SEMI-AUTOMATIC CORONARY ARTERY CENTERLINE EXTRACTION IN COMPUTED TOMOGRAPHY ANGIOGRAPHY DATA Coert Metz1∗ , Michiel Schaap1∗, Alina van der Giessen2, Theo van Walsum1, Wiro Niessen1 1
Departments of Radiology and Medical Informatics 2 Department of Biomedical Engineering Erasmus MC - University Medical Center Rotterdam E-mail:
[email protected] * Both authors contributed equally to this paper
ABSTRACT This paper presents a semi-automatic coronary centerline extraction algorithm for computed tomography angiography data. The method applies region growing to computed tomography angiography data and incorporates bifurcation and leak detection. The presented method is evaluated either on the original data and on data in which vessel-like structures have been enhanced. Semiautomatically extracted centerlines of the three main coronary arteries are compared with centerlines manually annotated by three observers, using an overlap and distance measure. The method successfully extracted vessel centerlines in up to 15 out of 18 evaluated cases, with a localization accuracy which was in the range of the interobserver variability. Vessel enhancement prior to centerline extraction did not improve the results. Index Terms— coronary arteries, computed tomography angiography, vessel enhancement, centerline extraction, quantitative evaluation 1. INTRODUCTION Vessel centerline extraction is a useful image processing technique, which can be used for constructing application specific vessel visualization and as a first step towards quantitative vascular image analysis. The most notable example is vessel stenosis evaluation and grading in multiplanar reformatted images [1]. A number of authors have presented techniques for vessel axis tracking, see e.g. Aylward and Bullit [2] and Wink et al. [3], and have successfully applied these methods to carotic, iliac, pulmonary, brain, portal and hepatic vessels. Automatic tracking of the coronaries in 3D computed tomography angiography (CTA) has proven more challenging, owing to the limited image resolution relative to the size of the coronaries, the presence of pathologies, such as severe stenoses and calcifications, motion artifacts, and the proximity of other enhancing structures such as the aorta and the heart. In this paper we introduce a region growing based method for coronary vessel centerline extraction. The contribution of our paper is threefold. First, the region growing method includes branch and leak detection, in order to segment multiple coronary segments with a single seed point, and to limit leakage into the aorta and non-vascular structures. Second, it is investigated whether vessel enhancement prior to centerline extraction improves the results. Hereto, a recently introduced technique, coined vessel enhancing diffusion (VED) [4], is applied. To the best of our knowledge, VED has not been evaluated on its potential to improve the performance of vessel centerline extraction. Third, the
method is quantitatively evaluated by comparing semi-automatically extracted vessel centerlines with centerlines that were manually annotated by three observers using an overlap and distance measure. 2. METHOD 2.1. Vessel enhancing diffusion The vessel centerline extraction algorithm is applied both to the original image data, and to data preprocessed with VED. Manniesing et al. [4] propose a smoothed version V of the multiscale vessel filter defined by Frangi et al. [5] to tune a non-linear anisotropic diffusion scheme such that vascular structures are enhanced in the image. The diffusion tensor D, which is used in the diffusion equation ∂t u = ∇ · (D∇u) [6], is defined as D , QΛ′ QT
(1)
with Q the eigenvectors of the Hessian matrix H corresponding to the scale for which the vesselness response is maximal and Λ′ having the following functions on its diagonal λ′1 , 1 + (ω − 1) · V
1 s
(2) 1
λ′2 = λ′3 , 1 + (ǫ − 1) · V s .
(3)
Parameter ω indicates the strength of anisotropic diffusion, ǫ ensures the positive definiteness of the tensor and s denotes the sensitivity to the vesselness response. For non-vessel structures (V goes to zero) diffusion is high and isotropic, and background noise is reduced, whereas for vessel structures (V goes to one) diffusion is maximal (ω) along the vessel. Two examples of VED processed coronary CTA data are shown in Figure 1. 2.2. Vessel centerline extraction Centerline extraction is achieved via region growing based segmentation and subsequent path extraction. Starting from a user-defined seed point, region growing iteratively adds neighboring voxels within a pre-defined greyscale range to the segmentation. The algorithm keeps track of vessel bifurcations by performing connected component analysis after every two region grow iterations on the voxels added to the segmentation in these two iterations. If these voxels are not connected, a bifurcation has been detected, but if these voxels are connected, it is assumed that no vessel bifurcation is present (see Figure 2). This procedure results in a
(a)
(b)
(c)
(d) (a)
Fig. 1. Unprocessed coronary CTA data (a, c) and VED filtered equivalents (b, d).
(a)
(b)
Fig. 2. Bifurcation detection. Region growing is started from seed point s. Voxels added in the last two region grow iterations are marked with N, N1 and N2 . Voxels added in the two previous region grow iterations are marked with P. If the voxels in N are connected, it is assumed that the vessel does not bifurcate. In the case of a bifurcation, voxels in N1 and N2 are not connected.
tree representation of the segmented vasculature. To prevent leaking into the aorta or non-vascular structures during region growing, a threshold is imposed on the number of voxels that may be added to a vessel tree segment every two region grow iterations. When this threshold is exceeded, the corresponding segment is deleted from the segmentation. The coronary path is extracted by applying six sub-iteration thinning [7] to the segmented vessel tree. The user can subsequently interactively define vessel segments by selecting a second point in this thinned tree. The corresponding centerline is defined by the longest path with minimal Euclidain length through the second point and the thinning result, see Figure 3. Resulting centerlines are smoothed with a Gaussian kernel (σ = 1mm). 3. EXPERIMENTS 3.1. Data Twelve CTA data sets of patients who were referred for coronary intervention were randomly selected. Retrospectively ECG gated CTA scans were acquired in the Erasmus MC, Rotterdam with a Siemens R Sensation 64 scanner. Mean voxel sizes of the reconstructed CTA
(b)
Fig. 3. Path construction. A segmentation is constructed by region growing from a user-defined seedpoint s1 , light grey in (a). The skeleton of this segmentation is created by thinning, black in (a). Points in the thinned segmentation closest to the user-defined points (s1 , s2 ) are automatically selected and the centerline is defined by the longest possible path with minimal Euclidian length through these points and the thinned tree. In this example, the resulting path is path P from s1 to f , which is displayed in (b).
data are 0.3x0.3x0.4mm3 . Six of these data sets are used for optimization of the region growing parameters and six are used for the evaluation.
3.2. Evaluation method The semi-automatic centerline extraction algorithm is evaluated by comparing the extracted centerlines of the three main coronary arteries, namely the right coronary artery (RCA), left anterior descending (LAD) and left circumflex (LCX), with manually annotated centerlines. Furthermore, semi-automatically extracted centerlines are compared with centerlines resulting from the presented method without VED preprocessing. The vessel centerline for manual annotation is defined as the center of the lumen. Three trained observers annotated these centerlines in two steps. In the first step vessel points were annotated globally using axial, sagittal and coronal views of the data. In the second step the locations of these points were improved locally using multiplanar reformats orthogonal to the initial annotated path. In order to define the interobserver variability and the accuracy of the semi-automatic centerline extraction method, two distance measures are defined: the similarity index and the mean distance. First, a third order spline is fitted between each pair of neighboring points of the centerlines, which is subsequently sampled very densely (0.03mm). CTP ij is defined as the parts of centerline Ci that have corresponding points in centerline Cj , CFP ij as the parts of centerline Cj that do not have corresponding points in centerline Ci and CFN ij as the parts of centerline Ci that do not have overlap with centerline Cj (see Figure 4). The sum and mean of minimal Euclidian distances between cor¯ responding points on CTP ij are defined as sdij and did . The similarity index between centerlines Ci and Cj (SIij ) is used as a measure for
100
Fig. 4. Similarity index: the part of centerline (Ci ) that does not have overlap with centerline (Cj ) is marked as false negative (FN). The part in Ci that has overlap with Cj is marked as true positive (TP) and the part in Cj that has no overlap with centerline Ci is marked as false positive (FP).
Similarity index (%)
80
60
40
20
the amount of overlap SIij =
0
2|CTP ij | |CFP ij |
+
|CFN ij |
+
2|CTP ij |
× 100%.
(4)
The distance from a semi-automatically extracted centerline c to the manually annotated centerlines of the three observers is defined as P3 sdic (5) Dc = P3i TP . |C i ic | ¯ c for a semi-automatically extracted and the mean similarity index SI centerline c with respect to the manually annotated centerlines as the mean of the three similarity indices. The interobserver variability is expressed in the interobserver distance P3 P3 i j:j,i sdij (6) ID = P3 P3 TP i j:j,i |Cij | and the mean of the six similarity indices between the manually an¯ i ). notated centerlines of the three observers (SI 3.3. Parameters All VED parameters, except for the scale settings, were set to those reported by Manniesing et al. [4]. For the multiscale vessel response filter α = β = 0.5, λ = 5, c = 10−6 , σmin = 0.5 mm, σmax = 3.0 mm, with 5 different scales, exponentially distributed between σmin and σmax . Diffusion is performed with ω = 25 and s = 5.0 (Eq. 2 and 3), and time step ∆t = 10−3 . The leak detection parameter of the region grow procedure was set to 3000 voxels. This setting prevented leaking into the aorta and non-vascular structures. The threshold values (tmin , tmax ) for the region growing were determined by exhaustive parameter optimization. Centerlines for six training datasets, different from the datasets used during the evaluation phase, were manually annotated by two observers, each observer annotating centerlines for three datasets. The presented semi-automatic coronary centerline extraction algorithm was applied to these datasets using threshold values in the range [150-880] HU. Threshold values that maximized the similarity index were used for the evaluation experiment. Optimization was carried out separately for filtered and unfiltered data. The resulting lower bound values were 181 and 201 HU for respectively unprocessed and VED filtered data. The upper bound value was 876 HU for both unprocessed and filtered data. 4. RESULTS Figure 6 shows a typical example of manually annotated centerlines and the centerline resulting from our semi-automatic extraction.
Unfil. Fil. Int. LAD
Unfil. Fil. Int. LCX
Unfil. Fil. Int. RCA
Fig. 5. Mean similarity indices per subject for LAD, LCX and RCA on unfiltered and filtered data next to the interobserver variability. The six subjects are indicated by different symbols. Vessel LAD LCX RCA
Unfiltered ¯ c , Dc SI 73%, 0.48 79%, 0.49 78%, 0.40
VED filtered ¯ c , Dc SI 70%, 0.42 80%, 0.49 77%, 0.34
Interobserver ¯ i , ID SI 90%, 0.40 74%, 0.55 88%, 0.36
¯ c ) and mean distance (Dc in mm) beTable 1. Average overlap (SI tween successfully semi-automatically extracted and manually annotated centerlines for LAD, LCX and RCA.
Mean similarity indices per subject for semi-automatically extracted centerlines of the RCA, LCX and LAD in comparison with manual annotated centerlines are shown in Figure 5. For compari¯ i ) are also included in this son, the interobserver similarity indices (SI figure. According to a paired Student’s t-test the results of the presented method are not significantly different with or without VED preprocessing (P=0.29 for the similarity index and P=0.43 for the distances). Observations revealed that the semi-automatic centerline extraction algorithm found the proximal part of the vessels in 14 out of 18 evaluated cases when using VED as preprocessing step, and 15 out of 18 cases without VED filtering. The failed centerline extractions were caused by early leaking into parts of the heart due to the presence of stents or significant calcifications. Comparison with the ¯ values of these failed centerevaluation results showed that the SI ¯ > 50% on both unfillines are 50% or lower. Only subjects with SI tered and filtered data were taken into account in the average SI and mean distance values, which are shown in table 1. 5. DISCUSSION AND CONCLUSION A region growing based method for coronary artery vessel centerline extraction has been presented, which incorporates bifurcation and leak detection. Furthermore, the additional value of vessel enhancing diffusion as a preprocessing step has been evaluated. Vessel centerlines were successfully extracted in 14 or 15 out of 18 evaluated cases when respectively using vessel enhanced or unprocessed data. For these successfully extracted centerlines the over-
Attention should also be paid to the leak detection method which can fail for small leakages. In order to solve problems caused by contrast differences between different datasets we will investigate if the region growing step can be improved by taking, next to intensity information using fixed threshold values, first and second order differential structure of the image into account. In conclusion, the presented method is promising for automatic vessel analysis and VED preprocessing with parameter settings proposed by Manniesing et al. [4] does not improve the extraction results for this application. 6. REFERENCES [1] C. Caussin et al., “Comparison of coronary minimal lumen area quantification by sixty-four-slice computed tomography versus intravascular ultrasound for intermediate stenosis.,” Am. J. Cardiol., vol. 98, no. 7, pp. 871–876, 2006. [2] S.R. Aylward and E. Bullitt, “Initialization, noise, singularities, and scale in height ridge traversal for tubular object centerline extraction,” IEEE Trans. Med. Img., vol. 21, no. 2, pp. 61–75, 2002.
Fig. 6. Example of manually annotated (colors) and semiautomatically extracted (white) centerlines overlayed on a maximum intensity projection.
lap and mean distance between semi-automatically extracted and manually annotated centerlines is close to the interobserver variability. Furthermore, VED did not significantly change the overlap or distance between semi-automatically extracted and manually annotated centerlines. The additional failure when using VED happened in subject 5 and was caused by an early leak at the location of a stent in the image. Due to this stent the response of the vesselness filter is lower, leading to relatively more isotropic diffusion in the vessel neigborhood. The VED settings used were directly taken from the work of Manniesing et al. [4] and therefore might not be appropriate for coronary arteries. In future work we will investigate whether optimizing the VED parameters improves the performance of the extraction algorithm.
[3] O. Wink, W.J. Niessen, and M.A. Viergever, “Multiscale vessel tracking,” IEEE Trans. Med. Img., vol. 23, no. 1, pp. 130–133, 2004. [4] R. Manniesing, M.A. Viergever, and W.J. Niessen, “Vessel Enhancing Diffusion - A Scale Space Representation of Vessel Structures,” Med. Image Anal., 2006. [5] A.F. Frangi, W.J. Niessen, K.L. Vincken, and M.A. Viergever, “Muliscale Vessel Enhancement Filtering,” in Proc. of MICCAI, 1998, pp. 130–137. [6] J. Weickert, Anisotropic diffusion in image processing, Ph.D. thesis, Dept. of Mathematics, University of Kaiserslautern, Germany, 1996. [7] K. Palagyi and A. Kuba, “A 3d 6-subiteration thinning algorithm for extracting medial lines,” Pattern Recogn. Lett., vol. 19, no. 7, pp. 613–627, 1998.
