DOI: 10.1148/radiol.2473070436
(Radiology 2008;247:841-846.)
© RSNA, 2008
Cerebral Arteries: Fully Automated Segmentation from CT Angiography—A Feasibility Study1
Rashindra Manniesing, PhD,
Max A. Viergever, PhD,
Aad van der Lugt, MD, PhD, and
Wiro J. Niessen, PhD
1 From the Departments of Medical Informatics (R.M., W.J.N.) and Radiology (R.M., A.v.d.L., W.J.N.), Erasmus MC–University Medical Center Rotterdam, Dr. Molewaterplein 40/50, 3015 GE Rotterdam, the Netherlands; and Image Sciences Institute, University Medical Center Utrecht, Utrecht, the Netherlands (M.A.V.). From the 2006 RSNA Annual Meeting. Received March 15, 2007; revision requested May 23; revision received July 16; accepted August 16; final version accepted October 22.
Address correspondence to R.M. (e-mail: r.manniesing{at}erasmusmc.nl).
 |
ABSTRACT
|
|---|
The purpose of this study was to retrospectively assess the feasibility of a fully automated image postprocessing tool for the segmentation of the arterial cerebrovasculature from computed tomographic (CT) angiography in 27 patients (nine men, 18 women; mean age, 55 years; age range, 33–76 years) with subarachnoid hemorrhage. The institutional review board approved this study, and informed consent was waived. The proposed method, which does not require the acquisition of an additional CT scan for bone suppression, consists of the following: (a) automatic detection of the main arteries for initialization, (b) segmentation of these arteries through the skull base, and (c) suppression of the large veins near the skull. The parameters of this method were optimized on the training subset of nine patients, and the method was successful at segmentation of the arteries in 15 (83%) of the 18 remaining patients. The difference between automatic and manual diameter measurements was 0.0 mm ± 0.4 (standard deviation). The study results showed that fully automated segmentation of the cerebral arteries is feasible.
Supplemental material: http://radiology.rsnajnls.org/cgi/content/full/247/3/841/DC1
© RSNA, 2008
|
INTRODUCTION
|
|---|
Digital subtraction angiography has long been the reference standard for the examination of intracranial arteries. Because digital subtraction angiography is an invasive procedure associated with a small complication risk (1–4), it is gradually being replaced by noninvasive imaging techniques like computed tomographic (CT) angiography and magnetic resonance angiography. Multiple studies have compared CT angiography with digital subtraction angiography in the assessment of aneurysms, stenoses, and collateral arteries. Study results have demonstrated that CT angiography is superior to digital subtraction angiography in the depiction of vessel patency in the posterior circulation (5) and that CT angiography has a sensitivity similar to that of digital subtraction angiography in the depiction of large intracranial aneurysms (6–8). One study (9) showed encouraging results when CT angiography was preferred over digital subtraction angiography as the only modality for the diagnosis of cerebral aneurysms and pretreatment planning for patients with cerebral aneurysms.
It has been stated that CT angiography will become the new reference standard for vascular imaging in many applications (10). This poses new challenges for clinical imaging, because the sheer size of state-of-the-art CT angiographic data sets makes their analysis time consuming. For this reason, there is a strong interest in automatic and reproducible techniques for detection and quantification of vascular disease. A first step toward an effective vessel analysis tool is segmentation of the vasculature.
Automatic segmentation of the cerebral vessels is challenging for a number of reasons. First, it is difficult to achieve good segmentation results without manual initialization. Second, the close proximity of the internal carotid arteries (ICAs) and vertebral arteries to the bony skull base makes it difficult to separate these tissues. Third, leakage of the segmentation into the venous structures may occur.
For cerebral vasculature visualization and segmentation from CT angiography, bone masking prior to vessel segmentation is often used and has been shown to be successful (11–14). However, bone masking requires the acquisition of an additional CT scan, thereby increasing the radiation dose to the patient. Alternatives are methods that directly extract the vessel axis from CT angiographic data by using an automated vessel tracking procedure. While published studies (15,16) have addressed some issues, to our knowledge no integral solution has been reported for automatic cerebral artery segmentation of CT angiographic data. Thus, the purpose of our study was to retrospectively assess the feasibility of a fully automated image postprocessing tool for the segmentation of arterial cerebrovasculature from CT angiography.
|
MATERIALS AND METHODS
|
|---|
Patients and Scanning Protocol
The institutional review board of the University Medical Center Utrecht approved our retrospective study, and informed consent was waived. CT angiographic data from 27 patients (nine men, 18 women; mean age, 55 years; age range, 33–76 years) who were examined from February 2003 to March 2004 at the radiology department of the University Medical Center Utrecht were included. These patients had acute subarachnoid hemorrhage and underwent CT angiography as part of their diagnostic work-up. In a previous larger study (17), these patients were retrospectively selected on the basis of the following criteria: the presence of perimesencephalic nonaneurysmal hemorrhage or the presence of subarachnoid hemorrhage from a posterior circulation aneurysm and the quality of the CT angiogram. The quality was considered good if the CT angiogram had sufficient contrast and was without movement artifacts that could hamper the evaluation of the cerebral vessels. The data from 27 patients used in our study were collected from the larger study on three dates during the above-mentioned period. On these dates, all consecutive scans that were not yet archived to the storage system were included, except for two postoperative scans because of large deformations of the skull.
The data were acquired with a 16-section CT scanner (Mx8000 LDT; Philips Medical Systems, Best, the Netherlands) with the following parameters: 16 x 0.75-mm collimation; pitch, 0.3; section thickness, 1.0 mm; reconstruction interval, 0.5 mm; 120 kV; 200 mAs; field of view, 160 mm; and without angulation from the anthropologic baseline. Nonionic contrast agent (300 milligrams of iodine per milliliter, iopromide, Ultravist; Schering, Berlin, Germany) was injected into the antecubital vein: 50 mL at a rate of 5 mL/sec and 20 mL at a rate of 3 mL/sec. A 30-mL saline flush was administered at a rate of 3 mL/sec. The total scanning time was approximately 2 minutes. More details of the applied selection criteria and the scanning protocol can be found in van der Schaaf et al (17).
Segmentation Method
Our proposed segmentation method consists of five steps. First, several regions of interest are automatically defined on the basis of anatomic information (the skull, the skull base, and the ICAs) contained in the patient data set. This is required because, depending on the spatial locations of the region of interest, different image processing techniques will be applied. To define the regions of interest, we use the concept of entropy profile (Fig 1). On the basis of the classic Shannon entropy measure (18), we plot the entropy value of each section as a function of the axial section number. The entropy is a measure of the amount of information that is contained in the intensity histogram of a section. The section with maximum entropy is used as an anchor point for the definition of other regions. In results of experiments, it has been found that this section consistently corresponds with an axial section through the skull base.
View larger version (128K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1a: (a) Sagittal view of a volume-rendered cerebral CT angiographic data set image and (b) graph of entropy as function of section (slice) coordinate z, with z being zero denoting the caudal side of the head. The section with maximum entropy (Zmax) corresponds with an axial section through the skull base.
|
|
View larger version (16K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1b: (a) Sagittal view of a volume-rendered cerebral CT angiographic data set image and (b) graph of entropy as function of section (slice) coordinate z, with z being zero denoting the caudal side of the head. The section with maximum entropy (Zmax) corresponds with an axial section through the skull base.
|
|
Second, all nonarterial structures are suppressed as much as possible, without affecting the arterial vasculature. These nonarterial structures include background, air, bone, and veins close to the skull and are suppressed by combining intensity information, spatial information, and morphologic operations. Third, seed points are placed in the feeding arteries of the circle of Willis (ie, the ICAs and the basilar artery) by searching for circular structures in cross-sectional sections by using applications of the Hough transform (19). The Hough transform uses a parametric description of a circle to map the original image to a feature space in which maxima, representing the circles in the original image, are searched. By using this transform, noncircular structures can be effectively suppressed. The center points of the circles are used as seed points to initialize the segmentation process.
Fourth, a rough segmentation of the arterial vasculature is obtained by combining segmentations of the ICAs and the intracranial vasculature. The segmentation of the ICAs through the region of the skull base is challenging owing to close proximity of bone. Segmentation in this region is achieved by using a so-called level-set evolution (20)—which is a three-dimensional surface that evolves into an image to capture the boundaries of interest—extended by using a tubular shape constraint on its skeleton representation. In results of previous work (21), this method was a viable solution to the segmentation problem of the ICAs in this anatomic region. Fifth, the obtained rough segmentation of the vasculature is used as initialization of another level-set evolution to achieve accurate vessel boundary localization. The segmentation method was implemented in C++ with a 2.4-GHz Linux computer system, without code optimization. The average computation time for segmentation was 292 minutes ± 85 (standard deviation), with most of the time spent on segmentation of the ICAs (232 minutes ± 86).
A complete technical description of the method is provided in Appendix E1 (http://radiology.rsnajnls.org/cgi/content/full/247/3/841/DC1).
Qualitative Analysis
The method was qualitatively evaluated with visual inspection of the segmentation results. To this end, we divided the total patient data into two subsets: one subset of nine patients (the training set) on which the parameters of the method were optimized and one subset of the remaining 18 patients (the test set) on which the method (with the optimized parameters) was evaluated. Optimization was performed by setting a range of values for the most influential parameters of the method; these parameters are described in Appendix E1 (http://radiology.rsnajnls.org/cgi/content/full/247/3/841/DC1). The evaluation criterion for both subsets was the number of successful segmentations. This was determined with visual inspection by one observer (R.M., with 4 years of CT angiographic experience) with respect to the number of correctly found sections containing the skull base, the number of correctly found seed points, the inclusion of the skull base, and the suppression of the veins located close to the skull. Segmentation was classified as a full success if no parts of the skull base were included and if the veins were suppressed at the segmentation.
Quantitative Analysis
The method was quantitatively evaluated with vessel diameter measurement. We choose to validate the method by using diameter measurement because it indicates the accuracy with which the method has captured the vessel boundaries, and it provides a basis for quantification of pathologic disease, such as stenoses or aneurysms, in later studies. We selected data from eight patients from the training set that were used in the qualitative analysis. One patient was excluded from this training set because the segmentation was not successful. The data from eight patients were then divided into two equally sized subsets: the training set and the test set. Parameters were selected (Appendix E1, http://radiology.rsnajnls.org/cgi/content/full/247/3/841/DC1) that minimized the difference between the measurement of the observer and that of the method.
Automatic measurements were compared with manual measurements performed by one experienced observer on a clinical workstation (MxView; Philips) of the larger arterial vessels forming the circle of Willis. The following vessels were included: the anterior cerebral arteries (A1 and A2 segments), the middle cerebral arteries (M1 and M2 segments), the posterior communicating arteries, and the posterior cerebral arteries (P1 and P2 segments), which gave 14 points for each patient. The procedure for manual measurement was as follows: First, the window width and window level settings were fixed at 150 and 400 HU, respectively. Several slabs that included all the vessels of interest were then selected. The observer drew a line approximately perpendicular to the central vessel axis, and the workstation then automatically provided the length of this line. The procedure for the automatic diameter measurement was as follows: The same observer was asked to place a seed point at the location where the manual measurement had been performed. At this location, the automatic segmentation of the method was reformatted along the central vessel line, and the diameter was determined by the average diameter, with the assumption of a cylindric shape. Further details of both procedures are available (22).
Statistical Analysis
For the qualitative analysis, a 95% confidence interval of the success rate was provided by using a binomial distribution. For the quantitative analysis, the measurement by using the method and that of the observer were compared with a Bland-Altman analysis (23). In addition, a paired t test was performed on the two series of measurements. The P value was two-tailed. A P value less than .05 was considered to indicate a significant difference. The statistical analysis was performed by using a software package (SPSS, version 11.0.1 for Windows; SPSS, Chicago, Ill).
|
RESULTS
|
|---|
Qualitative Evaluation
For the test set of data from 18 patients, the section containing the skull base, selected by using the proposed entropy measure, was found in all 18 patients. Seed point detection failed once for the ICAs and four times for the basilar artery, which amounts to a success rate of 91% (49 of 54). Failure was mainly caused by the absence of a circular shape of the vessel cross sections. The failures were manually corrected in the axial sections. In three patients, the segmentation contained parts of the skull base, owing to a failure of segmentation in one or both ICAs. However, these segmentation results were still useful for improved visualization (Fig 2). None of the segmentations contained the veins near the bony skull. The straight sinus, vein of Galen, internal cerebral veins, inferior sagittal sinus, and basal veins of Rosenthal have no bone tissue around them; thus, our vein suppression approach will not work for them. Overall, the vessel segmentation method had a success rate of 83% (15 of 18 patients) (95% confidence interval: 59%, 96%) (Fig 3). A rotating view of a three-dimensional surface rendering of cerebral arteries from one of the data sets can be viewed (Movie, http://radiology.rsnajnls.org/cgi/content/full/247/3/841/DC1).
View larger version (61K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3: Axial maximum intensity projections (left column) and axial (second to left column), coronal (second to right column), and sagittal (right column) isosurface renderings in five patients from the test set of 18 patients. Arrow = straight sinus, BA = basilar artery.
|
|
Quantitative Evaluation
The diameter measurements for the test set of four patients showed a difference between the observer and the automated method of 0.0 mm ± 0.4 (95% confidence interval: –0.9, 0.8) (Fig 4). The paired t test results showed that the accuracy of the method was comparable to the observers. (The hypothesis that the accuracy of the methods was different was rejected with a P value of .59.)
View larger version (9K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 4: Bland-Altman plot of diameter measurements for test set of four patients. The difference between segmentation method (M) and observer (O) was 0.0 mm ± 0.4. SD = standard deviation.
|
|
|
DISCUSSION
|
|---|
To our knowledge, we have presented the first integral method for fully automated segmentation of the complete arterial vasculature from CT angiography. In contrast to some methods for vessel segmentation at CT angiography (12,22), our method does not require an additional scan for bone masking and thus avoids additional radiation exposure to the patient.
Masking, based on intensity information combined with morphologic operations, generally worked well. Vein masking, which was restricted to the largest veins near the skull, substantially improved visualization of the arterial vasculature. Suppression of the straight sinus, vein of Galen, internal cerebral veins, inferior sagittal sinus, and basal veins of Rosenthal at the postprocessing stage will require development and evaluation of additional image processing techniques.
The use of entropy for the characterization of the location of anatomic structures in CT angiographic data, and in particular for selection of the region containing the skull base, was a straightforward method. For all data sets, the correct region of interest containing the skull base was found. On the basis of the entropy profile, cross-sectional sections were selected for automated detection of the ICAs and the basilar artery. Ninety-one (49 of 54) percent of the arteries in the selected cross sections were correctly identified. Failure to detect the arteries was caused by the absence of a circular shape of the arteries.
If the arteries were not found in the selected cross sections, which happened in five of 54 carotid arteries, a minimum user interaction was required to select the correct vessel center in the axial sections. In a clinical setup, only the axial sections need to be shown for manual evaluation and possible correction. Automatic detection of the arteries may be improved by using majority voting of seed point detection in multiple axial sections or by including spatial information.
Segmentation of the ICAs through the skull base was achieved by using a shape-constrained level set. An important aspect of this method is that the cross-sectional shape does not need to be circular or ellipsoidal; therefore, vessel tracking is less hampered by pathologic conditions. However, in some patients, tracking of the ICAs was not able to proceed beyond the skull base or leaked into background, owing to the strong curvature of vessel in this region and strong local intensity variation. This step may be improved by designing a more advanced speed function, which dynamically updates itself based on the current vessel information and by prior filtering of the vessel structures.
The final step of the method is a level set–based segmentation aimed at accurate boundary localization. A level-set framework was selected because it provides a continuous representation of the evolving surface which can be controlled by imposing curvature constraints. Results of a previous study (22) showed that proper weighting of internal and external constraints can be used to achieve high accuracy in diameter estimation. In most of our patients (15 of 18), the complete arterial vasculature was successfully segmented without inclusion of the bony skull base. Successive vessel segmentation by using region growing with decreasing intensity bounds allowed inclusion of even small intracranial vessels.
Our study had limitations. One limitation was the required computation time of approximately 5 hours to obtain a complete segmentation. Therefore, the segmentation was not readily available after scanning. However, the primary focus of our study was to show the feasibility of a fully automated segmentation; therefore, no special effort has been made in reducing the computation time. By revisiting the different steps of the method, the computation time can be reduced to be in the order of minutes. In particular, the general level set–based method for segmentation of the ICAs can be replaced by a much faster region-growing process by optimizing the source code and by using state-of-the art workstations possibly equipped with dedicated hardware. A second limitation of our method was that veins not surrounded by bone were not suppressed.
In addition, in a small number of cases, the initialization of our method failed. In these circumstances, manual intervention was required. We have focused on evaluation of the feasibility of automated segmentation. A more extensive evaluation is desirable, with a strong focus on the diagnostic value of the segmentation results (eg, by comparing these results to the original CT angiographic data sets and/or to digital subtraction angiographic data and by addressing the question of whether the detection of pathologic conditions can be improved).
To conclude, we have shown that segmentation of the cerebral arteries in CT angiographic data is feasible without human interaction and without the acquisition of an additional CT scan for bone suppression. We believe that the proposed segmentation method may provide a basis for the development of advanced clinical applications, which includes detection, characterization, and quantification of plaques, calcifications, stenoses, aneurysms, and vasospasm.
|
ADVANCE IN KNOWLEDGE
|
|---|
- Fully automated segmentation of the cerebral arteries from CT angiography is feasible without an additional CT scan for bone suppression.
|
IMPLICATION FOR PATIENT CARE
|
|---|
- The segmentation method does not require an additional CT scan for bone suppression; this will reduce scanning time and radiation dose.
|
FOOTNOTES
|
|---|
Abbreviations: ICA = internal carotid artery
Author contributions: Guarantors of integrity of entire study, R.M., W.J.N.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; manuscript final version approval, all authors; literature research, R.M.; clinical studies, R.M.; statistical analysis, R.M.; and manuscript editing, all authors
Authors stated no financial relationship to disclose.
|
References
|
|---|
- Waugh JR, Sacharias N. Arteriographic complications in the DSA era. Radiology 1992;182(1):243–246.[Abstract/Free Full Text]
- Warnock NG, Gandhi MR, Bergvall U, Powell T. Complications of intraarterial digital subtraction angiography in patients investigated for cerebral vascular disease. Br J Radiol 1993;66(790):855–858.[Abstract/Free Full Text]
- Heiserman JE, Dean BL, Hodak JA, et al. Neurologic complications of cerebral angiography. AJNR Am J Neuroradiol 1994;15(8):1401–1407.[Abstract]
- Pryor JC, Setton A, Nelson PK, Berenstein A. Complications of diagnostic cerebral angiography and tips on avoidance. Neuroimaging Clin N Am 1996;6(3):751–758.[Medline]
- Bash S, Villablanca JP, Jahan R, et al. Intracranial vascular stenosis and occlusive disease: evaluation with CT angiography, MR angiography, and digital subtraction angiography. AJNR Am J Neuroradiol 2005;26(5):1012–1021.[Abstract/Free Full Text]
- Kouskouras C, Charitanti A, Giavroglou C, et al. Intracranial aneurysms: evaluation using CTA and MRA—correlation with DSA and intraoperative findings. Neuroradiology 2004;46(10):842–850.[CrossRef][Medline]
- Karamessini MT, Kagadis GC, Petsas T, et al. CT angiography with three-dimensional techniques for the early diagnosis of intracranial aneurysms: comparison with intra-arterial DSA and the surgical findings. Eur J Radiol 2004;49(3):212–223.[CrossRef][Medline]
- Dammert S, Krings T, Moller-Hartmann W, et al. Detection of intracranial aneurysms with multislice CT: comparison with conventional angiography. Neuroradiology 2004;46(6):427–434.[Medline]
- Hoh BL, Cheung AC, Rabinov JD, Pryor JC, Carter BS, Ogilvy CS. Results of a prospective protocol of computer tomography angiography in place of catheter angiography as the only diagnostic and pretreatment planning study for cerebral aneurysms by a combined neurovascular team. Neurosurgery 2004;54(6):1329–1342.[CrossRef][Medline]
- Prokop M. Multislice CT angiography. Eur J Radiol 2000;36(2):86–96.[CrossRef][Medline]
- Venema HW, Hulsmans FJ, den Heeten GJ. CT angiography of the circle of Willis and intracranial internal carotid arteries: maximum intensity projection with matched mask bone elimination—feasibility study. Radiology 2001;218(3):893–898.[Abstract/Free Full Text]
- Kwon SM, Kim YS, Kim T, Ra JB. Digital subtraction CT angiography based on efficient 3D registration and refinement. Comput Med Imaging Graph 2004;28(7):391–400.[CrossRef][Medline]
- Tomandl BF, Hammen T, Klotz E, Ditt H, Stemper B, Lell M. Bone-subtraction CT angiography for the evaluation of intracranial aneurysms. AJNR Am J Neuroradiol 2006;27(1):55–59.[Abstract/Free Full Text]
- Lell M, Anders K, Klotz E, Ditt H, Bautz W, Tomandl BF. Clinical evaluation of bone-subtraction CT angiography (BSCTA) in head and neck imaging. Eur Radiol 2006;16(4):889–897.[CrossRef][Medline]
- Suryanarayanan S, Mullick R, Mallya Y, Kamath V, Nagaraj N. Automatic partitioning of head CTA for enabling segmentation. In: Fitzpatrick JM, Sonka M, eds. Proceedings of SPIE: medical imaging 2004—image processing. Vol 5370. Bellingham, Wash: International Society for Optical Engineering, 2004; 410–419.
- Shim H, Yun ID, Lee KM, Lee SU. Partition-based extraction of cerebral arteries from CT angiography with emphasis on adaptive tracking. Inf Process Med Imaging 2005;19:357–368.[Medline]
- van der Schaaf IC, Velthuis BK, Gouw A, Rinkel GJ. Venous drainage in perimesencephalic hemorrhage. Stroke 2004;35(7):1614–1618.[Abstract/Free Full Text]
- Shannon CE. A mathematical theory of communication. Bell Syst Tech J 1948;27:379–423; 623–656.
- Hough PVC. Methods and means for recognizing complex patterns. U.S. patent 3,069,654. 1962.
- Sethian JA. Level set methods and fast marching methods. 2nd ed. Cambridge, England: Cambridge University Press, 1999.
- Manniesing R, Viergever MA, Niessen WJ. Vessel axis tracking using topology constrained surface evolution. IEEE Trans Med Imaging 2007;26(3):309–316.[CrossRef][Medline]
- Manniesing R, Velthuis BK, van Leeuwen MS, van der Schaaf IC, van Laar PJ, Niessen WJ. Level set based cerebral vasculature segmentation and diameter quantification in CT angiography. Med Image Anal 2006;10(2):200–214.[CrossRef][Medline]
- Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986;1(8476):307–310.[CrossRef][Medline]
TOP
ABSTRACT
INTRODUCTION
