4D flow MRI—Automatic assessment of blood flow in cerebral arteries

Objective: With a 10-minute 4D flow MRI scan, the distribution of blood flow to individual arteries throughout the brain can be analyzed. This technique has potential to become a biomarker for treatment decisions, and to predict prognosis after stroke. To efficiently analyze and model the large dataset in clinical practice, automatization is needed. We hypothesized that identification of selected arterial regions using an atlas with a priori probability information on their spatial distribution can provide standardized measurements of blood flow in the main cerebral arteries. Approach: A new method for automatic placement of measurement locations in 4D flow MRI was developed based on an existing atlas-based method for arterial labeling, by defining specific regions of interest within the corresponding arterial atlas. The suggested method was evaluated on 38 subjects with carotid artery stenosis, by comparing measurements of blood flow rate at automatically selected locations to reference measurements at manually selected locations. Main results: Automatic and reference measurement ranged from 10 to 580 ml min−1 and were highly correlated (r = 0.99) with a mean flow difference of 0.61 ± 10.7 ml min−1 (p = 0.21). Out of the 559 arterial segments in the manual reference, 489 were correctly labeled, yielding a sensitivity of 88%, a specificity of 85%, and a labeling accuracy of 87%. Significance: This study confirms that atlas-based labeling of 4D flow MRI data is suitable for efficient flow quantification in the major cerebral arteries. The suggested method improves the feasibility of analyzing cerebral 4D flow data, and fills a gap necessary for implementation in clinical use.

. Today, cerebral 4D flow MRI is processed manually. This requires a skilled operator and is time-consuming if many arteries are to be investigated. By automatizing this process, flow values can be obtained in a systematic manner, and the operator is free to perform other tasks in the meantime, but this type of automated post-processing tools is missing.
For a method to be considered fully automatic, it is not enough that the delineation of the vessel and separation of vessel lumen from background is automated (van der Geest et al 1998, Jiang et al 2015, or to automatically find a measurement plane perpendicular to the artery, starting from a manually selected seed point (Wåhlin et al 2013, Schrauben et al 2015. A fully automatic method should also find the correct placement of the measurements, presenting flow values for all arteries of interest. There has been work presented on fully automatic identification of cerebral arteries (Bogunovic et al 2013, Bilgel et al 2013), but not aimed specifically on flow quantification, and not adapted for 4D flow MRI.
We have previously developed a cerebral arterial atlas (UBA167), describing the spatial distribution of the main cerebral arteries connected to the Circle of Willis (Dunås et al 2016). The atlas consists of a 3D probability maps for each artery, describing the probability that the artery can be found at a particular location in the brain. We have also constructed an arterial identification method (AAIM) utilizing the information from the atlas to enable accurate identification of the main cerebral arteries (Dunås et al 2017). We hypothesized that this framework could be expanded to include standardized and automatic assessment of blood flow in these arteries. The challenge for the automatic method is to cope with the inter patient variability in cerebral arterial anatomy and find a measurement position in the identified artery that produce a flow rate estimate that correspond to the one produced when an investigator manually selects the well specified measurement position. Importantly, functionality of such an automatic method should also be robust for deviating flow patterns, e.g. compromised flow caused by a stenosis or occlusion in stroke patients.

Methods
In this study, a new post-processing tool for automatic analysis of 4D flow MRI is presented. This method is based on a previously developed method for arterial labeling where the whole artery is labeled (Dunås et al 2016), which in this study is extended to allow for standardized flow measurements.
This new post-processing tool was validated on subjects with carotid artery stenosis (CAS), both in terms of labeling accuracy and flow assessment. For the vessel segmentation and flow quantification we used a previously developed method (Wåhlin et al 2013). Automatic measurements were compared to measurements at manually selected locations.

Subjects
Thirty-eight stroke patients with CAS (27 men, 11 women, age 72.5±5.7 years), admitted to the tertiary stroke center at Umeå university hospital during 2012-2015, were included in this study and investigated with magnetic resonance imaging, including 4D flow MRI.
Inclusion criteria were a transient ischemic attack or stroke, with a corresponding CAS50% on the symptomatic side, and eligible for carotid endarterectomy. Patients with Mini-Mental State Exam <24, modified Rankin scale >2 or ICA occlusion contralateral to the symptomatic side were excluded. Ongoing atrial fibrillation, severe aphasia or other previous neurological disease in the central nervous system were also exclusion criteria. Patients with contraindication for MRI examination were excluded as well. The CAS grading was performed using computed tomography angiography (GE Light Speed VCT 64, Waukesha, WI, USA) data in 31 patients (82%) and carotid artery ultrasound (GE Logiq E9, Linear probe 9L, 9900 Innovation Drive, Wauwatosa, WI, USA or Philips iU22, Linear probe L9-3, 22100, Bothell Everett Highway, Bothell, WA, USA) in 7 patients. The regional ethics review board at Umeå University approved the study and all subject gave informed consent.

Novel approach
The new method is based on a previously developed atlas-based method for artery identification (AAIM). The challenge tackled in this study was to extend this framework to identify specific arterial regions, and to locate seed points for flow quantification corresponding to these regions. Finding such well-defined regions is crucial to obtain reliable flow measurements at standardized locations.

Flow rate quantification algorithm
The flow rate quantification method (FRQ) used in this study (Wåhlin et al 2013) was based on selection of a seed point within the vessel. The velocity vector in the selected seed point was used to determine the direction of the vessel, and blood flow was calculated in a volume defined by two cross-sectional planes perpendicular to this vessel direction, three voxels apart, surrounding the seed point. The equation used to determine which voxels to include in the selected volume is shown in figures 1(b) and (c). The vessel lumen was separated from the background by thresholding the complex difference angiographic image at ten percent of the maximum intensity value. Within the selected volume, flow rates (Q) in x-, y-and z-direction were calculated from the corresponding velocity images, by integrating across the volume and dividing the value by the distance between the two planes.
The flow rate through the arterial volume was calculated as the projection of the total flow on the directional vector of the seed voxel, where θ is the angle between the flow direction in the seed voxel and the average flow direction within the plane. This approach compensates for misalignment between the initial approximation of the vessel direction and an approximation based on the average flow direction in the vessel.

Manual reference measurements
Two independent raters viewed axial images of the angiographic volume to manually place the seed points for the measurements. The FRQ described above was used to calculate blood flow. In general, the mean flow rate across raters was used as the reference value, although in cases where the difference between the measurements from the two rates in an artery was over 20%, a consensus measurement was made. Because the flow rate in the posterior communicating artery (PCoA) can be very low, it was not feasible to apply the 20% limit to determine if measurements were correct. Therefore, for PCoA, the mean flow rate across raters was used for all cases.

Basic principle
The underlying principle of the AAIM was to label voxels within a vascular skeleton according to atlas probabilities. The vascular skeleton of a subject consisted of centerline branches, connected by junction points, where each branch corresponded to a vascular segment and had a unique identification number. To construct this skeleton, the complex difference angiographic image was thresholded to create a binary image, which was then gradually thinned and pruned (Palàgyi and Kuba 1998, Chen and Molloi 2003).
The atlas used for labeling was the UBA167 (Dunås et al 2017), which consists of probability maps describing the spatial probability of sixteen large cerebral arteries in Montreal Neurological Institute (MNI) space (Evans et al 2012). The labeling took place in the native space of the subject, therefore, using the subjects T1 weighted magnitude image, SPM8's DARTEL (Ashburner 2007) was used to transform the atlas from MNI-space into native space (Dunås et al 2016).

Defining measurement sites
To position the seed points in a standardized way, we defined artery specific measurement sites with corresponding regions in the UBA167. Thirteen of the sixteen probability maps in the UBA167 (all except distal ACA and distal MCA) were used in this implementation, and in ICA and PCA, two regions were defined in the same probability map. These atlas regions where formed as planes with 9 mm thickness and an orientation perpendicular to the arterial segment in question.
The following positions were used for measurements: All arteries except BA are measured bilaterally, giving a total of 17 defined atlas regions.

Selection of seed voxel
An illustration of the arterial labeling and selection of the seed voxel is presented in figure 1. As for the manual measurements, the FRQ was used for flow calculations. Branch voxels were first labeled according to the probabilities in UBA167, assigning each voxel to the artery with the highest probability. The labeling was then refined by removing voxels that did not fall within the defined atlas regions. Since this transform smoothens the edges of the atlas regions, probability values under 0.3% (half of the lowest value found in the original atlas) were removed before labeling, for both the full UBA167 and for the defined regions. For each region, the longest continuous segment of labeled voxels was identified. Segments shorter than eight voxels (∼5 mm) were considered unreliable and were therefore discarded. The midpoint of each identified segment was used as the seed point for the FRQ.

Evaluation
The output from the automatic method, in addition to the flow values, was a 3D image of the vascular system with color-coded regions corresponding to the labeled segments, an example of this labeling is shown in figure 2. These images were visually inspected and each labeled segment was classified as correct or incorrect, based on if the labeled segment corresponded to the correct arterial branch, and accuracy of the labeling method was calculated for each of the nine specified positions. Note that the aim of the labeling process was not to find the whole artery, but to identify a specific segment, selected for accurate flow measurements. Flow values from the automatic method were compared to the manual reference measurements in terms of flow difference.

Statistical analysis
All data processing and analyses were done in Matlab (Mathworks, Natick, MA, USA). Correlation between automatic and reference measurements was calculated, as well as mean and standard deviation of the difference. Differences between flow rates were tested with paired t-test, significance level was set to p<0.05.
Agreement between the two raters, and between manual and automatic measurements, was evaluated using intraclass correlations (ICC). In both cases, twoway absolute agreement analysis was selected. Since the mean value from the two raters was used as the reference, the multiple measurement option (ICC(2,k)) was used for that analysis, while single measurement (ICC(2,1)) was used when comparing automatic to reference measurements.

Results
Manual reference measurements were obtained for 559 arterial segments, whereof 489 were correctly identified by the automatic method. For the 87 arteries were no reference measurements where obtained, 74 were correctly identified as non-existing, yielding a sensitivity of 88%, a specificity of 85%, and an accuracy of 87% when looking at all arterial segments together. Labeling results for each arterial segment are presented in table 1, where the accuracy of 100% in BA and C4 should be specifically noted. The average length of the identified segments was 16.4±3.6 voxels (11.3 mm).
Results from flow quantification are presented in table 2. On average, the flow rate obtained with the automatic measurement was 0.61 ml min −1 higher than the manual reference, but this difference was not significant (p=0.21). The standard deviation of the difference compared to the manual reference measurements, across all arteries, was 10.7 ml min −1 . The measurements were highly correlated (r=0.99), agreement between the two methods are presented in figure 3.
Intraclass correlation for the different arteries were between 0.95 and 0.99 for the two raters, and between 0.95 and 1.00 for the automatic versus reference measurements. Calculated on all arteries together, both tests gave a intraclass correlation of 0.99. For the comparison of the two raters, all arteries had a 95% confidence interval lower bound over 0.90, and for the automatic versus reference measurements, this was true for all arteries but PCoA.

Discussion
In this study, we described and evaluated a new postprocessing tool for 4D flow MRI, where blood flow   Values are calculated from the correctly identified arteries, total cerebral blood flow (tCBF) is calculated from BA and C4, note that arteries with zero flow have not been included in the flow rate calculations for individual arteries, but are included when calculating tCBF, * p<0.05. ICA -Internal carotid artery, VA -Vertebral artery, BA -Basilar artery, PCA -Posterior cerebral artery, MCA -Middle cerebral artery, ACA -Anterior cerebral artery, PCoA -Posterior communicating artery.
rates are automatically calculated in 17 specified arterial regions. Automatic blood flow measurements had an excellent agreement with the manual reference measurements, confirming that atlas-based labeling is suitable for automatic blood flow quantification. An advantage of 4D flow MRI compared to previous methods such as ultrasound or 2D PCMRI is that data for the entire brain is sampled simultaneously, in a ten-minute MRI sequence. Thus, contrary to 2D PCMRI investigations, it is not necessary for a radiologist to decide in advance which arteries should be examined, or to place and angulate the measurement plane during the MRI acquisition. This work is instead done during the post-processing, and with the method suggested in this study, the time spent on this manual processing could be substantially reduced. The 4D flow MRI collection, together with the post processing described, are so straight forward that it would be possible to implement this new method in clinical routine.
The suggested method had an overall accuracy of 87%, and this accuracy does not seem to be affected by the likely more challenging inter-subject variation in flow rate and morphology in patients with CAS. The automatic seed point selection is supposed to function as an initial placement, with the option to make manual corrections by re-selecting or moving the seed point along the artery if the initial placement was incorrect or unsatisfying.
If flow values are required for all seventeen regions included in this study, manual editing would be needed for 13% of the arteries, but if only the larger cerebral arteries are of interest, this number could be substantially reduced. The new method revealed 100% labeling accuracy in the basilar and distal carotid arteries, making it possible to calculate total cerebral blood flow in all evaluated subjects. For the analysis of the blood supply to the cerebral vascular territories from the Circle of Willis (MCA, ACA, PCA), we have an accuracy between 88 and 95% (table 1). These values are slightly lower than what have been observed for full artery labeling, where these arteries had an accuracy of 96% or higher (Dunås et al 2016, Dunås et al 2017. This decrease in accuracy is mainly seen in ACA, due to the restricted atlas region, and the MCA, due to the new criteria that only pre-bifurcation segments are considered correctly labeled, which is necessary for correct flow measurements. Even though these values are slightly reduced, they are still considered sufficient for the purpose of this tool. For ICA, BA, VA and PCoA, results are at the same level as previous studies, while P1 have not been previously evaluated.
For MCA, errors occur when the first bifurcation of the M1 segment arises very close to the origin of the artery, resulting in the labeling of a post-bifurcation branch. This error would be easily fixed by manual editing. Lower identification accuracy was obtained for P1, PCoA and VA. If these arteries are required in a specific analysis, for example collateral circulation in the Circle of Willis or cerebellum blood supply, one can expect the need of slightly more editing time.
When looking at flow differences for specific arteries, a significant difference between automatic and manual reference measurements were obtained for ACA and P2, indicating a systematic difference in the placement of the reference measurements compared to the atlas regions. However, the difference was small, and none of the arteries had a systematic difference of more than 4%. The largest difference in flow between automatic and reference measurements was found in the VA. This was likely a manifestation of the highly variable anatomy and the branching of arteries like posterior inferior cerebellar artery, visible in figure 2. Branching arteries that are not included in the atlas does increase the risk of mislabeling, and is the main reason for the low labeling accuracy for VA.
The reported tCBF values are a bit lower than previously reported in healthy elderly (Zarrinkoob et al 2015), this is expected in subjects with CAS (Fang et al 2016), since the ICA flow is restricted. This is partly, but not completely, compensated by increased blood flow in BA.
The FRQ used in this study is robust and simple, flow values are calculated in a volume with a length of three voxels along the artery. By resampling the neighborhood of the measurement voxel, calculations could be done in a true 2D plane rather than an angled plane through a 3D volume (Schrauben et al 2015), which could result in more accurate measurements. It should also be investigated if using a more advanced segmentation than a fixed 10% threshold could give more accurate flow assessment. However, to develop and evaluate the optimal segmentation method, a study designed with an independently measured reference flow, for example with high-resolution 2D PCMRI is needed. Figure 4 shows a printout report of a 4D flow MRI investigation, produced by the automatic method described in this paper. Blood flow is reported in ml/ min for each of the labeled arteries. Although the postprocessing of each 4D flow MRI investigation is challenging and consists of a large set of data, calculations are obtained in less than half an hour with the described automated method, not including reconstruction of data, which is necessary for manual measurements as well. Since the process is automated, the only active processing time needed is selecting the cases to analyze, and reviewing the results, the rest of the processing time can be devoted to other tasks.
The post-processing tool described in this paper allows the investigator to quickly get a profile of the distribution of blood flow to the major arteries of the brain. This could for instance be used to study effects of stenosis or carotid artery dissections, or to obtain a risk profile for how well collateral circulation is able to maintain blood flow to the brain's vascular territories In conclusion, atlas-based labeling was suitable for automatic quantification of cerebral blood flow in 4D flow MRI. In total, 87% of all arteries were correctly labeled, with 100% accuracy for the large supplying arteries needed to calculate total cerebral blood flow. Agreement between automatic and manual reference measurements was excellent, with no systematic difference and a higher stability than repeated manual measurements. This study demonstrates the feasibility and utility of the atlas-based approach to assess cerebral arterial blood flow. It provides the tool for analyzing cerebral 4D flow MRI data that is needed for efficient in vivo study of arterial blood flow in the brain and should have important applications in various neurological diseases.