hemodynamic parameters across complex vascular networks. Here, we introduce the quantitative velocity tool (QVT), a freely available post-processing analysis platform that further automates and advances the previously established CPS. Specifically, we (1) implement an automated threshold-based segmentation method for improved vessel segmentation, (2) improve efficiency of vessel centerline generation and hemodynamic quantification, and (3) create an interactive, 3D user interface that allows for intuitive vessel selection and color map visualization of hemodynamic metrics such as flow, area, etc. We validate segmentation performance and flow quantification of the QVT in a flow phantom and 10 human subjects and compare it to CPS processing. Additionally, the total processing times for both tools are measured and compared.

2. Materials and Methods

2.1. Reconstruction

All 4D flow MRI data were acquired with a radially undersampled trajectory, PCVIPR[13, 14]. An automated offline reconstruction was performed which included: PILS (Parallel Imaging with Localized Sensitivities) to reduce the spread of undersampling artifacts[15], Maxwell term phase offsets corrections[16], 4D Laplacian unwrapping to correct velocity aliasing[17], and 3^rd-order polynomial background phase correction[18]. Time-resolved velocity, magnitude, and complex difference datasets were reconstructed into 20 cardiac phases using retrospective cardiac gating with radial temporal view sharing of higher spatial frequencies[19]. Additionally, time-averaged datasets of velocity, magnitude, and complex difference were obtained.

2.2. Post-Processing

A newly developed quantitative velocity tool (QVT) was implemented in MATLAB (version 2018b, MathWorks, Natick, MA, USA) building upon a previously established CPS[10]. The post-processing workflow is completely automated and is briefly outlined as follows: (1) global angiogram generation, (2) vessel centerline creation (skeletonization), (3) centerline branch labeling and smoothing, (4) tangent plane creation, (5) in-plane segmentation, and (6) computation of blood flow parameters at each centerline point. Following post-processing, a user selects vessel segments of interest through an interactive volumetric display and local, vessel-specific hemodynamics are subsequently saved. A flow chart describing the steps used to complete data reconstruction, post-processing, and analysis/visualization is shown in Figure 1. The automated post-processing chain will be described in more detail in the following paragraphs.

Figure 1:

Flow chart representing each step in the quantitative velocity tool (QVT) processing pipeline, including data reconstruction, post-processing, and analysis/visualization.

For global angiogram generation, a threshold-based segmentation algorithm was applied to the time-averaged complex difference dataset[20]. To automatically find an appropriate threshold value, a novel “sliding threshold” algorithm was developed. First, the complex difference dataset was normalized from 0 to 1. Next, the upper threshold for inclusion in the angiogram was iteratively varied from 0 (all voxels included) to 1 (only the voxel with the highest signal intensity included) in increments of 0.001. The number of voxels included within the imaging volume for each threshold value was then recorded. This data was then smoothed with a Gaussian kernel and the point of maximum curvature was chosen as the final global threshold. This process was empirically found to consistently minimize the inclusion of background voxels even in high-noise datasets.

Using the 3D global angiogram, a vessel skeleton was created with a homotopic thinning algorithm[21] creating vessel centerlines with 1 voxel diameters. Dilation and thinning algorithms were used consecutively to connect broken centerline segments and remove vessel spurs prior to vessel labeling. The QVT identified vessel endpoints and junctions, facilitating automated generation of unique classification labels for each vessel branch. All branches were spatially smoothed using 3D spline interpolation to improve orthogonal plane generation. A tangent vector was created using neighboring pixels along the centerline to generate orthogonal cutplanes for all centerline locations. A 20×20 pixel plane was used to capture the large range of arterial and venous cranial vessel diameters while also minimizing extravascular background signal. Similar to global segmentation, local segmentation was performed within each cutplane using the threshold-based method (sliding threshold algorithm) described previously. Because the cutplane was generated about the centerline, the vessel of interest was centered within the plane. To remove additional extravascular signal (e.g., closely adjacent vessels near the edge of the plane), only the segmented region with the centroid closest to the center was kept.

After performing local segmentation, hemodynamic parameters were automatically calculated at all centerline points, including mean and maximum velocity over the cardiac cycle, total volumetric flow rate, vessel area, pulsatility index[22], and resistivity index[23]. To reduce noise in calculated parameters, quantitative values were spatially averaged between 5 adjacent centerline points (current point plus 2 proximal and 2 distal points). Standard deviation of each calculated parameter was also computed as a metric of stability in the measurement region. After computing vessel areas and hemodynamics, a compressed MATLAB structure file containing all relevant data for all centerline points in the entire imaging volume was automatically stored to allow for fast reloading of 4D flow MRI studies.

Finally, data was loaded into the QVT, shown in Figure 2. The QVT is composed of 2 main graphical user interfaces, the interactive 3D interface and the control window. The interactive 3D interface (Figure 2A) is used to display the global angiogram and vessel centerlines, allowing the user to rotate, zoom, and pan the angiogram, as well as interactively select vessels for vessel-specific hemodynamic analysis. To display quantitative parameters over the entire vasculature, centerlines are color-coded with hemodynamic parameters of choice (Figure 2A, Figure 3). Upon selecting a centerline point of interest from the 3D interface, 2D orthogonal images of time-averaged magnitude, complex difference, through-plane velocity, and time-resolved velocity (with overlaid vessel segmentation masks) are shown along with flow curves plotted over the cardiac cycle, as shown in Figure 2B. Cutplane images and flow curves are updated in real-time in the control window upon selecting a new centerline point on the interactive 3D interface, allowing the user to visually inspect image, segmentation, and flow data quality at each centerline point before saving results. Data can then be saved at multiple vessel segments of interest by labelling each analysis point from a pre-defined list of vessel names. This generates tabulated hemodynamic values (time-averaged and time-resolved) for each vessel segment, along with screenshots of the interactive 3D interface, control window, and cut-planes images from 5 neighboring vessel segments to allow for easy trouble shooting during data analysis. Lastly, an additional graphical user interface (Visual Tool) was incorporated to allow for more advanced visualization of the global angiogram, with capabilities of overlaying velocity vector glyphs as well as axial, coronal, or sagittal magnitude slices (Figure 2C). A video demonstration of the QVT is provided in Supplementary Video 1.

Figure 2:

Quantitative velocity tool (QVT) graphical user interfaces. (a) Interactive 3D interface for vessel selection, showing the vessel centerlines color-coded by flow rates with the overlaid semi-transparent global angiogram. In the figure, a centerline point is selected in the left cavernous ICA (red box). (b) QVT control window that updates local segmentations and flow profiles in real-time as the centerline point of interest is changed on the interactive 3D interface. Vessel-specific hemodynamic information can be saved under unique vessel labels. Additional options to adjust visualizations, such as changing centerline width, are available in this window. Note that the five flow profiles shown represent upstream (red), current (black), and downstream (blue) profiles relative to the current point of interest. (c) Double-oblique and sagittal views of cranial vascular anatomy with the global angiogram are shown in red as opaque and semi-transparent, smoothed isosurfaces. Magnitude slices from the 4D flow MRI acquisition can be simultaneously displayed as axial, sagittal, and/or coronal 2D slices. Additionally, 3D velocity vector arrows, or “glyphs”, can be displayed within the angiogram, which are color-coded by velocity magnitude and point in the direction of blood flow in each voxel.

Figure 3:

Color-coded centerline displays for automatically computed hemodynamic parameters of vessel area (a), flow (b), mean velocity (c), and pulsatility index (d). These interactive 3D displays allow the user to view a complex vascular network in a single image and visualize both local and global hemodynamic changes from any view angle and zoom factor.

2.2. In Vitro Validation

In vitro validation was performed in a silicon-based arterial cranial flow phantom (Model H+N-R-A-002+, Shelley Medical Imaging Technologies, London, ON, Canada) to assess flow and vessel area quantification in a controlled environment (Figure 4). A pulsatile displacement pump (Model PD-1100 BDC Laboratories, Wheat Ridge, CO, USA) was connected via distensible tubing to the flow phantom to form a closed loop system free of air. A pulsatile flow profile mimicking arterial blood flow was produced at the pump outlet and was maintained at a rate of 60 beats per minute. In order to improve signal in the flow phantom, the fluid used for the experiment was gadolinium-doped water with 2 ml MultiHance (gadobenate dimeglumine, Bracco Diagnostics Inc, Milan, Italy) for 3 L of water. Pulsatile waveforms were generated at five physiologically realistic input flow rates (0.8, 0.9, 1.0, 1.1, and 1.2 L/min). Prior to MR imaging, a 0.25-inch non-intrusive ultrasonic flow sensor (PXL Clamp-On Flowsensor Transonic Systems Inc., Ithaca, NY, USA) was used to measure flow rates at the phantom’s inlet and six outlets to ensure equal distribution of flow throughout the phantom. During MRI acquisitions, a similar 0.75-inch ultrasonic flow sensor was placed at the pump outlet in the console room to record and verify pump flow rates, well before the phantom input for compatibility with the MRI system.

Figure 4:

(a) Realistic cranial arterial silicone phantom model used for in vitro studies. (b) CT angiogram of the cranial phantom used for reference area measurements and to display locations for area and flow calculations.

A total of seven 4D flow MRI scans were acquired at each flow rate, including two repeat scans performed at 1.0 L/min. Complete volumetric coverage of the cranial phantom was accomplished with the following imaging parameters: TR/TE = 7.7/2.6 ms; flip angle = 8°; number of projections = 11,000; reconstruction matrix= 320×320×320; acquired isotropic resolution = 0.69 mm; imaging volume = 22×22×16 cm³; scan time = 5:40 min; encoding scheme = 4-point referenced. For all flow rates, velocity encoding sensitivity (VENC) was maintained at 80 cm/s to be consistent with our currently used cranial 4D flow MRI protocol. Cardiac gating was recorded using an MRI-compatible LED connected to the scanner’s photoplethysmogram sensor for peripheral gating (Shelley Medical Imaging Technologies, London, ON, Canada) that simulated the ECG waveform output directly from the flow pump. All MRI scans were acquired on a clinical 3T Discovery 750 system (GE Healthcare, Waukesha, WI, USA) with a 32-channel head coil.

The QVT and CPS were used for the assessment of all quantitative parameters derived from the acquired 4D flow MRI scans. Segmentation performances between the CPS (using a k-means clustering approach[24]) and the QVT (using a local threshold-based approach) were evaluated by comparing obtained cross-sectional vessel areas to high-resolution computed tomography (CT) area measurements, using CT as the reference standard. A 3D cone-beam CT of the phantom was obtained with a Siemens Artis zee biplane system using a Syngo DynaCT reconstruction (Siemens Healthineers, Forcheim, Germany). The scan parameters for the CT included: acquisition matrix = 512×512; isotropic resolution = 0.38 mm; number of slices = 488; image volume = 19.3×19.3×18 cm³. The contrast between the silicon mold and the air-filled vessels was sufficient for signal separation, eliminating the need for iodinated contrast agents to avoid potential complications from air bubbles. The CT angiogram was created with manual thresholding using dedicated segmentation software (Mimics, Materialise, Brussels, Belgium). The CT angiogram was registered to the MR coordinate system in MATLAB using an intensity-based, regular-step gradient descent registration with mutual information as a similarity metric. For all scans, 29 vessel locations (Figure 4) were chosen to compare k-means and threshold-based vessel areas to CT measurements for a total of 203 points (7 scans × 29 vessel locations). A flow comparison between QVT-derived flow rates and the reference standard ultrasonic flow sensor was performed. Because flow could not be directly measured at the phantom inlets or outlets during MRI acquisition, output pump flow from the 0.75-inch probe was compared to the sum of flow from the 3 proximal vessels near the phantom inlet (shown in Figure 4) derived from the QVT.

2.3. In Vivo Validation

In vivo validation was performed on 10 healthy participants (8 women, mean age: 56.4 years, age range: 48–67 years) after informed consent and Institutional Review Board approval. Non-contrast 4D flow MRI scans were acquired with the same magnet, coils, and scan parameters as the in vitro studies. Data reconstruction and pre-processing steps were also consistent with the in vitro study. QVT and CPS were used for the evaluation of area and flow parameters from the acquired 4D flow MRI scans. To obtain ground-truth vessel segmentations, 1 user (C.A.H, 6 years of 4D flow processing experience) manually segmented vessels from 13 arterial and venous vessel locations: inferior and superior internal carotid arteries (ICA, 4 planes), basilar artery (BA, 1 plane), middle cerebral arteries (MCA, 2 planes), posterior cerebral arteries (PCA, 2 planes), straight sinus (SS, 1 plane), superior sagittal sinus (SSS, 1 plane), and transverse sinuses (TS, 2 planes). Additionally, 5 total planes (4 adjacent to center) were extracted from each location giving a total of 650 vessel segmentations. Vessel areas were obtained in these same locations using the CPS and QVT to allow for quantitative comparison to ground-truth manual segmentation.

Due to the absence of ground-truth flow measurements in vivo, internal consistency of flow was used for validation. Specifically, conservation of flow at vessel junctions and consistency of flow along continuous vessel segments was assessed. A total of three vessel junctions were analyzed: ICA (inlet), MCA (outlet), and anterior cerebral artery (outlet) for both the left and right sides, as well as SSS (inlet), SS (inlet), and left and right TS (outlets). Flow along a vessel was measured at every 5^th point vessel centerline point for three vessel segments: left ICA, right ICA, and SSS. Values were normalized by the mean and converted to a percentage from the mean to allow flow values from different ranges and magnitudes to be compared equally.

In vivo flow quantification was performed by 2 users (C.A.H. and G.S.R.) with 6 and 5 years of experience in MRI flow analysis respectively. All QVT analyses were performed independently with the following desktop system specifications: Dell Precision 5820 Tower, Intel Xeon W-2123 3.60 GHz 4-core CPU, 32 GB RAM, 500 GB SATA hard drive, and Windows 10 operating system. Times to complete each post-processing step including vessel selection were recorded for the QVT and CPS for all 13 vessel locations. If the vessel of interest was not present or selectable, the vessel was recorded as missing and was excluded from analysis. Interobserver repeatability was assessed for both CPS and QVT methods using measures of mean flow at all vessel locations for each participant.

2.4. Statistical Analysis

To statistically assess vessel area and flow measures derived from QVT and CPS relative to reference standards in vitro, linear regression was used to compare: (1) CPS segmented vessel areas against CT areas, (2) QVT segmented vessel areas against CT areas, and (3) QVT flow against ultrasound flow. For the in vivo study, linear regression was used to compare: (1) CPS areas against manually segmented areas, (2) QVT areas against manually segmented areas, and (3) QVT flow at junction inlets against QVT flow at junction outlets. Associated model p-values, 95% confidence intervals, and Pearson’s correlation coefficient (r) were calculated for each linear regression model. In vivo segmentation performance was also assessed with Dice coefficients for both methods relative to ground-truth manual segmentation. Conservation of flow along a vessel was assessed by fitting the percent flow variation for all vessels to a Gaussian curve using a minimum variance unbiased estimator and providing the 95% CI for the estimated parameters (mean and standard deviation). Interobserver repeatability for QVT and CPS flow measures was assessed with a Bland-Altman analysis. Paired t-tests were used to compare post-processing times between the CPS and QVT. Statistical significance was defined as p<0.05. All statistical analyses were conducted in MATLAB.

3. Results

4D flow MRI reconstructions, including additional data corrections, were successfully completed for all scans. ECG gating files did not contain significant variations in heart rates and visual inspection of cross-sectional planes used for analysis, both in vitro and in vivo, showed no visible velocity aliasing and included all relevant vasculature. All reconstructed data were automatically post-processed and analyzed successfully with the QVT and CPS.

3.1. In Vitro Validation

Average flow rates from the pulsatile flow pump measured at the pump outlet during MRI acquisition ranged from 0.82 – 1.21 L/min. Pulsatile flow was visible in both the ultrasound probe measurements and throughout the phantom vasculature as visualized by the QVT. The linear regression coefficients between CPS and CT-segmented areas were calculated as: A_CPS = 0.71·A_CT + 4.35, where A_CPS is the automatic k-means vessel area and A_CT is the CT area (in units of mm²). The 95% confidence intervals for the slope and intercept were [0.68, 0.74] and [3.93, 4.77], respectively. The linear regression coefficients between QVT and CT-segmented areas were calculated as: A_QVT = 1.03·A_CT + 3.51, where A_QVT is the threshold-based vessel area. The 95% confidence intervals for the slope and intercept were [0.99, 1.07] and [2.91, 4.11], respectively. A very strong correlation between the variables was observed for both the CPS and QVT segmentation methods (CPS: r = 0.95, p<0.001; QVT: r = 0.95, p<0.001). The linear regression coefficients between QVT-derived flow and ultrasound flow were calculated as: F_QVT = 0.98·F_US + 0.05, where F_QVT is the QVT-derived inlet flow and F_US is ultrasound flow measured at the pump outlet (units of L/min). The 95% confidence intervals for the slope and intercept were [0.74, 1.22] and [−0.20, 0.29], respectively. A very strong correlation between the variables was observed for quantitative flow measures (r = 0.96, p<0.001). Linear regression plots for all in vitro comparisons are shown in Figure 5.

Figure 5:

Regression plots showing the relationship between in vitro vessel areas obtained using high-resolution CT and (a) the centerline process scheme (CPS) using k-means segmentation, and (b) the quantitative velocity tool (QVT) using threshold-based segmentation. Both techniques were highly correlated with CT, with the QVT method giving a slope closer to unity. However, there appears to be a slight overestimation of vessel areas using threshold-based segmentation. (c) Linear regression plot of flow derived from the ultrasonic probe and the QVT showed a strong correlation with a slope near unity.

3.2. In Vivo Validation

A total of 645 cut-planes (1 missing transverse sinus) were successfully segmented manually for comparison. Average Dice coefficients were 0.77 ± 0.07 (standard deviation) for the CPS and 0.91 ± 0.06 for the QVT compared to manual segmentation. Several illustrative examples of the differences between each segmentation method are provided in Figure 6.

Figure 6:

Segmentations were completed for both k-means and threshold-based methods (top) on the curve of the left internal cerebral artery, (middle) in the right transverse sinus, (bottom) and in the anterior cerebral artery. In all cases, the QVT threshold-based segmentation outperformed k-means relative to manual segmentation. Arrows were placed in the images to indicate potential sources of error related to slow flow or other vessels present in the region of interest.

The linear regression coefficients between CPS and manually segmented cross-sectional areas were calculated as: A_CPS = 0.61·A_MAN + 0.38, where A_CPS is the automatic k-means segmented area and A_MAN is the manually segmented vessel area (in units of mm²). The 95% confidence intervals for the slope and intercept were [0.60, 0.63] and [0.08, 0.68], respectively. The linear regression coefficients for QVT and manually segmented areas were calculated as: A_QVT = 0.93·A_MAN – 0.00, (units of mm²), where A_QVT is the threshold-based segmented area. The 95% confidence intervals for the slope and intercept were [0.92, 0.95] and [−0.44, 0.43], respectively. A very strong correlation between the variables was observed for both the CPS and QVT segmentation methods (CPS r = 0.96, p<0.001; QVT r = 0.96, p<0.001). Linear regression between the inlet and outlet flow from vessel junctions was calculated as: F_IN = 1.05·F_OUT – 0.21, where F_IN is QVT-derived inlet flow and F_OUT is QVT-derived outlet flow (in units of mL/s). The 95% confidence intervals for the slope and intercept were [0.98, 1.13] and [−0.68, 0.26], respectively. A very strong correlation between the inlet and outlet flow was observed (r = 0.98, p<0.001). Flow values along a vessel were successfully converted to a percentage from the mean and plotted in a histogram to assess variation of flow along single vessel segments. The histogram Gaussian curve fit resulted in a mean and standard deviation of 9.33e-16% [−0.35, 0.35] and 3.05% [2.82, 3.31], respectively. Linear regression and histogram plots for in vivo area and flow assessments are shown in Figure 7.

Figure 7:

Regression plots showing the relationship between in vivo vessel areas obtained with manual segmentation and (a) the centerline process scheme (CPS) using k-means segmentation, and (b) the quantitative velocity tool (QVT) using threshold-based segmentation. Both techniques were highly correlated to manually segmented areas, with the QVT method giving a slope closer to unity. (c) Linear regression of the conservation of flow at junctions using the QVT method showed a strong correlation with a slope near unity for the in vivo experiments. (d) The normalized percent variation of flow was nearly zero along continuous vessel segments, indicating excellent internal consistency of flow measures.

Between both users, the percentage of missed vessels from the 13 total vessel locations over the 10 participants was 8.1% (21/260 measurement points) for the CPS and 1.9% (5/260 measurement points) for the QVT. Smaller vessels and vessels with slow flow, namely the MCA, PCA, and non-dominant TS, accounted for the majority of missed vessels for both methods. Bland-Altman analysis of interobserver repeatability for QVT flow quantification demonstrated a mean bias of 0.042 mL/s with 95% limits of agreement of [−0.37, 0.45 mL/s], shown in Figure 8. Interobserver repeatability for the CPS showed a mean bias of 0.037 mL/s with 95% limits of agreement of [−1.05,1.13]. Average times for angiogram generation, data loading, vessel selection, as well as total case analysis times, are given in Table 1 for both CPS and QVT. While data loading time was significantly greater for the QVT (p<0.001), the time required for angiogram generation, vessel selection, and to complete a total case was significantly reduced (all p<0.001).

Figure 8:

Bland-Altman plot and analysis from 10 cranial cases (13 vessel segments) analyzed using the quantitative velocity tool (QVT) by two independent users. This analysis showed a mean bias of 0.042 ml/s with 95% limits of agreement of [−0.37, 0.45 ml/s].

Table 1:

Post-Processing Times for CPS and QVT Methods

Method	Angiogram (min)	Load Data* (min)	Vessel Select (min)	Total Case (min)	Per Plane (min)
CPS	0.82 ± 0.14	1.02 ± 0.18	15.6 ± 3.36	17.5 ± 3.36	1.20 ± 3.20
QVT	0.20 ± 0.02	2.34 ± 0.42	4.71 ± 0.88	7.94 ± 0.98	0.36 ± 0.97

Processing times are reported as mean ± 1 standard deviation (minutes). QVT = Quantitative Velocity Tool; CPS = Centerline Processing Scheme; RAM = random-access memory.

^*Data loading for QVT included saving reloadable MATLAB file structures.

4. Discussion

This work presents a novel, freely available quantitative velocity tool (QVT) that allows for automated quantification of hemodynamic parameters generated from cranial 4D flow MRI. The accuracy and repeatability of quantitative metrics derived from the QVT, specifically vessel area and flow, were validated both in vitro and in vivo. This tool provides a unique, interactive 3D visualization interface that can simultaneously display anatomic and color-coded quantitative vascular information while also allowing users to visualize data from cutplanes and flow profiles on a vessel-by-vessel basis and supports more stable estimation of parameters by incorporating measures from adjacent slices. Combined, these features allow for quick, repeatable, and robust analyses to investigate hemodynamic parameters throughout the brain which would otherwise be difficult and time-consuming to perform using alternative post-processing software or flow-sensitive imaging techniques. Furthermore, the improvements in automation, accuracy, and time savings may make the application of 4D flow MRI more widely available in the clinical setting. The tool and its source code have been made freely available for the research community (see Data Availability Statement).

Quantitative 4D flow MRI-derived parameters, such as blood flow, mean velocity, pulsatility index, and resistivity index, directly rely on accurate segmentation of vessel cross-sections. Previous studies have utilized automatic segmentation methods, including global thresholding[25], k-means clustering[10], and local thresholding[12] to help improve accuracy required for clinical use. A recent study comparing each of these 3 methods suggested that threshold-based methods may improve segmentation stability[12]. Here, we implemented an automated threshold-based segmentation methodology with a novel sliding threshold algorithm that provided accurate and efficient global segmentation (for centerline generation) combined with subsequent local in-plane segmentation (for vessel area and flow analysis). While both k-means and threshold-based segmentation techniques were highly correlated to reference standards for in vitro and in vivo studies, linear regression and Dice coefficients indicated that the segmentation from QVT outperformed CPS. The k-means segmentation consistently underestimated vessel area by approximately 30–40% while the threshold-based method reduced this to less than 10%. However, it should be noted that in vitro results showed a slight overestimation of threshold-based vessel areas relative to high-resolution CT segmentations. Morphological operations, such as erosion, may be useful in addressing this bias but more rigorous validation is required. Furthermore, threshold-based segmentation could be further improved by incorporating additional spatial and intensity-based image information[26]. While doing this would greatly increase the time and memory requirements for k-means clustering, there would be little to no changes in processing times for the threshold-based technique presented here.

Intracranial blood flow is an important hemodynamic parameter in the assessment of many vascular diseases[27, 28]. Doppler ultrasound can provide quantitative values but is limited by a small field of view, operator dependence, and cranial bone windows. Radiographic techniques such as CT and digital subtraction angiography provide excellent spatial resolution and qualitative flow information but require the use of radiation exposure and iodinated contrast agents. The application of 4D flow MRI allows for non-invasive quantitative analysis of blood flow while maintaining large volumetric coverage in a single, clinically-feasible scan. We investigated the accuracy and internal consistency of blood flow both in vitro and in vivo. The QVT-derived flow values from 4D flow MRI showed strong correlation and agreement with ultrasonic flow probe measurements for in vitro studies. In vivo analyses demonstrated that flow was conserved along continuous vessel segments and at vessel junctions, signifying good internal consistency. This was performed in both arterial and venous intracranial vasculature, providing a range of vessel sizes and flow rates for conservation of flow analysis.

The QVT improved upon a previously presented centerline process scheme (CPS) by implementing various technical and graphical developments that led to significantly faster post-processing and improved vessel selection. By automating post-processing steps (aside from vessel selection), user interaction time is decreased, and potential user-dependent errors can be mitigated. Additionally, after completing post-processing, hemodynamic information was saved into MATLAB file structures which allows for cases to be easily re-loaded into the QVT within seconds. These file structures contain only relevant 4D flow MRI data needed to run the tool which reduces memory and storage requirements and allows the QVT to run on less powerful systems such as laptops, which was not previously possible. With the QVT, we found that data loading, segmentation, vessel selection, flow assessment, and data saving for a single case (13 analyzed vessel segments) required around 8 minutes to complete from start to finish, compared to the CPS which took approximately 18 minutes. While data loading did take longer for the QVT compared to the CPS, this additional time was the result of saving reloadable MATLAB file structures (not available in the CPS). The QVT also reduced the number of vessels that could not be analyzed, primarily smaller vessels that are difficult to identify or select on the single-view CPS interface. Such reductions in processing time and resources, as well as improvements in vessel selection, allow for the possibility of performing comprehensive, large-scale flow studies.

A limitation of this study is that the QVT relies on accurate global angiogram generation. If the VENC is incorrectly set or vessels of interest include abnormal flow patterns (e.g., stenosis, aneurysm, etc.), vessel centerlines may not be properly created which can lead to missed vessel segments. Our reconstruction pipeline included a 4D phase unwrapping technique[17] to help reduce phase wrapping errors associated with VENC underestimation. The addition of dual-VENC 4D flow MRI[29] or manually segmented angiograms could also be used to address poor global segmentation due to velocity aliasing or neurovascular pathology at the cost of longer scans or post-processing times. Another limitation of this study is the lack of ground truth measurements for vessel area and velocities in vivo. To address this, we performed manual vessel segmentation and utilized internal consistency measurements of flow. The QVT was designed specifically for cranial 4D flow MRI scans and has yet to be adapted for other regions of the body. However, it could be transferred into other vascular territories with the addition of time-resolved segmentation and modifications to several post-processing steps, but the effects of larger fields of view and motion need to be thoroughly investigated. Lastly, post-processing with the QVT is currently limited to our institution’s data format but will be modified to accept more universal data formats (e.g., DICOM series) to allow for widescale usage across vendors.

5. Conclusion

In conclusion, we have developed and validated a fully automated 4D flow analysis tool (QVT) that allows for 3D visualization and quantification of cerebrovascular hemodynamics with cranial 4D flow MRI. This technique improved upon an established CPS, further reducing post-processing times through the addition of an interactive 3D interface, allowing for quick vessel selection and verification of quantitative blood flow parameters. Global and local vascular segmentation was improved through the application of a threshold-based segmentation technique. The completely automated post-processing pipeline reduced the QVT computer memory and storage requirements, allowing cranial 4D flow MRI analysis to be completed with minimal resources and increasing the repeatability of 4D flow MRI studies by reducing user-dependent errors.

Supplementary Material

2

Video S1: Video demo of the tool.

Highlights.

• 4D flow MRI allows for non-invasive assessment of cerebral hemodynamics

• We develop, validate a quantitative velocity tool (QVT) that improves existing tools

• Improved segmentation and visualization reduce 4D flow MRI analysis time

• Automated post-processing tools are well-suited for fast and repeatable flow analysis

Data Availability Statement

Our cranial 4D flow MRI analysis software has been made publicly available on Github: : https://github.com/uwmri/QVT.