Conference PaperPDF Available

Automatic Artery-Vein Separation from Thoracic CT Images Using Integer Programming


Abstract and Figures

Automated computer-aided analysis of lung vessels has shown to yield promising results for non-invasive diagnosis of lung diseases. In order to detect vascular changes affecting arteries and veins differently, an algorithm capable of identifying these two compartments is needed. We propose a fully automatic algorithm that separates arteries and veins in thoracic computed tomography (CT) images based on two integer programs. The first extracts multiple subtrees inside a graph of vessel paths. The second labels each tree as either artery or vein by maximizing both, the contact surface in their Voronoi diagram, and a measure based on closeness to accompanying bronchi. We evaluate the performance of our automatic algorithm on 10 manual segmentations of arterial and venous trees from patients with and without pulmonary vascular disease, achieving an average voxel based overlap of 94.1% (range: 85.0% – 98.7%), outperforming a recent state-of-the-art interactive method.
Content may be subject to copyright.
Automatic Artery-Vein Separation from
Thoracic CT Images using Integer Programming
Christian Payer1,2,?, Michael Pienn2, Zolt´an B´alint2,
Andrea Olschewski2, Horst Olschewski3, and Martin Urschler4,1
1Institute for Computer Graphics and Vision,
BioTechMed, Graz University of Technology, Austria
2Ludwig Boltzmann Institute for Lung Vascular Research, Graz, Austria
3Department of Pulmonology, Medical University of Graz, Austria
4Ludwig Boltzmann Institute for Clinical Forensic Imaging, Graz, Austria
Abstract. Automated computer-aided analysis of lung vessels has shown
to yield promising results for non-invasive diagnosis of lung diseases. In
order to detect vascular changes affecting arteries and veins differently,
an algorithm capable of identifying these two compartments is needed.
We propose a fully automatic algorithm that separates arteries and veins
in thoracic computed tomography (CT) images based on two integer pro-
grams. The first extracts multiple subtrees inside a graph of vessel paths.
The second labels each tree as either artery or vein by maximizing both,
the contact surface in their Voronoi diagram, and a measure based on
closeness to accompanying bronchi. We evaluate the performance of our
automatic algorithm on 10 manual segmentations of arterial and venous
trees from patients with and without pulmonary vascular disease, achiev-
ing an average voxel based overlap of 94.1% (range: 85.0% – 98.7%),
outperforming a recent state-of-the-art interactive method.
1 Introduction
The automatic extraction of vascular tree structures is highly relevant in medical
image analysis. Especially the investigation of the pulmonary vascular tree, the
complex lung vessel network responsible for oxygen uptake and carbon dioxide
release, is of great interest in clinical practice. Applications include detection
of early stage pulmonary nodules [1], pulmonary embolism [2], or pulmonary
hypertension [3]. The recent trend towards quantitative pulmonary computer-
aided diagnosis (CAD) was enabled by the rapid progress in medical imaging
technologies, with state-of-the-art computed tomography (CT) scanners allow-
ing depiction of anatomical structures in the lung at a sub-millimeter level at
low radiation doses. This remarkable resolution can help in CAD of diseases like
chronic obstructive pulmonary disease [4] or pulmonary hypertension [5]. How-
ever, the higher amount of data leads to an increased demand for fully automatic
?We thank Peter Kullnig for radiological support, Vasile F´oris for medical support and
data analysis and Martin Wurm for his help in the manual artery-vein separations.
2 Payer et al.
(a) CT slice (b) Unlabeled vessels (c) Our A/V separation
Fig. 1. Thoracic CT case where arteries (blue) and veins (red) are close to each other.
extraction and computer-aided quantitative analysis of pulmonary structures, in-
cluding the automated separation of arterial and venous trees (see Fig. 1). This
can improve the diagnosis of lung diseases affecting both trees differently.
Automatic artery-vein (A/V) separation is a very complex problem, due to
the similar intensity values of both vessel trees in CT. Further, the pulmonary
arterial and venous trees are intertwined and the vessels are in close proximity,
making their distinction even harder. Most of the few existing A/V separation
algorithms start with vessel segmentation, often using tubularity filters combined
with region growing or fast marching methods based on seed points [6]. The
work of [7] proposes to solely detect pulmonary arteries by using the anatomical
constraint that arteries usually run along the bronchi, whereas the method of [8]
involves global structural information by constructing a minimum-spanning-tree
from weights derived from local vessel geometry measures and a cutting step for
A/V separation. However, using this method, an interactive refinement is often
necessary to finalize the separation. In [9] A/V separation utilizes the close
proximity of arteries and veins. By morphological operations with differently
sized kernels, equal intensity structures are split and locally separated regions
are traced. [10] extended this method with a GUI enabling efficient refinement.
Another promising method is [11], who formulate an automatic voxel labeling
problem based on root detection for both trees. However, it requires a training
step and, due to locally restricted image features, it still has problems near the
hilum of the lung, i.e. where arteries and veins are in close proximity.
In this work we present a novel, automatic A/V separation algorithm for
thoracic CT images, which requires no manual correction and takes the global
structural information about vascular trees as well as local features like vessel
orientation and bronchus proximity into account. Based on a vessel segmentation
step we formulate both the extraction of subtrees and the labeling of arteries and
veins as integer programs. We evaluate our method on a database of 10 thoracic
CT images with manually segmented A/V trees as a reference and demonstrate
the benefits of our method compared to the state-of-the-art method in [8].
Automatic Artery-Vein Separation from Thoracic CT 3
4D paths
Subtree A/V
CT image
Fig. 2. Overview of our proposed A/V separation algorithm.
2 Method
Our proposed algorithm for A/V separation from contrast-enhanced thoracic CT
images is shown in Fig. 2. After a lung segmentation from [5], subsequent process-
ing is performed for both lungs independently. A multi-scale vessel enhancement
filter [12] produces a vessel orientation estimate as well as a 4D tubularity im-
age for the three spatial coordinates and the radius. Next, we calculate a graph
G= (V, E ) of regularly spaced local maxima Vof the vessel enhanced image,
which are connected by edges Ein a local neighborhood similar to [13]. For
every edge, a path between its two endpoints is extracted, which minimizes the
geodesic distance, penalizing small tubularity values along the path [12]. To dras-
tically prune these edges, a filtering step is performed removing all paths that
fully contain any other path. The resulting graph still contains many spurious
edges, but also the real arterial and venous vessel paths. These are subsequently
organized in subtrees and separated using two integer programs.
2.1 Subtree Extraction
In order to identify anatomically meaningful vascular trees and prepare the input
for A/V separation, we contribute a novel method to extract a set of connected
subtrees from the overcomplete maxima graph G, using an optimization proce-
dure based on an integer program. Different from [13], we do not need explicitly
declared root nodes, but include their search into the optimization. Formally,
we find multiple tree-like structures in G, defined by edge tuples heij , tij i, where
binary variable tij = 1 indicates that the path from node ito node jis active,
i.e. contained in one of the resulting subtrees. The quadratic objective function
is a sum of weights wijk Rof adjacent oriented edge pairs described by tij and
tjk to model the tree structure, and a term controlling the creation of subtrees
formalized by a binary variable rij ∈ {0,1}indicating if eij is a root of a subtree:
4 Payer et al.
arg min
t,r X
eij , ejk E
wijk tij tj k +σX
eij E
s.t. Pehi Ethi +rij tij ,Pehi Ethi +rij 1,
tij rij , tij +tji 1,eij E
The linear constraints in (1) enforce tree-like structures by ensuring that an
active edge tij = 1 has exactly one predecessor thi = 1 in the set ehi of all
preceding edges, or is a root node rij = 1. The fourth constraint guarantees that
a directed edge and its opposite are not active simultaneously. With σR+
number of created subtrees is globally controlled, where in contrast to [13] we
allow growing of numerous subtrees throughout the whole local maxima graph,
and select the best root nodes of anatomically meaningful subtrees implicitly.
The key component of the integer program is the weight wijk of adjacent ori-
ented edge pairs. It consists of three parts, the first one derived from the costs
of traveling along the path from node ito kvia jaccording to the tubularity
measure, the second one penalizing paths with strong orientation changes com-
puted by the mean of the dot products of directions along a path, and the final
part penalizing radius increases from start node ito end node k. While all other
components are positive, a global parameter δRis further added to wijk to
allow for negative weights, otherwise the trivial solution, i.e. no extracted paths
and subtrees, would always minimize the objective function (1).
After minimizing the objective function, the integer program results in a list
of individual connected subtrees, which is post-processed to locate the anatom-
ical branching points instead of local tubularity maxima.
2.2 Subtree A/V Labeling
Next, each individual subtree is labeled as either artery or vein using two anatom-
ical properties. First, we exploit that arteries and veins are roughly uniformly
distributed in the lung. The second property uses the fact, that bronchi run par-
allel and in close proximity to arteries, as previously proposed in [7]. The main
contribution of our work lies in a novel optimization model calculating the final
A/V labeling by an integer program, that assigns to every individual subtree ti
either ai= 1 (artery) or vi= 1 (vein). This is achieved by maximizing
arg max
a,v X
ti, tjT
ij +λX
s.t. ai+vi= 1,tiT,
where the first term counts the number of voxels wborder
ij R+
0on the contact sur-
face between artery and vein regions modeled by a generalized Voronoi diagram.
For each voxel inside the lung segmentation this Voronoi diagram determines
the nearest subtree. By maximizing the sum of all wborder
ij of neighboring artery
Automatic Artery-Vein Separation from Thoracic CT 5
and vein regions, a uniform distribution of arteries and veins throughout the
whole lung is ensured. The second term of (2) uses a measure of arterialness
0for every tree tito incorporate a distinction between arteries and
veins. Our arterialness measure is inspired by [7], but instead of searching for
bronchus points in the input CT image, we employ the multi-scale tubularity
filter from [12] for locating dark on bright bronchus structures. At each voxel
along the segments of a subtree, we locally search for similarly oriented bronchial
structures giving high tubularity response in a plane orthogonal to the vessel di-
rection. After fitting a line through all bronchus candidate locations of a vessel
segment, we compute an arterialness measure from their distance and deviation
in direction. This gives higher values for arteries running closer and in parallel
to bronchi, while veins, typically more distant and deviating stronger from the
bronchus direction, will receive lower values. The arterialness value wartery
a tree tiis the sum of all arterialness values of its vessel segments. Finally, the
constraint from (2) ensures, that not both labels are active at the same time for
the same tree ti. A factor λweights the sums. The result of solving this integer
program is the final labeling of arteries and veins for all subtrees.
3 Experimental Setup and Results
Experimental Setup: To validate our algorithm, we used 10 datasets from pa-
tients (6 female/4 male) with and without lung vascular disease who underwent
thoracic contrast-enhanced, dual-energy CT examinations. The CT scans were
acquired either with a Siemens Somatom Definition Flash (D30f reconstruction
kernel) or with a Siemens Somatom Force (Qr40d reconstruction kernel) CT
scanner. The size of the isotropic 0.6mm CT volumes was 512×512 ×463 pixels.
Manual reference segmentations of all 10 patients, that include pulmonary
artery and left atrium, as well as A/V trees down to a vessel diameter of 2mm,
were created requiring 5–8 h per dataset. These data served as basis for vali-
dating our algorithm together with a re-implementation of [8], which is similar
to our proposed method, as it extracts and labels subtrees. The interactive step
of [8], i.e. the final A/V labeling of subtrees, was done manually by looking for
connections to the heart for every subtree. Similarly, we additionally performed
user-defined labeling of the extracted subtrees of our algorithm, to evaluate the
A/V labeling part. As the compared segmentations may differ substantially in
the included vessel voxels, we compared only those voxels which are present in
the segmentation and the manual reference.
The development and testing platform for our C++ algorithm consisted of a
Windows 7 Intel Core i5-4670 @ 3.40 GHz with 16 GB RAM. For multi-scale
vessel enhancement and 4D path extraction, the publicly available code from [12]
was used. Gurobi Optimizer1was applied to solve integer programs. The param-
eters σ= 0.2 and δ=0.2 were determined empirically within a few try-outs
providing satisfactory results for the subtree extraction. The parameter λ= 6.0
was determined by grid search, as the A/V labeling is not time consuming.
1Gurobi Optimizer Version 6.0 with academic license from
6 Payer et al.
Table 1. Overlap ratio in % of automatic, Park et al. [8] and user-defined (UD) labeling
methods with manual reference segmentations for the 10 datasets.
Patient # 1 2 3 4 5 6 7 8 9 10 µ
Automatic 95.0 93.3 98.4 87.3 97.4 85.0 95.0 98.7 98.5 92.4 94.1
Park et al. [8] 90.2 93.9 91.2 90.4 91.9 88.0 90.6 94.7 95.3 93.3 91.9
UD labeling 99.1 99.4 98.6 98.5 97.9 97.6 99.2 99.8 98.6 99.1 98.8
Results: The automatic algorithm generated an average of 1210 individual ves-
sel segments, composed of 619 arteries and 591 veins, with diameters ranging
from 2 to 10mm. The average voxel-based overlap of correct labels for all 10
datasets between automatic and manual reference segmentation was 94.1%. The
re-implementation of [8] achieved 91.9%, whereas the user-defined subtree label-
ing based on the output of our subtree extraction achieved an overlap of 98.8%.
The individual values are listed in Table 1, with two datasets visualized in Fig. 3.
Furthermore, in order to evaluate the subtree A/V labeling in more detail, we
validated our automatic segmentation against the user-defined labeling (Table 2).
The average ratio of voxels, where arteries and veins were switched (mislabeled)
was 4.9%. As additional voxels cannot be added by the manual labeling, the
number of mistakenly detected vessels can be quantified as well. The average
ratio of misclassified structures (non-vessel), i.e., the ratio of voxels that are not
in the manually labeled result, but present in the automatic one, was 1.7%.
The average time needed for generating a single, fully automatic A/V seg-
mentation with our unoptimized method was 5 hours, while [8] required 0.5
hours of computation time with additional 2 hours of interaction time.
Table 2. Comparison of user-defined subtree labeling and fully automatic segmenta-
tion for the 10 datasets. The amount of mislabeled vessel voxels (Mislabeled) and the
amount of detected structures, which are not vessels (Non-vessel) is provided in %.
Patient # 1 2 3 4 5 6 7 8 9 10 µ
Mislabeled 3.9 6.1 0.2 11.6 1.8 10.4 4.6 1.4 0.6 7.9 4.9
Non-vessel 5.1 0.2 1.4 2.2 4.7 3.0 0.1 0.3 0.4 0.0 1.7
4 Discussion and Conclusion
Our proposed novel algorithm for separating arteries and veins in thoracic CT
images achieves a higher average correct overlap compared to the method of [8],
although the latter method includes explicit manual correction, while our ap-
proach is fully automated. We assume that our integer program is better mod-
eling the anatomical constraints involved in A/V separation compared to the
Automatic Artery-Vein Separation from Thoracic CT 7
(a) Reference (b) Automatic (c) Overlap
(d) Reference (e) Automatic (f) Overlap
Fig. 3. Visualization of reference segmentation (left), automatic segmentations (cen-
ter) and their overlap (right). Only voxels, that are set in both segmentations, are
visualized in the overlap image. Red: vein, blue: arteries, yellow: disagreement between
segmentations. Top row: 98.4% agreement (#3), bottom row: 85% agreement (#6).
minimum-spanning-tree of [8] by exploiting the uniform distribution of arteries
and veins throughout the lungs as well as proximity and similar orientation of
arteries and bronchi. Our extracted vascular subtrees are very well separated,
which can be observed in Fig. 3 or by comparing the user-defined labeling with
the manual reference in Table 2. Furthermore, our main contribution, the auto-
matic A/V labeling step, removes the need for manual post-processing.
We observed that mislabeled subtrees often are neighbors in the generalized
Voronoi diagram, due to the maximization of their contact surfaces. This may
lead to switched labels of all subtrees within a lung lobe. Therefore, restricting
the uniform distribution of arteries and veins to lung lobes instead of whole
lungs could further improve performance and robustness. Because of the lack
of a standardized dataset and openly available A/V separation algorithms, our
algorithm was only compared to the most recently published work in [8]. The
evaluation of our algorithm on a larger dataset is ongoing. As our algorithm is
not optimized in its current state, we expect that improvements in runtime are
still possible. Another interesting idea could be to combine subtree extraction
and labeling into one integer program, like in [14]. We conclude, that our novel
method provides an opportunity to become an integral part of computer aided
diagnosis of lung diseases.
8 Payer et al.
1. Murphy, K., van Ginneken, B., Schilham, A.M.R., de Hoop, B.J., Gietema, H.A.,
Prokop, M.: A large-scale evaluation of automatic pulmonary nodule detection in
chest CT using local image features and k-nearest-neighbour classification. Med.
Image Anal. 13(5) (2009) 757–770
2. Masutani, Y., MacMahon, H., Doi, K.: Computerized detection of pulmonary
embolism in spiral CT angiography based on volumetric image analysis. IEEE
Trans. Med. Imaging 21(12) (2002) 1517–1523
3. Linguraru, M.G., Pura, J.A., Van Uitert, R.L., Mukherjee, N., Summers, R.M.,
Minniti, C., Gladwin, M.T., Kato, G., Machado, R.F., Wood, B.J.: Segmentation
and quantification of pulmonary artery for noninvasive CT assessment of sickle cell
secondary pulmonary hypertension. Med. Phys. 37(4) (2010) 1522–1532
4. Est´epar, R.S.J., Kinney, G.L., Black-Shinn, J.L., Bowler, R.P., Kindlmann, G.L.,
Ross, J.C., Kikinis, R., Han, M.K., Come, C.E., Diaz, A.A., Cho, M.H., Hersh,
C.P., Schroeder, J.D., Reilly, J.J., Lynch, D.A., Crapo, J.D., Wells, J.M., Drans-
field, M.T., Hokanson, J.E., Washko, G.R.: Computed tomographic measures of
pulmonary vascular morphology in smokers and their clinical implications. Am. J.
Respir. Crit. Care Med. 188(2) (2013) 231–239
5. Helmberger, M., Pienn, M., Urschler, M., Kullnig, P., Stollberger, R., Kovacs, G.,
Olschewski, A., Olschewski, H., B´alint, Z.: Quantification of tortuosity and fractal
dimension of the lung vessels in pulmonary hypertension patients. PLoS One 9(1)
(2014) e87515
6. van Rikxoort, E.M., van Ginneken, B.: Automated segmentation of pulmonary
structures in thoracic computed tomography scans: a review. Phys. Med. Biol. 58
(2013) R187–R220
7. ulow, T., Wiemker, R., Blaffert, T., Lorenz, C., Renisch, S.: Automatic extraction
of the pulmonary artery tree from multi-slice CT data. In: Proc. SPIE 5746, Med.
Imaging Physiol. Funct. Struct. from Med. Images. (2005) 730–740
8. Park, S., Lee, S.M., Kim, N., Seo, J.B., Shin, H.: Automatic reconstruction of the
arterial and venous trees on volumetric chest CT. Med. Phys. 40(7) (2013) 071906
9. Saha, P.K., Gao, Z., Alford, S.K., Sonka, M., Hoffman, E.A.: Topomorphologic
separation of fused isointensity objects via multiscale opening: Separating arteries
and veins in 3-D pulmonary CT. IEEE Trans. Med. Imaging 29(3) (2010) 840–851
10. Gao, Z., Grout, R.W., Holtze, C., Hoffman, E.A., Saha, P.K.: A new paradigm
of interactive artery/vein separation in noncontrast pulmonary CT imaging using
multiscale topomorphologic opening. IEEE Trans. Biomed. Eng. 59(11) (2012)
11. Kitamura, Y., Li, Y., Ito, W., Ishikawa, H.: Adaptive higher-order submodular
potentials for pulmonary artery-vein segmentation. In: Proc. Fifth Int. Work.
Pulm. Image Anal. (2013) 53–61
12. Benmansour, F., T¨uretken, E., Fua, P.: Tubular Geodesics using Oriented Flux:
An ITK Implementation. The Insight Journal (2013)
13. uretken, E., Benmansour, F., Andres, B., Pfister, H., Fua, P.: Reconstructing
Loopy Curvilinear Structures Using Integer Programming. In: IEEE Conf. Com-
put. Vis. Pattern Recognit. (2013) 1822–1829
14. Robben, D., T¨uretken, E., Sunaert, S., Thijs, V., Wilms, G., Fua, P., Maes, F.,
Suetens, P.: Simultaneous Segmentation and Anatomical Labeling of the Cerebral
Vasculature. MICCAI (2014) 307–314
... In recent years, several new ideas for automatic A/V separation have been proposed. Payer et al. [9] combined their previous ideas and proposed a fully automated algorithm for arterial and venous separation in thoracic CT images based on two integer programming. Through integer programming, parameter tuning was carried out to extract subtrees, and the A/V separation was performed. ...
... Payer [9] 96.3% median voxel enhanced Charbonnier [10] 89.0% median voxel noncontrast Daniel [11] 89.1% mean particle noncontrast Nardelli [12] 94.0% mean particle noncontrast Zhai [13] 77.8% mean particle enhanced Yulei Qin [14] 90.3% mean voxel noncontrast Our approach 96.2% mean particle noncontrast Table I presents the comparison results of our methods in recent years for pulmonary artery-vein segmentation. As shown in Table I, Payer [9] implemented A/V separation on enhanced CT based on two integer programming with an interactive accuracy of 96.3%. ...
... Payer [9] 96.3% median voxel enhanced Charbonnier [10] 89.0% median voxel noncontrast Daniel [11] 89.1% mean particle noncontrast Nardelli [12] 94.0% mean particle noncontrast Zhai [13] 77.8% mean particle enhanced Yulei Qin [14] 90.3% mean voxel noncontrast Our approach 96.2% mean particle noncontrast Table I presents the comparison results of our methods in recent years for pulmonary artery-vein segmentation. As shown in Table I, Payer [9] implemented A/V separation on enhanced CT based on two integer programming with an interactive accuracy of 96.3%. Charbonnier [10] used tree partitioning and peripheral vessel matching to classify arteries and veins with median accuracy of 89.0% on noncontrast CT. ...
With the development of medical computer-aided diagnostic systems, pulmonary artery-vein(A/V) reconstruction plays a crucial role in assisting doctors in preoperative planning for lung cancer surgery. However, distinguishing arterial from venous irrigation in chest CT images remains a challenge due to the similarity and complex structure of the arteries and veins. We propose a novel method for automatic separation of pulmonary arteries and veins from chest CT images. The method consists of three parts. First, global connection information and local feature information are used to construct a complete topological tree and ensure the continuity of vessel reconstruction. Second, the multitask classification network proposed can automatically learn the differences between arteries and veins at different scales to reduce classification errors caused by changes in terminal vessel characteristics. Finally, the topology optimizer considers interbranch and intrabranch topological relationships to maintain spatial consistency to avoid the misclassification of A/V irrigations. We validate the performance of the method on chest CT images. Compared with manual classification, the proposed method achieves an average accuracy of 96.2% on noncontrast chest CT. In addition, the method has been proven to have good generalization, that is, the accuracies of 93.8% and 94.8% are obtained for CT scans from other devices and other modes, respectively. The result of pulmonary artery-vein reconstruction obtained by the proposed method can provide better assistance for preoperative planning of lung cancer surgery.
... We evaluate our method on a database of 25 thoracic CT images with manually segmented and labeled A/V trees as a reference and demonstrate the benefits of our method compared to a related method from (Park et al., 2013). The method described in this paper extends a recent conference paper (Payer et al., 2015) by giving a more detailed technical description, proposing a dedicated, more efficient solver for the time-consuming subtree extraction, and finally presenting a more comprehensive experimental evaluation. ...
... The paper is organized as follows: Section 2 places our work in the context of the current related work. In Sec. 3 we explain in detail our novel A/V separation approach which is formalized more compactly and solved more efficiently compared to (Payer et al., 2015). Section 4 defines our experimental setup and the dataset we are evaluating on, while Sec. 5 describes the results of our quantitative evaluation. ...
... Additionally, in Appendix B we discuss possible strategies to tackle this problem in case cycles would occur (existing approaches and a simple heuristic). It can be seen that the formulation (2)-(4) is fully equivalent (including the possibility of cycles) to the one we proposed in (Payer et al., 2015). There, an explicit variable indicating a root edge was used. ...
Full-text available
Automated computer-aided analysis of lung vessels has shown to yield promising results for non-invasive diagnosis of lung diseases. To detect vascular changes which affect pulmonary arteries and veins differently, both compartments need to be identified. We present a novel, fully automatic method that separates arteries and veins in thoracic computed tomography images, by combining local as well as global properties of pulmonary vessels. We split the problem into two parts: the extraction of multiple distinct vessel subtrees, and their subsequent labeling into arteries and veins. Subtree extraction is performed with an integer program (IP), based on local vessel geometry. As naively solving this IP is time-consuming, we show how to drastically reduce computational effort by reformulating it as a Markov Random Field. Afterwards, each subtree is labeled as either arterial or venous by a second IP, using two anatomical properties of pulmonary vessels: the uniform distribution of arteries and veins, and the parallel configuration and close proximity of arteries and bronchi. We evaluate algorithm performance by comparing the results with 25 voxel-based manual reference segmentations. On this dataset, we show good performance of the subtree extraction, consisting of very few non-vascular structures (median value: 0.9%) and merged subtrees (median value: 0.6%). The resulting separation of arteries and veins achieves a median voxel-based overlap of 96.3% with the manual reference segmentations, outperforming a state-of-the-art interactive method. In conclusion, our novel approach provides an opportunity to become an integral part of computer aided pulmonary diagnosis, where artery/vein separation is important.
Full-text available
This document describes an ITK implementation of an interactive method for tracing curvilinear structures. The basic tools provided in this framework are an oriented flux-based tubularity measure and a geodesic path tracer that uses the fast marching algorithm. The framework is efficient and requires minimal user interaction to trace curvilinear structures such as vessels and neurites in 2D images and 3D image stacks.
Full-text available
Rationale: Angiographic investigation suggests that pulmonary vascular remodeling in smokers is characterized by distal pruning of the blood vessels. Objectives: Using volumetric computed tomography scans of the chest we sought to quantitatively evaluate this process and assess its clinical associations. Methods: Pulmonary vessels were automatically identified, segmented, and measured. Total blood vessel volume (TBV) and the aggregate vessel volume for vessels less than 5 mm(2) (BV5) were calculated for all lobes. The lobe-specific BV5 measures were normalized to the TBV of that lobe and the nonvascular tissue volume (BV5/T(issue)V) to calculate lobe-specific BV5/TBV and BV5/T(issue)V ratios. Densitometric measures of emphysema were obtained using a Hounsfield unit threshold of -950 (%LAA-950). Measures of chronic obstructive pulmonary disease severity included single breath measures of diffusing capacity of carbon monoxide, oxygen saturation, the 6-minute-walk distance, St George's Respiratory Questionnaire total score (SGRQ), and the body mass index, airflow obstruction, dyspnea, and exercise capacity (BODE) index. Measurements and main results: The %LAA-950 was inversely related to all calculated vascular ratios. In multivariate models including age, sex, and %LAA-950, lobe-specific measurements of BV5/TBV were directly related to resting oxygen saturation and inversely associated with both the SGRQ and BODE scores. In similar multivariate adjustment lobe-specific BV5/T(issue)V ratios were inversely related to resting oxygen saturation, diffusing capacity of carbon monoxide, 6-minute-walk distance, and directly related to the SGRQ and BODE. Conclusions: Smoking-related chronic obstructive pulmonary disease is characterized by distal pruning of the small blood vessels (<5 mm(2)) and loss of tissue in excess of the vasculature. The magnitude of these changes predicts the clinical severity of disease.
Full-text available
Pulmonary hypertension (PH) can result in vascular pruning and increased tortuosity of the blood vessels. In this study we examined whether automatic extraction of lung vessels from contrast-enhanced thoracic computed tomography (CT) scans and calculation of tortuosity as well as 3D fractal dimension of the segmented lung vessels results in measures associated with PH. In this pilot study, 24 patients (18 with and 6 without PH) were examined with thorax CT following their diagnostic or follow-up right-sided heart catheterisation (RHC). Images of the whole thorax were acquired with a 128-slice dual-energy CT scanner. After lung identification, a vessel enhancement filter was used to estimate the lung vessel centerlines. From these, the vascular trees were generated. For each vessel segment the tortuosity was calculated using distance metric. Fractal dimension was computed using 3D box counting. Hemodynamic data from RHC was used for correlation analysis. DISTANCE METRIC, THE READOUT OF VESSEL TORTUOSITY, CORRELATED WITH MEAN PULMONARY ARTERIAL PRESSURE (SPEARMAN CORRELATION COEFFICIENT: ρ = 0.60) and other relevant parameters, like pulmonary vascular resistance (ρ = 0.59), arterio-venous difference in oxygen (ρ = 0.54), arterial (ρ = -0.54) and venous oxygen saturation (ρ = -0.68). Moreover, distance metric increased with increase of WHO functional class. In contrast, 3D fractal dimension was only significantly correlated with arterial oxygen saturation (ρ = 0.47). Automatic detection of the lung vascular tree can provide clinically relevant measures of blood vessel morphology. Non-invasive quantification of pulmonary vessel tortuosity may provide a tool to evaluate the severity of pulmonary hypertension. NCT01607489.
Full-text available
Computed tomography (CT) is the modality of choice for imaging the lungs in vivo. Sub-millimeter isotropic images of the lungs can be obtained within seconds, allowing the detection of small lesions and detailed analysis of disease processes. The high resolution of thoracic CT and the high prevalence of lung diseases require a high degree of automation in the analysis pipeline. The automated segmentation of pulmonary structures in thoracic CT has been an important research topic for over a decade now. This systematic review provides an overview of current literature. We discuss segmentation methods for the lungs, the pulmonary vasculature, the airways, including airway tree construction and airway wall segmentation, the fissures, the lobes and the pulmonary segments. For each topic, the current state of the art is summarized, and topics for future research are identified.
Full-text available
The purpose of this paper is to present an automated method for the extraction of the pulmonary vessel tree from multi-slice CT data. Furthermore we investigate a method for the separation of pulmonary arteries from veins. The vessel tree extraction is performed by a seed-point based front-propagation algorithm. This algorithm is based on a similar methodology as the bronchial tree segmentation and coronary artery tree extraction methods presented at earlier SPIE conferences. Our method for artery/vein separation is based upon the fact that the pulmonary artery tree accompanies the bronchial tree. For each extracted vessel segment, we evaluate a measure of "arterialness". This measure combines two components: a method for identifying candidate positions for a bronchus running in the vicinity of a given vessel on the one hand and a co-orientation measure for the vessel segment and bronchus candidates. The latter component rewards vessels running parallel to a nearby bronchus. The spatial orientation of vessel segments and bronchi is estimated by applying the structure tensor to the local gray-value neighbourhood. In our experiments we used multi slice CT datasets of the lung acquired by Philips IDT 16-slice, and Philips Brilliance 40-slice scanners. It can be shown that the proposed measure reduces the number of pulmonary veins falsely included into the arterial tree.
Full-text available
Distinguishing pulmonary arterial and venous (A/V) trees via in vivo imaging is a critical first step in the quantification of vascular geometry for the purpose of diagnosing several pulmonary diseases and to develop new image-based phenotypes. A multiscale topomorphologic opening (MSTMO) algorithm has recently been developed in our laboratory for separating A/V trees via noncontrast pulmonary human CT imaging. The method starts with two sets of seeds-one for each of A/V trees and combines fuzzy distance transform and fuzzy connectivity in conjunction with several morphological operations leading to locally adaptive iterative multiscale opening of two mutually conjoined structures. In this paper, we introduce the methods for handling "local update" and "separators" into our previous theoretical formulation and incorporate the algorithm into an effective graphical user interface (GUI). Results of a comprehensive evaluative study assessing both accuracy and reproducibility of the method under the new setup are presented and also, the effectiveness of the GUI-based system toward improving A/V separation results is examined. Accuracy of the method has been evaluated using mathematical phantoms, CT images of contrast-separated pulmonary A/V casting of a pig's lung and noncontrast pulmonary human CT imaging. The method has achieved 99% true A/V labeling in the cast phantom and, almost, 92-94% true labeling in human lung data. Reproducibility of the method has been evaluated using multiuser A/V separation in human CT data along with contrast-enhanced CT images of a pig's lung at different positive end-expiratory pressures (PEEPs). The method has achieved, almost, 92-98% agreements in multiuser A/V labeling with ICC for A/V measures being over 0.96-0.99. Effectiveness of the GUI-based method has been evaluated on human data in terms of improvements of accuracy of A/V separation results and results have shown 8-22% improvements in true A/V labeling. Both qualitative and quantitative results found are very promising.
Conference Paper
We present a novel algorithm for the simultaneous segmentation and anatomical labeling of the cerebral vasculature. The method first constructs an overcomplete graph capturing the vasculature. It then selects and labels the subset of edges that most likely represents the true vasculature. Unlike existing approaches that first attempt to obtain a good segmentation and then perform labeling, we jointly optimize for both by simultaneously taking into account the image evidence and the prior knowledge about the geometry and connectivity of the vasculature. This results in an Integer Program (IP), which we solve optimally using a branch-and-cut algorithm. We evaluate our approach on a public dataset of 50 cerebral MRA images, and demonstrate that it compares favorably against state-of-the-art methods.
We propose a novel approach to automated delineation of linear structures that form complex and potentially loopy networks. This is in contrast to earlier approaches that usually assume a tree topology for the networks. At the heart of our method is an Integer Programming formulation that allows us to find the global optimum of an objective function designed to allow cycles but penalize spurious junctions and early terminations. We demonstrate that it outperforms state-of-the-art techniques on a wide range of datasets.
Purpose: This paper introduces a novel approach to classify pulmonary arteries and veins from volumetric chest computed tomography (CT) images. Although there is known to be a relationship between the alteration of vessel distributions and the progress of various pulmonary diseases, there has been relatively little research on the quantification of pulmonary vessels in vivo due to morphological difficulties. In particular, there have been few efforts to quantify the morphology and distribution of only arteries or veins through automated algorithms despite the clinical importance of such work. In this study, the authors classify different types of vessels by constructing a tree structure from vascular points while minimizing the construction cost using the vascular geometries and features of CT images. Methods: First, a vascular point set is extracted from an input volume and the weights of the points are calculated using the intensity, distance from the boundaries, and the Laplacian of the distance field. The tree construction cost is then defined as the summation of edge connection costs depending on the vertex weights. As a solution, the authors can obtain a minimum spanning tree whose branches correspond to different vessels. By cutting the edges in the mediastinal region, branches can be separated. From the root points of each branch, the cut region is regrouped toward the entries of pulmonary vessels in the same framework of the initial tree construction. After merging branches with the same orientation as much as possible, it can be determined manually whether a given vessel is an artery or vein. Our approach can handle with noncontrast CT images as well as vascular contrast enhanced images. Results: For the validation, mathematical virtual phantoms and ten chronic obstructive pulmonary disease (COPD) noncontrast volumetric chest CT scans with submillimeter thickness were used. Based on experimental findings, the suggested approach shows 9.18 ± 0.33 (mean ± SD) visual scores for ten datasets, 91% and 98% quantitative accuracies for two cases, a result which is clinically acceptable in terms of classification capability. Conclusions: This automatic classification approach with minimal user interactions may be useful in assessing many pulmonary disease, such as pulmonary hypertension, interstitial lung disease and COPD.