Abstract
We present a method to extract coronary vessels from fluoroscopic x-ray sequences. Given the vessel structure for the source frame, vessel correspondence candidates in the subsequent frame are generated by a novel hierarchical search scheme to overcome the aperture problem. Optimal correspondences are determined within a Markov random field optimization framework. Post-processing is performed to extract vessel branches newly visible due to the inflow of contrast agent. Quantitative and qualitative evaluation conducted on a dataset of 18 sequences demonstrate the effectiveness of the proposed method.
This work was supported by the Institute for Information & communications Technology Promotion (IITP) Grant (No. R0101-16-0171) and by the National Research Foundation (NRF) Grant (2015R1A5A7036384), both funded by the Korean Government (MSIP).
You have full access to this open access chapter, Download conference paper PDF
Similar content being viewed by others
Keywords
1 Introduction
Fluoroscopic X-ray angiograms (XRA, Fig. 1) are used to evaluate stenosis in coronary arteries and provide guidance for percutaneous coronary intervention. Here, vessel extraction enables registration of pre-operative CT angiograms (CTA) for visualization of 3-D arterial structure. For chronic total occlusion, this can visualize otherwise invisible arteries due to blockage of contrast agent.
Many works focus on vessel extraction from a single image. Pixelwise enhancement [1], and segmentation methods with sophisticated optimization [2, 3] or learning [4] are some examples. While these methods are applicable to a wide variety of vessels, they do not consider temporal continuity and thus may give inconsistent results for a sequence. Many works use an accurate vessel structure extracted from a detailed 3D CTA to extract an accurate and consistent vessel structure for XRA sequences [5, 6]. Relatively few works have been proposed that do not require 3D CTA to detect and track curvilinear structures such as vessels [7] or guide-wires [8] from fluoroscopic image sequences.
Thus, we present a method, which we term vessel correspondence optimization (VCO), to extract coronary vessels from fluoroscopic x-ray angiogram sequences. Given the vessel structure of a source frame (obtained by manual annotation or automatic methods [2–4]), the detailed global and local motion is estimated by determining the optimal correspondence for vessel points in the subsequent frame. Local appearance similarity and structural consistency are enforced within a Markov random field (MRF) optimization framework. Essentially, VCO performs registration of the vessel structure. Post-processing is performed to deal with vessel branches newly visible due to the inflow of contrast agent. The proposed method is summarized in Fig. 2.
The main contributions are the development of (i) an MRF optimization method for optimal registration of the vessel structure, and (ii) an accurate vessel extraction method for XRA sequences based on accurate motion estimation. Experiments show that VCO is robust to complicated vessel structures and abrupt motions. We believe that VCO can be used to provide analytic visual information to the clinician without CTA acquisition. It can also be used in the automatic registration of vessels between 2D XRA sequences and 3D CTA.
2 Vessel Correspondence Optimization
2.1 Markov Random Field Representation
A pairwise MRF graph is constructed from the vessel centerlines of the source frame. Nodes correspond to sampled points from the centerlines and edges represent connectivities between the vessel points. The MRF energy is defined as follows:
where \(\mathbf x \) is the vector comprising the set of all random variables \(x_i\) at each node with index i. Each \(x_i\) can be labeled by \(N_p +1\) different values, and the optimal \(\mathbf x \) is determined by minimizing (1). The \(N_p +1\) labels comprise \(N_p\) correspondence candidates and 1 dummy label which will be assigned to a node when there are no candidates with consistent local shape or appearance. If the dummy label is found to be optimal, that node is excluded from the resulting VCO point set.
The unary cost function \(\varphi (x_{i})\) depends on the similarity between the local appearance of the ith node and the \(x_i\)th correspondence candidate. Since we seek corresponding points with similar appearance, we define \(\varphi (x_{i})\) to decrease as local appearance similarity increases as:
\(F_{src}\) and \(F_{dst}\) denote the source and destination frames, \(p_i\) and \(\pi _i (x_i )\) denote the coordinate of the ith node of the source vessel structure and the \(x_i\)th correspondence candidate from the destination frame, respectively. D is a function for a local feature descriptor. Note that \(\varphi (x_i )\) is truncated by \(T_\varphi \) to ensure robustness to outliers. Outliers may occur when there is no corresponding point due to severe local deformations.
The pairwise cost \(\psi (x_i ,x_j )\) enforces similar displacement vectors between neighboring points, and thus consistent local shape. It is defined similar to that of [9], as follows:
where \(p_i\) and \(p_j\) are coordinates of the ith and jth source node, while \(\pi _i (x_i )\) and \(\pi _j (x_j )\) are coordinates in the destination frame of the correspondence candidates of \(x_i\) and \(x_j\), respectively. \(p_i - \pi _i (x_i )\) is the displacement vector of the ith source node. Again, truncation is included based on threshold \(T_\psi \). The parameter \(\lambda \) controls the amount of this regularization in (1).
2.2 Hierarchical Search of Vessel Correspondence Candidates
The tubular shapes of vessels, together with the aperture problem, make it very challenging to distinguish different local regions. We thus propose a hierarchical correspondence search scheme comprising global, branch, and point searches.
We define vessel junctions, including bifurcations and crossings from 3D-2D projection, and endpoints, both of which have distinctive appearances, as vessel keypoints. A vessel branch refers to the line connecting two vessel keypoints. In the following, we denote the \(\alpha \)th keypoint as \(p^\alpha \), with a superscript, to distinguish it from general vessel point \(p_i\). The mth branch is denoted as \(b_m\).
Global Search by Chamfer Matching: We perform chamfer matching [10] to estimate large global translational motion from heart beating, breathing, or viewpoint change. The template shape is the set of source vessel points. The target shape is constructed by sequentially applying vessel enhancement [1], thresholding, and skeletonization to the destination frame. We find the global displacement vector that minimizes the sum of distances between each template point and target shape by brute force search on the distance transform (DT) of the target shape. Figure 2(a) shows an example result of this step.
Branch Search by Vessel Keypoint Correspondence: We search for corresponding points \(\pi ^\alpha \) and \(\pi ^\beta \) for both keypoints \(p^\alpha \) and \(p^\beta \) of a branch \(b_m\), each within a local search region of size \(w_{k} \times h_{k}\). Correspondences are determined by similarity of local appearance, measured using (2). Non-max suppression is applied to avoid nearby matches, and up to \(N_k\) possible correspondences are obtained for both \(p^\alpha \) and \(p^\beta \). The set of candidate branches is generated for \(b_m\) by simply applying all displacement vectors \(\delta ^\alpha _{1} = \pi ^\alpha _{1}-p^\alpha , ... ,\delta ^\alpha _{N_{k}} =\pi ^\alpha _{N_{k}}-p^\alpha \) and \( \delta ^\beta _{1} = \pi ^\beta _{1}-p^\beta , ... ,\delta ^\beta _{N_{k}} =\pi ^\beta _{N_{k}}-p^\beta \) to \(b_m\). We note that there can be up to \(N_k \times 2\) candidate branches, depending on the number of keypoint correspondences. We also include the branch with no displacement, in case all keypoint matches are unreliable, which results in at most \(N_k \times 2 + 1\) branch candidates.
Correspondence Candidate Generation by Vessel Point Search: For a vessel point \(p_{i} \in b_{m}\), the set of corresponding points based on candidate branches are \(\{p_{i}+\delta ^\alpha _{1}, ... ,p_{i}+\delta ^\alpha _{N_{k}}, p_{i}+\delta ^\beta _{1}, ... ,p_{i}+\delta ^\beta _{N_{k}}, p_{i} \}\). We define local search regions of size \(w_{p} \times h_{p}\) at each of these points and determine the \(N_l\) best corresponding points, again, based on (2). Figure 3 shows an example where the correspondence candidates are greatly improved with a smaller local search range based on the prior branch search.
The resulting maximum number of candidates is \(N_p = N_l \times (N_k \times 2 + 1)\). Due to non-max suppression, the actual candidate number \(N_c\) can be less than \(N_p\) depending on the image, which can complicate implementation. Thus, we fix the number of labels to \(N_p\) for all vessel points, but nullify labels larger than \(N_c\) without an actual corresponding candidate by assigning an infinite unary cost.
2.3 Post-processing for Complete Centerline Extraction
After VCO, obtained vessel points are connected to construct vessel centerline structure \(\mathcal V\) by the fast marching method [11]. To extract newly visible branches, a binary segmentation mask is first obtained by thresholding vesselness [1], and regions not connected to the vessel centerlines are excluded. The DT is computed for this mask, with the non-vessel regions as seeds. Based on the mask and the DT, the following is iterated until no branch longer than the maximum vessel radius is found: (i) perform fast marching method, with \(\mathcal V\) as the seeds and the DT values as marching speed; (ii) find the shortest path from the pixel with the latest arrival time to \(\mathcal V\), and add to \(\mathcal V\). This method is adopted from [12].
3 Experimental Results
Evaluation Details: The dataset comprises 18 XRA sequences of total 617 frames from 5 different patients. All sequences were acquired by Philips digital fluoroscopic systems at \(512 \times 512\) resolution, 8 bit depth, and 15 fps frame rate. Parameter values were manually selected and fixed to \(N_k=2\), \(w_{k}=101\), \(h_{k}=101\), \(N_l=5\), \(w_{p}=21\), \(h_{p}=21\), and \(N_{p} = N_l \times (N_k \times 2 + 1) = 25\). Source vessel points are sampled from the centerline at a 5 pixel interval. The VLFeat [13] library for SIFT [14] is used for D in (2). TRW-S is used for MRF optimization [15]. VCO without post-processing took a few minutes by an unoptimized Matlab implementation on a Intel Zeon processor with over \(80\,\%\) of computation on SIFT matching. Ground truth vessel centerlines and corresponding bifurcation point coordinates in frame pairs were obtained by expert annotation. For centerlines, the semi-automatic method of [2] was used.
Comparisons are made with two relevant methods: (1) that by Fallavollita et al. [7], which combines optical flow (OF) and an active contour model (ACM), modified to handle whole vessel structures, denoted as OF+ACM, and (2) a variant of OF+ACM where OF is substituted with global chamfer matching, denoted as CM+ACM. Further comparisons are made with two variants of VCO: (3) VCO-HS VCO without hierarchical search, with enlarged search regions instead, and (4)VCO-DL, VCO without dummy labels.
Quantitative Evaluation: We perform two different experiments. In Exp1, vessel extraction is performed for all source-destination frame pairs, using the ground truth source vessel structure. The average precision, recall, F-measure, along with the target registration error (TRE) for bifurcation points, of 599 sequential frame pairs are presented in Table 1. TRE is defined as the average distance between the estimated and ground truth point coordinate. A vessel point is true positive if there is a ground truth point within a two pixel radius. VCO achieves the highest accuracy, with additional improvement from hierarchical search and dummy labels. Here, post-processing is applied only up to point connection to exclusively evaluate the VCO accuracy.
In Exp2, we evaluate the average number of consecutive frames with sufficient vessel extraction when iteratively applying all methods for a single initial frame. Here, sufficient is defined as F-measure higher than 0.7. Evaluation on the 18 sequences showed that sufficient vessel extraction was obtained for 2.0, 5.3, and 8.6 subsequent frames, by the OF+ACM, CM+ACM, and VCO methods, respectively. This demonstrates the practical usefulness of VCO.
Qualitative Evaluation: Figure 4 presents representative sample results of Exp1. Significant improvements compared to previous methods are visible. Figure 1 presents results of Exp2 for a sample sequence. Both figures highlight the effectiveness of the proposed method. Figure 5 shows one limitation of VCO, where VCO is not able to handle the topology change due to vessel superimposition.
4 Conclusion
We have proposed a method to determine optimal vessel correspondences for registration and extraction of vessel structures from fluoroscopic x-ray sequences. Experiments show promising results on complicated structures. In future work, we plan to investigate optimization measures including GPU implementation for real-time performance as well as other measures required for actual clinical application such as overlaying the extracted vessel structure to sequences acquired without the use of contrast agents.
References
Frangi, A.F., Niessen, W.J., Vincken, K.L., Viergever, M.A.: Multiscale vessel enhancement filtering. In: Wells, W.M., Colchester, A.C.F., Delp, S.L. (eds.) MICCAI 1998. LNCS, vol. 1496, pp. 130–137. Springer, Heidelberg (1998)
Poon, K., Hamarneh, G., Abugharbieh, R.: Live-vessel: extending livewire for simultaneous extraction of optimal medial and boundary paths in vascular images. In: Ayache, N., Ourselin, S., Maeder, A. (eds.) MICCAI 2007, Part II. LNCS, vol. 4792, pp. 444–451. Springer, Heidelberg (2007)
Honnorat, N., Vaillant, R., Duncan, J.S., Paragios, N.: Curvilinear structures extraction in cluttered bioimaging data with discrete optimization methods. In: ISBI, pp. 1353–1357 (2011)
Becker, C., Rigamonti, R., Lepetit, V., Fua, P.: Supervised feature learning for curvilinear structure segmentation. In: Mori, K., Sakuma, I., Sato, Y., Barillot, C., Navab, N. (eds.) MICCAI 2013, Part I. LNCS, vol. 8149, pp. 526–533. Springer, Heidelberg (2013)
Rivest-Henault, D., Sundar, H., Cheriet, M.: Nonrigid 2D/3D registration of coronary artery models with live fluoroscopy for guidance of cardiac interventions. IEEE Trans. Med. Imaging 31(8), 1557–1572 (2012)
Sun, S.-Y., Wang, P., Sun, S., Chen, T.: Model-guided extraction of coronary vessel structures in 2D X-Ray angiograms. In: Golland, P., Hata, N., Barillot, C., Hornegger, J., Howe, R. (eds.) MICCAI 2014, Part II. LNCS, vol. 8674, pp. 594–602. Springer, Heidelberg (2014)
Fallavollita, P., Cheriet, F.: Robust coronary artery tracking from fluoroscopic image sequences. In: Kamel, M.S., Campilho, A. (eds.) ICIAR 2007. LNCS, vol. 4633, pp. 889–898. Springer, Heidelberg (2007)
Honnorat, N., Vaillant, R., Paragios, N.: Graph-based geometric-iconic guide-wire tracking. In: Fichtinger, G., Martel, A., Peters, T. (eds.) MICCAI 2011, Part I. LNCS, vol. 6891, pp. 9–16. Springer, Heidelberg (2011)
Glocker, B., Komodakis, N., Paragios, N., Tziritas, G., Navab, N.: Inter and intra-modal deformable registration: continuous deformations meet efficient optimal linear programming. In: Karssemeijer, N., Lelieveldt, B. (eds.) IPMI 2007. LNCS, vol. 4584, pp. 408–420. Springer, Heidelberg (2007)
Barrow, H., Tenenbaum, J., Bolles, R., Wolf, H.: Parametric correspondence and chamfer matching: two new techniques for image matching. In: Proceedings of International Joint Conference on Artificial Intelligence, vol. 2, pp. 659–663 (1977)
Yatziv, L., Bartesaghi, A., Sapiro, G.: O(N) implementation of the fast marching algorithm. J. Comput. Phys. 212(2), 393–399 (2005)
Van Uitert, R., Bitter, I.: Subvoxel precise skeletons of volumetric data based on fast marching methods. Med. Phys. 34(2), 627–638 (2007)
Vedaldi, A., Fulkerson, B.: VLFeat: an open and portable library of computer vision algorithms (2008). http://www.vlfeat.org/
Lowe, D.G.: Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision 60(2), 91–110 (2004)
Kolmogorov, V.: Convergent tree-reweighted message passing for energy minimization. IEEE Trans. Pattern Anal. Mach. Intell. 28(10), 1568–1583 (2006)
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2016 Springer International Publishing AG
About this paper
Cite this paper
Shin, S.Y., Lee, S., Noh, K.J., Yun, I.D., Lee, K.M. (2016). Extraction of Coronary Vessels in Fluoroscopic X-Ray Sequences Using Vessel Correspondence Optimization. In: Ourselin, S., Joskowicz, L., Sabuncu, M.R., Unal, G., Wells, W. (eds) Medical Image Computing and Computer-Assisted Intervention - MICCAI 2016. MICCAI 2016. Lecture Notes in Computer Science(), vol 9902. Springer, Cham. https://doi.org/10.1007/978-3-319-46726-9_36
Download citation
DOI: https://doi.org/10.1007/978-3-319-46726-9_36
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-46725-2
Online ISBN: 978-3-319-46726-9
eBook Packages: Computer ScienceComputer Science (R0)