Automated Segmentation and Anatomical Labeling of Abdominal Arteries Based on Multi-organ Segmentation from Contrast-Enhanced CT Data Yuki Suzuki1,2 , Toshiyuki Okada2 , Masatoshi Hori2 , Futoshi Yokota3 , Marius George Linguraru4 , Noriyuki Tomiyama2 , and Yoshinobu Sato2,1 1

Graduate School of Information Science and Technology, Japan 2 Graduate School of Medicine, Osaka University, Japan [email protected] 3 Graduate School of Engineering, Kobe University, Japan 4 Sheikh Zayed Institute for Pediatric Surgical Innovation, Children’s National Medical Center, Washington DC, USA

Abstract. A fully automated method is described for segmentation and anatomical labeling of the abdominal arteries from contrast-enhanced CT data of the upper abdomen. By assuming that the regions of the organs and aorta have already been automatically segmented, the problem is formulated as extracting and selecting the optimal paths between the organ and aorta regions based on a basic anatomical constraint that arteries supplying blood to an organ consist of tree structures whose root nodes are located in the aorta region and leaf nodes in the organ region. Using the constraint, the proposed method solves both of artery segmentation and anatomical labeling. In addition, the method is robust against topological variability of the branching patterns. Experimental results using 10 datasets demonstrate that the proposed method was effectively applied to several kinds of the abdominal arteries, which include the hepatic, splenic, and renal arteries. The average F -measure, which is a normalized accuracy measure taking both false positives and true negatives into account, was 0.89 for the proposed and 0.74 for the previous methods. The method could also effectively deal with topological variability of the hepatic and renal arteries.

1

Introduction

Segmentation and anatomical labeling of the abdominal vessels from 3D data are key issues for computer-aided diagnosis and therapy planning. Especially in organ transplantation and catheterization, understanding the artery branches of the target organ is essential. In previous works, machine learning and statistical classification techniques were effectively utilized for anatomical labeling of segmented vascular trees and other branching structures such as the bronchi [1–3]. However, two problems were not sufficiently addressed in the previous K. Drechsler et al. (Eds.): CLIP 2012, LNCS 7761, pp. 67–74, 2013. c Springer-Verlag Berlin Heidelberg 2013 

68

Y. Suzuki et al.

works. Firstly, the previous methods assumed that vessel regions have already been segmented from 3D images although vessel segmentation is still a nontrivial task from the contrast-enhanced abdominal CT images. Secondly, the inter-patient topological variability in the branching patterns of the abdominal arteries is highly large compared with other parts of the body. In the previous methods, there were limitations on dealing with the topological variability in the branching patterns. In this paper, we address the above two problems, that is, (1) solving both of segmentation and anatomical labeling and (2) dealing with the topological variability. The main assumption of the proposed method is that the regions of the aorta and target organs have already been segmented from CT data. Due to recent advancement of abdominal multi-organ segmentation [4–7], the assumption is now practical. Based on the assumption, we utilize a basic anatomical constraint that the artery branches which supply blood to a target abdominal organ are imaged as a set of curvilinear structures connecting the organ region and aorta region. Using the constraint, we formulate a method of solving both of vessel segmentation and anatomical labeling, which is robust against the topological variability, from the abdominal contrast-enhanced CT data. While the proposed method is an extension of our previous method developed under the assumption that vascular trees have already been segmented [8], we deal with fragmented and over-extracted vessel candidate regions obtained by typical lowlevel processing so as to address segmentation as well as anatomical labeling, and compare the proposed method with our previous one.

2 2.1

Methods Assumption, Constraints, and Formulation

Our main assumption is that the abdominal organ and aorta regions have already been automatically segmented from the CT images. The initialization of our method is based on the recent development for fully automated CT segmentation of the aorta and the abdominal organs [7]. A typical segmentation result is shown in Fig. 1(a). Among these organs, we have found that the segmentation of the liver, spleen, and kidneys as well as aorta is now sufficiently stable and accurate. In our latest experiments, the average Jaccard index of these organs was 0.85 or higher using cross validation of 87 cases obtained at different hospitals and imaging protocols. Therefore, the assumption that the regions of the aorta and target organs have already been segmented is now practical for the artery branches. We use the basic constraint that the artery branches for a target organ are imaged as a set of curvilinear structures connecting the organ region and aorta region. Further, we combine an additional constraint that the artery branches should form a set of tree structures whose root nodes are located in the aorta region and leaf nodes in the target organ region. Based on the constraints, we formulate the problem of segmentation and anatomical labeling of the artery branches as extracting the minimum-cost paths between all the possible pairs

Automated Segmentation and Anatomical Labeling of Abdominal Arteries

(a)

(b)

69

(c)

Fig. 1. Typical input datasets and intermediate results of the proposed methods. (a) Automatically segmented organ and aorta regions from CT data; the image shows the liver in brown, kidneys in pink, spleen in purple, and aorta in red. (b) Candidate regions of vessels (blue) for hepatic arteries obtained using the liver and aorta regions. (c) Candidate root (yellow) and leaf (green) nodes for hepatic arteries. Regarding the root node candidates of this case, only the most upper root node was selected by the processes of the optimal path extraction and selection described in 2.4

of candidate source points (root nodes) in the aorta region and goal points (leaf nodes) in the target organ region, and then selecting the paths satisfying a necessary condition for tree structures among all the extracted paths. Given the above formulation, the remaining problems are (1) localization of the candidate root and leaf nodes in the aorta and organ regions, respectively, (2) definition of a cost function and algorithms for the minimum-cost path extraction and selection, and (3) anatomical interpretation of the selected paths. In the following, their specific implementations which we used in the experiments are described. 2.2

Extracting Candidate Regions of Arteries

We firstly extract candidate regions of arteries, which may be fragmented and over-extracted, to localize the candidate root and leaf nodes as well as define the cost function. The candidate regions are independently extracted for each pair of the aorta and organ. That is, different candidate regions, which may overlap each other, are extracted for different artery branches, for example, the hepatic artery (liver and aorta), left renal artery (left kidney and aorta), and so on. The details of the method for candidate region extraction are described below. In this work, multi-phase (such as arterial and portal venous phases) intensity information obtained by contrast-enhanced dynamic CT scan is used to enable easier discrimination of artery branches. One phase of the original image and one subtraction image between two phases were used. Our contrast-enhanced CT data consisted of three phases, early and late arterial, and portal phases. We selected the early arterial phase for the original image and the portal phase subtracted from the early arterial phase for the subtraction image.

70

Y. Suzuki et al.

By assuming that the intensities of the artery branches are similar to those of the aorta and aiming to artery region extraction even within the (contrastenhanced) organ region, the thresholds for the original and subtraction images were automatically determined based on discriminant analysis [9] for the union of the aorta and organ regions, and then the determined thresholds were used to binarize the entire CT images. In addition, multiscale vessel enhancement filtering [10] is applied to the selected original image. The threshold for vessel-enhanced images was manually adjusted, but it was fixed for all the cases. The candidate regions of arteries were defined as the intersection of the binary images of the original, subtraction, and line enhancement images. Fig. 1(b) shows candidate regions for the hepatic arteries which are considered for subsequent optimization. As described in 2.4, extended regions which are within n-voxel distance from the candidate regions are considered for potential (true) artery regions in the cost function (Eq. (1)). 2.3

Localizing Candidate Root and Leaf Nodes

To localize candidate root and leaf nodes, the candidate regions of arteries are skeletonized to extract their centerlines. The terminal points of the centerlines inside or within n-voxel distance from the aorta region are detected as the root node candidates. The leaf node candidates are the terminal points inside the target organ regions. Fig. 1(c) shows detected root and leaf node candidates. 2.4

Optimal Path Extraction and Selection

In order to define the cost function for the minimum-cost path extraction, we first obtain signed distance transform d(x) of the candidate regions of the vessels, which has negative distance inside the candidate regions and positive outside. By using d(x), we define the cost of position x, fcost (x), as ⎧ d(x) ≤ 0 ⎨ d(x) − dmin fcost (x) = k(d(x) − dmin ) 0 < d(x) ≤ n (1) ⎩ ∞ otherwise where k is an additional cost to fill a one-voxel gap between fragmented candidate regions, n is the maximum distance to be considered for filling gaps, and dmin is the minimum distance value in d(x). dmin is subtracted to make fcost nonnegative. Note that at maximum 2n-voxel gaps are potentially filled because n-voxel distance is considered from both fragmented candidate regions. The cost for the path, x1 , x2 , · · ·, xN , where xi and xi+1 are 26-neighborhood each other, is given by N −1  i=1

D(xi , xi+1 )fcost (xi )

(2)

Automated Segmentation and Anatomical Labeling of Abdominal Arteries

71

where D(xi , xi+1 ) is Euclidean distance between xi and xi+1 . The cost is designed so that the low-cost path goes along the centerline where low cost should be assigned in the signed distance transform. Given one root node candidate as the source point and one leaf node candidate as the goal point, the minimum-cost path between them is found by Dijkstra’s algorithm [11]. The minimum-cost paths are extracted for all the possible pairs of root and leaf node candidates. From each leaf node candidate, the paths are extracted to all the root node candidates. Among them, the path having the minimum cost is selected. That is, the unique path is selected for each leaf node, which also determines the uniquely corresponding root node. 2.5

Anatomical Interpretation for Selected Paths

For each pair of the target organ and aorta, the selected paths are labeled. If the target organ is the liver, the paths are basically “hepatic arteries” and typically consist of the celiac artery, the common and proper hepatic arteries, the left and right hepatic arteries, and their peripheral branches. The paths extracted using the spleen are basically “splenic arteries” and typically consist of the celiac and splenic arteries. Thus, the celiac artery is labeled for the shared potion of the above two arteries. The selected paths regarding the kidneys are naturally labeled as the renal arteries. The paths having the same label typically form one or more tree structures. Especially, it is not rare that the hepatic and renal arteries have irregular topology and form two or more tree structures. In such cases, our labeling scheme is not affected by the irregular topology as long as the paths are extracted using the organ and aorta regions.

3

Results

We evaluated the method using ten randomly selected cases of contrast-enhanced CT data of the upper abdomen acquired at Osaka University Hospital, which consisted of three phases, early and late arterial, and portal phases. The method was fully automated. The two phases of 3D datasets were automatically registered [12]. Automated segmentation of the liver, spleen, kidneys, and aorta was performed from the late arterial phase of CT data [7]. Fig. 2 shows a typical successful case. Fig. 3 shows illustrative results regarding irregular topology. In Fig. 3(a), one hepatic artery branch was directly bifurcated from the aorta without merging into the celiac artery. In Fig. 3(b), the right renal arteries consisted of two independent branches bifurcated directly from the aorta. The proposed method could effectively deal with the variability in branching topology. We evaluated the segmentation and labeling accuracy by the recall (sensitivity) R defined as R = TP /(TP + FN ) and precision P defined as P = TP /(TP + FP ) for the ten cases, where TP denotes the numbers of true positives, FN the number of false negatives, and FP the numbers of false positives. We

72

Y. Suzuki et al.

Fig. 2. Illustrative results of proposed method. Left: Manual traces. Right: Proposed method. Hepatic, splenic, and left and right renal arteries. The celiac artery is shown in green.

(a)

(b)

Fig. 3. Results for topologically irregular vessel branches. Left: Manual traces. Right: Proposed method. (a) Hepatic arteries. (b) Renal arteries.

defined true positives as any detected points within two-voxel distance from the skeletonized ground-truth points and false negatives as the skeletonized groundtruth points not within two-voxel distance from detected centerline points. Fig. 4 shows a summary of the accuracy evaluation results for n = 2 and k = 100 using a combined measure of the recall and precision, F -measure, defined as 2RP/(R + P ), which is a normalized accuracy measure taking both false positives and true negatives into account. The evaluations were performed for arteries only for outside the organ and aorta regions (that is, inter-organ arteries). In Fig. 4, we compared the proposed method with our previous method in which fragmentation and over-extraction of the extracted artery regions were not considered. When n = 0 and k = 1, the method is regarded as equivalent to the previous method. As shown in Fig. 4, the labeling accuracy improved by using the proposed method. Fig. 5 shows the effects of parameters n and k on the accuracy. As shown in Fig. 5(a), recall increased along with n because larger n filled larger gap between fragmented true artery regions. As shown in Fig. 5(a) and 5(b), sufficiently large k improved accuracy by reducing wrong connections.

Automated Segmentation and Anatomical Labeling of Abdominal Arteries

73

Fig. 4. Summary of accuracy evaluation of proposed method. The results of the previous method [8] are shown for comparison.

(a)

(b)

(c)

Fig. 5. Effects of parameters n and k on accuracy. (a) Recall. (b) Precision. (c) F measure.

Both n and k contributed to improve accuracy as shown in Fig. 5(c), and the best accuracy was obtained when n = 2 and k = 100. The average F -measure was 0.89 for the proposed method (n = 2 and k = 100) and 0.74 for the previous method (n = 0 and k = 1) [8].

4

Conclusion

We have developed a novel method for artery segmentation and labeling from contrast-enhanced CT data by using segmented organ and aorta regions. The method is based on a basic anatomical constraint that the artery branches supply blood to the organ. We demonstrate that the constraint can be effectively applied to artery segmentation and anatomical labeling using automatically segmented abdominal organ and aorta regions with typical segmentation accuracy. In addition, the method could successfully deal with irregular topology which is not rare in the abdominal artery branches. Future work will include extending the method so as to deal with the vein branches and large-scale validations using more datasets.

74

Y. Suzuki et al.

Acknowledgements. This work is partly supported by MEXT Grand-in-Aid for Scientific Research on Innovative Areas No. 21103003.

References 1. Mori, K., Hasegawa, J., Suenaga, Y., et al.: Automated anatomical labeling of the bronchial branch and its application to the virtual bronchoscopy system. IEEE Trans. Med. Imaging 19(2), 103–114 (2000) 2. Mori, K., Oda, M., Egusa, T., Jiang, Z., Kitasaka, T., Fujiwara, M., Misawa, K.: Automated nomenclature of upper abdominal arteries for displaying anatomical names on virtual laparoscopic images. In: Liao, H., Edwards, P.J.E., Pan, X., Fan, Y., Yang, G.-Z. (eds.) MIAR 2010. LNCS, vol. 6326, pp. 353–362. Springer, Heidelberg (2010) 3. Bogunovi´c, H., Pozo, J.M., C´ ardenes, R., Frangi, A.F.: Anatomical labeling of the anterior circulation of the circle of willis using maximum a posteriori classification. In: Fichtinger, G., Martel, A., Peters, T. (eds.) MICCAI 2011, Part III. LNCS, vol. 6893, pp. 330–337. Springer, Heidelberg (2011) 4. Shimizu, A., Ohno, R., Ikegami, T., et al.: Segmentation of multiple organs in non-contrast 3D abdominal CT images. Int. J. Comput. Assist. Radiol. Surg. 2(3), 135–142 (2007) 5. Okada, T., Yokota, K., Hori, M., Nakamoto, M., Nakamura, H., Sato, Y.: Construction of hierarchical multi-organ statistical atlases and their application to multiorgan segmentation from CT images. In: Metaxas, D., Axel, L., Fichtinger, G., Sz´ekely, G. (eds.) MICCAI 2008, Part I. LNCS, vol. 5241, pp. 502–509. Springer, Heidelberg (2008) 6. Linguraru, M.G., Pura, J.A., Chowdhury, A.S., Summers, R.M.: Multi-organ segmentation from multi-phase abdominal CT via 4D graphs using enhancement, shape and location optimization. In: Jiang, T., Navab, N., Pluim, J.P.W., Viergever, M.A. (eds.) MICCAI 2010, Part III. LNCS, vol. 6363, pp. 89–96. Springer, Heidelberg (2010) 7. Okada, T., Linguraru, M.G., Yoshida, Y., Hori, M., Summers, R.M., Chen, Y.-W., Tomiyama, N., Sato, Y.: Abdominal multi-organ segmentation of CT images based on hierarchical spatial modeling of organ interrelations. In: Yoshida, H., Sakas, G., Linguraru, M.G. (eds.) Abdominal Imaging. LNCS, vol. 7029, pp. 173–180. Springer, Heidelberg (2012) 8. Suzuki, Y., Okada, T., Hori, M., et al.: Automated anatomical labeling of abdominal arteries from ct data based on optimal path finding between segmented organ and aorta regions: A robust method against topological variability. Int. J. CARS 7(suppl. 1), s47–s48 (2012) 9. Otsu, N.: A Threshold Selection Method from Gray-Level Histograms. IEEE Trans. Syst. Man Cybern. 9(1), 62–66 (1979) 10. Sato, Y., Nakajima, S., Shiraga, N., et al.: Three-dimensional multi-scale line filter for segmentation and visualization of curvilinear structures in medical images. Med. Image Anal. 2(2), 143–168 (1998) 11. Wink, O., Niessen, W.J., Viergever, M.A.: Multiscale vessel tracking. IEEE Trans. Med. Imaging 23(1), 130–133 (2004) 12. Glocker, G., Komodakis, N., Tziritas, G., et al.: Dense Image Registration through MRFs and Efficient Linear Programming. Med. Image Anal. 12(6), 731–741 (2008)