Development of a Quantitative Intracranial Vascular Features Extraction Tool on 3D MRA Using Semi-automated Open-Curve Active Contour Vessel Tracing

Abstract

Purpose

To develop a quantitative intracranial artery measurement technique to extract comprehensive artery features from Time-of-flight (TOF) MRA.

Methods

By semi-automatically tracing arteries based on an open-curve active contour model in a graphical user interface 12 basic morphometric features and 16 basic intensity features for each artery were identified. Arteries were then classified as one of 24 types using prediction from a probability model. Based on the anatomical structures, features were integrated within 34 vascular groups for regional features of vascular trees. Eight 3D MRA acquisitions with intracranial atherosclerosis were assessed to validate this technique.

Results

Arterial tracings were validated by an experienced neuroradiologist who checked agreement at bifurcation and stenosis locations. This technique achieved 94% sensitivity and 85% positive predictive values (PPV) for bifurcations and 85% sensitivity and PPV for stenosis, respectively. Up to 1456 features such as length, volume, and averaged signal intensity for each artery, as well as vascular group in each of the MRA images, could be extracted to comprehensively reflect characteristics, distribution and connectivity of arteries. Length for the M1 segment of the middle cerebral artery extracted by this technique was compared with reviewer measured results, and the intra-class correlation coefficient was 0.97.

Conclusion

A semi-automated quantitative method to trace, label and measure intracranial arteries from 3D-MRA was developed and validated. This technique can be used to facilitate quantitative intracranial vascular research, such as studying cerebrovascular adaptation to aging and disease conditions.

INTRODUCTION

Vascular structure and function are important factors for maintaining brain health. Increasing evidence indicates that cerebrovascular disease is an under-recognized contributor to dementia, even in the absence of brain infarctions (1,2). A more comprehensive description of the whole intracranial vascular map, as well as an accurate quantification of vascular features could allow us to understand the effects of aging and other physiological/pathophysiological conditions on cerebral arteries and how such vascular changes impact brain health. 3D Time-of-flight (TOF) Magnetic Resonance Angiography (MRA) (3) is widely used clinically to depict vascular anatomy and pathology (4–6). Current clinical review of MRA focuses on distinct stenoses or aneurysms in only the major arterial segments. Comprehensive feature extraction using automated algorithms may provide a detailed map of intracranial arteries from routine MRA with global as well as regional information on blood supply that have been largely neglected.

However, it is challenging to comprehensively characterize cerebral arteries from MRA images because the cerebral vasculature is a complex network with substantial inter-individual variation (7). In addition, small but anatomically important arteries, such as communicating arteries, are harder to detect due to their relatively weak signal, resulting from slow or in-plane blood flow. An effective feature extraction method that can identify both large and small cerebral arteries, as well as provide quantitative cerebrovascular information on MRA is critically needed for complete feature extraction.

To date, there has been little effort to measure features from vascular maps beyond 3D visualization, vessel segmentation, and identification of stenosis or absent arteries (8–10). Some existing vascular feature extraction tools (11,12) are limited to four morphometric features from each of the four anatomic regions of the intracranial arteries. Wright et al (13) used L-Measure, an open source tool to quantify 19 morphometric features from six major arteries stemming from the circle of Willis (CoW). However, tracing more distal arteries and labeling arteries in more precise categories could facilitate more detailed analysis.

In this study, we aimed to develop a comprehensive semi-automated intracranial artery feature extraction technique using improved approaches for automated vessel tracing and labeling. Automated vessel tracing, which converts raw pixel or voxel content in planar or volumetric images to a topological and geometrical network with centerline and radius is the most important step for vascular feature extraction as demonstrated from applications of retinal fundus images (14,15) and computed tomography coronary angiography (16,17). Tracing using the Active Contour Model (Snake) is able to handle changes of topology and adapt locally to the shape of complex structures (18). The Open-Curve Active Contour model (19), as one of its improvements, driven by Gradient Vector Flow (GVF) and traces without initial boundary, has been successfully applied in tracing neural fibers (19–21) as well as tree-like tubular structures (22), including cerebral vessels from CT Angiography images (23). Recently, the application of a neuronal tree reconstruction method to MRA images has proved to be successful (13). With similar intention, the Open-Curve Active Contour model is likely to be suitable for MRA artery tracing after some modification and optimization, based on its adaptive stretching forces and smoothness preservation ability.

Moreover, the features provided by most existing 3D visualization and analysis tools (24,25) are limited to global geometric features such as length and volume, while regional features provide more direct information on of vascular diseases (26,27). In addition, flow dependent intensity features, which are often ignored in current segmentation algorithms, have potential value for the characterization of intracranial arteries. For example, high middle cerebral artery (MCA) velocity was found to be predictive of stroke in children with sickle cell disease (28).

Regional features are only available when all the arteries are correctly labeled. Therefore, we developed a new method for automated labeling to avoid labor intensive manual placement of branch points required by most current methods (11,12). While other automatic labeling methods exist, they cannot be easily applied to comprehensive intracranial artery labeling (29,30), as those models cannot differentiate enough artery types (>24 types) to describe a complete arterial tree.

Therefore, the goal of this study was to develop a semi-automated intraCranial artery feature extraction (iCafe) technique that can comprehensively quantify morphometric and intensity features of intracranial arteries on MRA images. Based on an automated approach to artery tracing and labeling, interactive human supervision can be performed to ensure accuracy.. Initial comparison with neuroradiologist review was performed to validate iCafe. By providing additional information on cerebrovascular structure and function beyond what is currently collected on MRA, iCafe presents a new opportunity to gain pathophysiological insights into the processes of vascular adaptation to aging and disease conditions in future population studies.

METHODS

MR Imaging

The algorithm was tested using eight prospectively recruited subjects (54±17 years, 5 males) with clinically documented intracranial arterial stenosis likely to be secondary to atherosclerosis on clinical CTA. Patients with clinically suspected Central Nervous System (CNS) vasculitis, intracranial dissection, previous cranial trauma, previous radiation, and reversible cerebral vasoconstriction syndrome or who underwent prior intracranial arterial interventions were excluded from the study. Patients were scanned on a 3T Philips Ingenia Scanner using a standard head coil. 3D TOF MRA imaging parameters were as follows: TR/TE = 14.7/3.5 ms, flip angle = 18°, in-plane resolution = 0.3 mm×0.3 mm, slice thickness = 1.4 mm, FOV = 190mm*190mm, matrix = 360*228, acquisition time = 116.79±21.57s. Local institutional review board approval was obtained for all study procedures and all subjects provided informed consent prior to enrollment.

Feature extraction

The Open-Curve Active Contour model was originally designed as a neural fiber tracing method. However, the neural fiber tracing method is not ideal for vessel tracing given the different nature and scale of vessels compared to neurons. Therefore, a series of modifications to the Open-Curve Active Contour model were applied in order to trace vessels: 1) re-slicing, intensity normalization and rudimentary segmentation before tracing, 2) a combined tracing approach with both original image and vesselness image, described in further detail below. In addition, arteries were labeled for feature calculation after tracing.

The iCafe feature extraction process can be divided into image preprocessing, rudimentary segmentation, vessel tracing, artery labeling, feature calculation and output. The workflow along with example images at each stage are shown in Fig 1.

Fig. 1.

Workflow for intraCranial artery feature extraction (iCafe)

Image preprocessing

MRA images were resliced using bi-cubic interpolation to achieve isotropic resolution in 3D space, as in most cases, the voxels were anisotropic. The absolute intensity values of MR images do not have a fixed meaning, so a fast and accurate intensity normalization method by Nyul (31) was used to adjust intensities among cases in the database. Based on the assumption that the histogram of each MRA image is similar, Intensities of Interest (IoI, ten percentiles as well as lower (0%) and higher (99.8%) percent intensity values) were trained to find parameters of standard scale. In the transformation stage, the IoI of each image was then mapped to the standard histogram by a lookup table (LUT). The original intensity of each voxel was then updated through the corresponding LUT.

Rudimentary segmentation

A clear separation between background and vessel pixels will enhance the gradient vector and benefit the detection of small arteries. For this reason, rudimentary segmentation of the artery region using intensity thresholding was applied before tracing. The Phansalkar local threshold method (32) was used due to its local thresholding ability with minimum loss of vascular regions in a reasonable processing time.

However, because the Phansalkar thresholding was not suitable for 3D visualization, the Renyi entropy threshold method (33), a method that can remove most background signal with the cost of some artery details, was also used. It should be noted that artery tracing was based on Phansalkar thresholded images to ensure capability of detecting small arteries, and traced results were reconstructed and superposed into Renyi thresholded images with clearer background for better visualization (Phansalkar thresholded image shown in Fig 1(b) and Renyi thresholded image shown in Fig 1(d)). With this approach, users can see an accurately traced arterial tree clearly.

Vessel tracing

We improved the Open-Curve Active Contour model (19) for a high-quality vessel tracing for MRA. With the original Open-Curve Active Contour model, it is likely that two vessels in close proximity would be traced as one, which is particularly problematic for parallel arteries such as the anterior cerebral arteries (ACA). As the Active Contour model (Snake) traces through iterations (18), the trace initialized by a seed point is stretched towards the direction of largest tubular probability calculated from gradient vector in trace end points. (19). This gradient based approach ensures elasticity and smoothness but is limited in its ability to differentiate parallel vessels since the intensity value between vessels is not low enough to form strong gradient vectors to confine traces to artery regions. Therefore, a combined tracing method using both raw and ‘vesselness images’ (Frangi vesselness filtered (34) image) was devised to improve tracing of parallel vessels which works as follows. Initially, the geometric center points of each connected region beyond a certain threshold in the vesselness image called seed points are found. In the first stage, the tracing process starts stretching from the seed point according to intensity value from the vesselness image. The current length of vessels is calculated in each iteration. If the length increases slower than 5 percent after an iteration, the trace is then stretched based on the raw image instead as a second stage until the length increase is below 5 percent after an iteration again, or a maximum iteration number is reached. The two stage process also overcomes a well-known disadvantage of the Frangi filter that vessels near an arterial bifurcation are weakened after filtering (35) as the locally tubular structure does not apply to the area of branching. The first stage of tracing ensures the accuracy in parallel vessels, and the second stage avoids the vessel being cut off near bifurcation points.

Manual editing

As a semi-automated technique, human supervision and correction are feasible when using iCafe to ensure the best quality of tracing. Three visualization methods (Maximum Intensity Projection (MIP) view, 3D view and Cross-section view) are provided for observing artery details, shown in Fig 2. Each trace in the MIP view or 3D view can be selected and edited. The operations available for users include the following: merging two traces together, splitting one trace into two, creating branch connections between traces, trimming end points of a trace, deleting traces, or removing outlier unconnected traces. Seed points can also be added manually from which additional arteries can be traced using the same automatic tracing method.

Fig. 2.

Manual editing is facilitated by three views (a) Maximum Intensity Projection (MIP) view with tracing results (green lines); (b) 3D view with tracing results (green lines); (c) Cross-section view

Artery labeling

Artery labeling and categorization are necessary after all vessels are traced in order to extract arterial segment-specific features. Therefore, a semi-automated labeling method was implemented. Twenty-four common artery types are assigned (definition of artery types in Supporting Table S1) based on 31 types of bifurcations of Interest (BoI) that are identified automatically. For those traces with “No label”, BoI types are excluded as they are possibly distal vessels beyond the named branches.

The identification of BoI type is based on a probability model using the positional, directional and topological features.

• Positional likelihood: Assumes BoI positions of each bifurcation type to follow Gaussian distribution, and calculates probability density function (pdf) of each bifurcation type from all the cases in the training set (previously processed data). Positional likelihood of the BoI for each bifurcation type was calculated from the corresponding pdf over the maximum probability in that function.

• Directional fitness: Averaged branch directions for each BoI type was calculated and normalized as model type based on previous labeled data. Each unlabeled BoI was then paired with an optimal model type by calculating and finding the maximum sum of inner product of all corresponding branch directions from each model type.

• Graph topological likelihood: a structural metric Network Structural Index (NSI) (36) of BoI, which is a weighted addition based on graph topological connection including node degree, radius, vesselness intensity as well as their nearby neighbor nodes. Large vessels with more connections to neighbor nodes had larger NSI. Similar to positional likelihood, the model assumes the NSI of each bifurcation type to follow a Gaussian distribution and calculates topological likelihood from the pdf.

Using Maximum a Posteriori (MAP) estimation, the possibilities of each bifurcation to be each BoI type were calculated. The labeling procedure was applied sequentially. Starting from the BoI with the highest probability and least ambiguity (largest difference between first and second possible types), BoI type was assigned in order. The prior probability of all neighbors was updated when a BoI type was assigned. Human inspection and correction were then performed to correct for errors.

Each artery could be labeled after all BoI types were assigned. Arteries with similar geometric and anatomical properties can be further grouped together based on anatomy and flow distribution. The 13 vascular groups are defined (definition of vascular group in Supporting Table S2). Some of the vascular groups also have sub-groups to differentiate left or right side features.

Feature calculation

Based on the labeled arteries, morphometric features and intensity features are calculated for each artery.

The twelve morphometric features are Length, Volume (assuming artery as a cylinder with changing radius), Tortuosity (length over head-tail distance), Surface area, Average Radius, Minimum Bifurcation angle, Cross-section area, Order (bifurcations towards ICA), Distance to COW, Width (minimum sagittal span), Depth (minimum coronal span), and Height (minimum vertical span).

The sixteen intensity features are Max, Min, Mean, Median, 10 percentiles (0%, 10%, …, 90%), Sigma, and Sum. The intensity values are sampled from centerline points along arteries from intensity normalized images for calculation.

With 12 morphometric features and 16 intensity features for each artery, grouped by each of the 24 artery types or 34 vascular groups (6 overlaps), there could be a total of (12+16)*(24+34−6)=1456 features extracted for each MRA.

Output and Analysis

Artery traces were saved in SWC format (37). Each line presents a point, which is comprised of Id, BoI type, X position, Y position, Z position, radius, and parent ID. iCafe is connected to MySQL database for data management and custom query. Features were arranged into five tables, recording case information, landmark coordinates, image intensity percentile for normalization (used for Nyul (31) normalization), vessel and point features. With user specified queries, any of the 1456 features for each case could be retrieved for analysis. Queries can be stored in the database to be applied to any set of data with ease. Advanced queries are also achievable through SQL scripts, such as calculating the difference of length in left and right anterior proximal arteries in subjects with a complete circle of Willis.

Validation

Accurate tracing results and features are important in the performance evaluation of this tool. However, validating the tracing results and features is a challenging task for human reviewers, as the complicated intracranial vascular structures in 3D space make it difficult for clinicians to segment and trace all the arteries through each slice of 2D images. Therefore, we used an indirect validation method with three steps to prove the accuracy of traces and features: 1) validate the 3D position of bifurcations identified as they determined the starting and ending points of most traces as well as the connections between them;2) validate the positions of stenosis of main arteries identified, as they evaluate the radius measurement of arteries; and 3) validate the length of labeled arteries (main arteries).

Validation of the 3D position of bifurcations

An experienced neuroradiologist marked the 3D position of bifurcations using 2D slices of the source MRA images and these were considered the reference standard. The bifurcation points to validate from one case were randomly chosen from all the bifurcations of the MCA, ACA or posterior cerebral arteries (PCA). All the bifurcations from the stemming bifurcation (bifurcation of ICA/MCA/ACA for MCA or ACA branches; bifurcation of VA/PCA for PCA branches) to the most distal branches were marked by the neuroradiologist. These positions were then compared with corresponding iCafe identified bifurcation positions. The bifurcation points that were marked by both the radiologist and detected by the algorithm were counted as agreement points, while those bifurcation points which were detected only by the algorithm (spurious) or only by the clinician (missed) were revisited and reviewed together by the same clinician and a separate iCafe operator to determine which were true bifurcations. In cases of non-resolution, a third rater was also consulted to adjudicate the results. Sensitivity and Positive Predictive Values (PPV) were calculated to evaluate the performance of bifurcation points found by the algorithm. Distance between the clinician and iCafe bifurcation positions was also calculated. The radius and order (level of bifurcations starting from ICA for MCA/ACA branches, VA for PCA branches) for agreed, missed and spurious branches were calculated to evaluate the location of agreed and disagreed bifurcations.

Validation of stenosis positions

Similar to the validation of bifurcation points, an experienced neuroradiologist independently marked the 3D position of stenosis using 2D slices of the original MRA images as ground truth while blinded to the iCafe analysis and patient clinical information. Due to the limitation of imaging resolution and quality, we only focused on the following large arteries or segments: 2xICA (carotid siphon to the bifurcation of ICA/M1/A1), 2xM1, 2xA1, 2xP1, BA, and 2xVA. Possible stenosis points located distal to bifurcation points were also excluded, as luminal measures change between parent and daughter arteries. Then the iCafe program generated all possible stenosis points based on the relative radius of arteries, with the calculation of the stenosis percentage based on the ratio of the radius at the maximum stenosis point over averaged radii in nearby normal artery points. We set the threshold of stenosis percentage of 10% to avoid the mild stenosis which neuroradiologists cannot detect confidently from 2D slices or may result from noise in radius estimation. The neuroradiologist-determined points of stenosis were compared to the iCafe generated stenosis points to determine the accuracy of the program.

Validation of artery length measurements

In the third step, for validation of features extracted from iCafe, length was selected for manual measurement. Vessel length was measured by manually selecting centerline points along the arteries. Experienced human reviewers measured M1 segments (centerlines of large arteries are easier to follow) using Philip Intellispace Portal Software (Philips Healthcare, Best, the Netherlands). After the centerline was drawn, length was measured on the corresponding curved planar reformat image and compared to length measurements using iCafe. The percentage of difference is calculated using equation [1] where N is the total number of measurements, m_1i and m_2i are the reviewer measured and iCafe measured length for the i th M1 segment.

$$\mathit{Diff}=\frac{1}{N}\sum _i^N\frac{\mid m_{1i}-m_{2i}\mid }{\frac{m_{1i}+m_{2i}}{2}}\ast 100\%$$ [1]

RESULTS

Rudimentary segmentation

To test the importance of applying the Phansalkar local threshold method, we compared the tracing results with and without using the threshold method before tracing, as shown in Fig 3. The positions marked by arrows in Fig 3(b) were verified by the clinician as missing traces or incorrect traces. It should be noted that some of the isolated arteries are also traced, but if a trace is labeled with “No label” in both ends, it did not affect feature extraction results.

Fig. 3.

(a) Maximum Intensity Projection (MIP) view for a TOF image; (b) Artery traces (green lines) using Open-Curve Active Contour model alone, some of the false traces marked in red arrows, some of the missing traces marked in white arrows; (c) Artery traces (green lines) using rudimentary segmentation before Open-Curve Active Contour model.

Improved tracing method

To test the effectiveness of the combined tracing method, we compared our tracing methods with tracing methods based only on the raw image and only on the vesselness image, shown in Fig 4. We found the tracing method based on the raw image alone led to two parallel arteries merged into one, and the tracing method based on the vesselness image alone led to dangling vessels. The red line is the erroneously traced artery. However, the combined method is able to depict arteries in parallel correctly for all four cases in this experiment.

Fig. 4.

(a) Tracing two parallel arteries directly in the native MRA image by the Open-Curve Active Contour model leads to the problem of two arteries merging into one; (b) Tracing in the vesselness image by the Open-Curve Active Contour model leads to dangling arteries and leaves some small arteries untraced; (c) Tracing by combining the vesselness and native images ensures arteries are correctly traced. Blue circles show the radius of each point.

Artery labeling

For artery labeling, the average BoI categorization accuracy by the prediction model was 98.51%. Manually changing the type of BoI allowed remaining inaccurate BoI types to be corrected. Each artery segment type was labeled with a color for better visualization in 3D view. An example is shown in Fig 5 with arteries labeled in color in MIP and 3D views.

Fig. 5.

Arteries labeled with their anatomical names shown with color lines in three directions of MIP (first row) and 3D view (second row)

Feature extraction

A total of 1456 features were extracted from each case, which were easily retrievable through the MySQL database. Examples of the key features include average radius of M1_L, distal branches length, of total valid arterial volume, average normalized intensity of proximal arteries, and average Circle of Willis segments tortuosity. Results of these features from eight cases can be found in Supporting Table S3.

2Processing time

iCafe was tested on an Intel^® Core^™ i5-4210H CPU with 2.9 GHz and 16 GB RAM. The processing time for preprocessing, automatic tracing and labeling was approximately 3 minutes. The human supervision and correction time varied by subject from 20 minutes to one hour.

Validation

Approximately 12 bifurcation points were marked on average by the human rater for each case. Among bifurcation points, the algorithm detected 127 points, and the human rater marked 99 points. Eighty-nine of them were agreed upon by the algorithm and human rater (having corresponding relations), the deviation distance of points was 0.328±0.444 millimeters, which is close to the image resolution of 0.297 millimeters per voxel. Thirty-eight of the bifurcation points detected by the algorithm were not marked by clinician, and 10 of the bifurcation points marked by the clinician were not detected by the algorithm. Nineteen of the 38 bifurcation points only detected by the algorithm were determined to be real or in other words, missed by the clinician in the first review. But the remaining 19 points were spurious points detected by the algorithm. For the 10 points that the algorithm failed to detect, it was concluded in the second review that 3 of them were incorrect and 7 were missed by algorithm. Based on this validation result, we could calculate the sensitivity of the algorithm for detecting bifurcations as (89+19)/(89+19+7)=94%, the PPV as (89+19)/(89+19+19)=85%, and the sensitivity of the human rater as (89+7)/(89+19+7)=83%, the PPV as (89+7)/(89+10)=97%.

The radius and order (level of bifurcations starting from ICA as 1^st order for anterior circulation branches and VA as 1^st order for posterior circulation branches with an addition of 1 order for each intervening segment prior to the targeted branch point) for agreed, missed and spurious branches is shown in Table 1. For missed bifurcations, five out of seven were tiny branches of lenticulostriate arteries near M1 branches, so their average radius was large and order was low. For spurious bifurcations, the smaller radius and higher order means they are most likely to be small branches in the distal area of the arterial tree.

Table 1.

Number, radius and order measurement for agreed, missed and spurious bifurcations (Average±standard deviation)

	Number	Radius of the bifurcation branch (millimeters)	Order of bifurcation branch
Agreed bifurcations	89	0.895±0.476	5.464±2.884
Missed bifurcations*	7	1.446±0.502	0.933±0.294
Spurious bifurcations**	19	0.545±0.200	5.895±2.075

^*Missed bifurcations: bifurcations identified by neuroradiologist (ground truth) but not by iCafe.

^**Spurious bifurcations: bifurcations identified by iCafe but not by neuroradiologist (ground truth).

The reviewer identified 13 stenosis positions among in the 8 cases. All of these stenosis positions were deemed by the algorithm to be somewhat stenotic (>0%). However, only 11 of them (sensitivity of 11/13=85%) had stenosis percentage higher than the pre-defined threshold. Meanwhile, there were two false positive positions (PPV of 11/13=85%) by the algorithm, which were considered by the reviewer as artefactual luminal narrowing due to flow dephasing resulting. We also compared average radius, vessel length, average normalized intensity, and tortuosity, which could be associated with atherosclerotic stenosis, between six stenotic artery segments and their contralateral non-stenotic counterparts (stenotic segments without contralateral counterparts were not included). The average radius on the diseased side was 0.21 millimeter less (p value 0.048 by paired t-test), and the average normalized intensity was only 81% of the non-stenosed side (p value 0.037 by paired t-test). Vessel length and tortuosity were not different between stenotic and non-stenotic side.

The reviewer-measured length of the M1 segment of the MCA was compared with the algorithm. The average difference between the two measures was 3.15%. The intra-class correlation coefficient of the two measurements was 0.97. The Bland Altman plot is shown in Fig 6.

Fig. 6.

Validation of length measurement of the M1 segment of the middle cerebral arteries

DISCUSSION

iCafe is able to trace and label intracranial arteries and extract quantitative features from widely available TOF MRA data. Therefore, this technique provides an important method for obtaining morphometric and signal intensity features of intracranial arteries globally and regionally. iCafe measurements include information on both large and small arteries and can be a supplement to perfusion imaging (38,39). With accurate regional features from precisely labeled artery groups, iCafe presents a quantitative tool for future research, such as examining the effects of aging and disease on intracranial cerebrovasculature.

The eight MRA images were obtained from prospective scans performed in patients with intracranial atherosclerosis, providing an equivalent imaging environment to vasculopathy patient populations. The performance of iCafe was evaluated on these more challenging diseased patients rather than healthy volunteers, proving its applicability on future analysis of clinical cases.

Algorithm

Other general-purpose automatic image segmentation or tracing tools like 3D Slicer (24) or ImageJ (40), are likely to fail in automatic processing under complicated situations, such as in regions with many intertwined distal MCA branches. These failures would also be difficult to identify by human supervision. iCafe is a human supervised semi-automated technique optimized for detecting and analyzing MRA. iCafe is implemented in a user-friendly graphical user interface where operators can monitor and supervise the artery tracing and labeling process through easy operations and three different visualization methods to ensure high accuracy of extracted features.

We chose the Open-Curve Active Contour model (19) for artery tracing for the following reasons: (1) as a deformable model, the Active Contour model (Snake) provides flexibility in tracing vessels with variable geometries; (2) the Open-Curve Active Contour model can be initialized automatically from a single seed point without a human specified contour; (3) there is precedent (13) that a neural tree reconstruction method could be adapted for MRA application; and (4) the Open-Curve Active Contour algorithm is open-sourced under Apache License, Version 2.0.

As the Open-Curve Active Contour model is based on gradient vector flow with low contrast MRA images, using the Open-Curve Active Contour model alone is inadequate to produce satisfactory results (shown in Error! Reference source not found.). Applying the Phansalkar local threshold method before tracing (in Error! Reference source not found. (c)) is effective in detecting smaller distal branches. The reason is that a clear background enhances gradient vectors and Frangi filtered images, which are critical in determining the seed points and to control stretching directions of traces in the model. In addition, the segmentation also roughly confines the artery traces within the foreground region, avoiding the traces leaking to the background.

Features

A large number of features can be extracted using iCafe based on artery types and features in contrast to previous methods. A detailed description of the artery tree and a comprehensive feature extraction may be helpful clinically and investigationally by identifying and characterizing pathology that may be missed with a limited number of features. In addition to common morphometric features, intensity features in each artery or vascular group are also extracted. As the intensity value is an indicator to blood flow, intensity normalization among all the data in the database may have potential value in analyzing these intensity features.

All the information of the vascular tree is stored in a well-organized database for easy retrieval. The features are not limited to 1456, as more customized features are available for specific tasks, such as, stenosis measurement, and vascular asymmetries between the left and right hemispheres.

Validation

Precise characterization of vascular anatomy is important for understanding cerebrovascular diseases (41).

Detailed human tracing of the intracranial arterial tree is difficult, which was the impetus for the development of iCafe. The challenges of tracing intracranial arteries lie in the precise identification of branching points. To validate a comprehensive intracranial artery map, previously used methods such as observing the tracings in a perpendicular view (13) are insufficient. In our study, bifurcation and stenosis points along arterial branches were marked and compared, which is more robust and allows tracing accuracy validation. Additionally, length measurements were compared with results from a commercially used measurement software. These validation experiments provide preliminary support that iCafe extracted features can be used in future analysis in large population-based studies.

The deviation between the expert-marked and iCafe-traced bifurcation points approximated the imaging resolution, indicating minimal difference. The high sensitivity and PPV indicate that bifurcation positions detected by iCafe are valid and accurate. We also found that most disagreements in the position of bifurcations were small branching arteries with relatively low signal.

It should be noted that the identification of bifurcation points is a challenging task, even for expert human reviewers. Missing and deviated bifurcation points arise for many reasons, such as low signal near branch ends caused by slow or in-plane flow, very small arteries located close to large arteries, and irregular branching patterns.

The validation of stenosis positions is aimed at proving the accuracy of radius measurements of traces and exploring its potential clinical uses. The high sensitivity and PPV of stenosis detection has shown that iCafe can accurately describe radius variations along artery segments. With further validation, iCafe could serve as an initial clinical evaluation tool to lessen the burden on radiologists in MRA.

Feature differences between diseased artery segments and the normal contralateral side indicate iCafe could potentially contribute biomarkers for vascular disease analysis in future research. More investigation is necessary however.

Due to the limitation of human measurement, only length of large arteries was assessed for feature validation. The difference between human-measured and iCafe-generated M1 segment length was low. Any difference in results may have occurred from the reconstruction when converting 3D images to 2D for manual measurement. In addition, the number of centerline points a human reviewer selected will also influence the measurement. Overall the validation showed a high level of agreement.

Limitations

While a large number of features can be derived using iCafe, only a limited set (bifurcation, stenosis positions and vessel length) were validated in the current study. This limitation is mainly due to the difficulty in obtaining human based tracing of the entire vasculature and the limitations of currently available measurement tools.

iCafe was implemented using C++ on the Windows platform only Since the underlying graphical user interface framework is cross-platform supported, iCafe can be further developed on other platforms to meet the needs of different clinical and research environments.

3D TOF MRA images were used to test iCafe, but we believe with some modifications, iCafe could be applied to other bright blood imaging techniques such as contrast-enhanced MRA, CT angiography or 3D digital subtraction angiography.

The processing time is between 20 minutes to an hour mostly due to time needed for human intervention and correction during the manual editing stage. A fully automated algorithm will help to reduce processing time, allowing for better clinical access. More powerful and accurate automated tracing and labeling methods are planned for future development.

CONCLUSION

A semi-automated quantitative measurement tool for intracranial arteries from 3D TOF MRA images has been developed and shown to provide accurate tracing of intracranial arteries. Through its ability to provide novel imaging biomarkers for cerebrovascular disease, it may facilitate new research and clinical approaches.

Supplementary Material

Supp info

Supporting Table S1: Intracranial artery segments classified in iCafe

Supporting Table S2: Definition of vascular groups

Supporting Table S3: Example features extracted from iCafe (Average±standard deviation of eight cases processed)