automated segmentation and morphometric analysis of the ......data with computational models. as...

493The Journal of Physiological Sciences Vol. 58, No. 7, 2008

REGULAR PAPER

Automated Segmentation and Morphometric Analysis of the Human Airway Tree from Multidetector CT Images

Masanori NAKAMURA1, Shigeo WADA2, Takahito MIKI3, Yasuhiro SHIMADA4,Yuji SUDA5, and Gen TAMURA6

1The Center for Advanced Medical Engineering and Informatics, Osaka University, Toyonaka, 560-8531 Japan; 2Department of Mechanical Science and Bioengineering, Graduate School of Engineering Science, Osaka University, Toyonaka, 560-8531 Japan; 3Graduate School of Biomedical Engineering, Tohoku University, Sendai, 980-8579 Japan; 4Graduate School of Engineering,

Kyoto University, Kyoto, 606-8051 Japan; 5Sendai City Medical Center, Sendai, 983-0824 Japan; and6Airway Institute in Sendai Co. Ltd., Sendai, 980-0871 Japan

J. Physiol. Sci. Vol. 58, No 7; Dec. 2008; pp. 493–498Online Dec. 6, 2008; doi:10.2170/physiolsci.RP007408

Abstract: Remarkable advances in computed tomography (CT) technology geared our research toward investigating the inte-grative function of the lung and the development of a database of the airway tree that incorporates anatomic and functional data with computational models. As part of this project, we are developing the algorithm to construct an anatomically realistic geometric model of airways from CT images. The basic concept of the algorithm is to segment as many airway trees as possible from CT images and later correct quantified parameters based on CT values. CT images are acquired with a 64-channel multi-detector CT, and the airway is then extracted from them by the region-growing method while maintaining connectivity. Using this method, we extracted 428 airways up to the 14th branch-

ing generation. Although the airway diameters up to the 4th generation showed good agreement with those reported in an autopsy study, those in later generations were all greater than the reported values because of the limited resolution of the CT images. We corrected the errors in diameters by assess-ing the relationship between the diameter and median value of Hounsfield unit (HU) intensity of each airway; consequently, the diameters up to generation 8 agreed well with the reported val-ues. Based on these results, we concluded that the use of HU-based correction algorithm combined with rough segmentation can be another way to improve data accuracy in the morpho-metric analysis of airways from CTs.

Key words: airway, automated segmentation, morphometric analysis, computed tomography.

Computed tomography (CT) is an important imaging modality. A 64-channel multidetector CT that is now broadly used is capable of simultaneously acquiring 64 slices of CT images 0.5 mm thick with each 400 ms gan-try revolution, yielding precise isotropic imaging of any region of the body within a single holding of breath. The helical imaging function allows real-time 3D reconstruc-tion of the organs.

With these advancements in CT technology, many imaging studies of both lung structure and lung function have been conducted [1–11]. For example, Tschirren et al. [10] presented algorithms that perform both the matching of branch points between two human airway trees and the assignment of anatomical names to the segments and branch points of the human airway tree without human intervention. Palágyi et al. [11] developed a skeletoniza-tion method of 3-dimensional tubular structures and reported subvoxel accuracy of branch point positioning,

insensitivity to changes of object orientation, and a high degree of reproducibility of derived quantitative indexes of the tubular structures. Wiemker et al. [12] presented a review of the work of other research groups in this area. Although these studies indicated the potential of CT im-aging, the accuracy of quantitative results is still limited by the CT resolution.

The aim of our physiome project is to investigate the integrative function of lung and to develop a database of the airway tree that incorporates anatomical and func-tional data with mathematical and computational models. The ultimate goal of this project is threefold:

1. To build an international database of lung anatomy to develop a more realistic lung model.

2. To study pulmonary f low dynamics to assess and predict particle deposition in relation to the subject’s anatomy.

3. To develop a visualization tool of the airway tree

Received on May 8, 2008; accepted on Dec 3, 2008; released online on Dec 6, 2008; doi:10.2170/physiolsci.RP007408Correspondence should be addressed to: Masanori Nakamura, the Center for Advanced Medical Engineering and Informatics,Osaka University, Toyonaka, Osaka, 560-8531 Japan. Tel: +81-6-6850-6173, Fax: +81-6-6850-6172,E-mail: [email protected]

494

M. NAKAMURA et al.

The Journal of Physiological Sciences Vol. 58, No. 7, 2008

in terms of 3D spatial distribution to provide better diagnostic methods.

To accomplish these goals, we must develop a method for automatic segmentation of a real airway tree to con-struct an anatomically realistic model of the airways. As described above, however, all conventional algorithms have diffi culties in accurately quantifying small airways because of the limited spatial resolution of CT images, which prompted us to develop a novel algorithm to ad-dress this problem.

This paper has two main sections. First, we brief ly describe the algorithm to segment the airway tree from CT images and to analyze morphometry and topology of this tree. The geometric model of the airway tree and measured diameters are presented as an example of an anatomic database. Also, we reveal that our algorithm inherently overestimates airway diameters. Based on this result, in the second section we introduce the algorithm to correct the diameters of lung airways by using the CT value of each airway. The corrected diameters were com-pared to those of a previous autopsy study by Weibel [13] to demonstrate the potential of the correcting algorithm.

1. Algorithm of segmenting and quantifying airways from CT images

1.1. CT scanning. CT scanning of the human lung was carried out using an Aquilion 64 (Toshiba Medical Systems Corp., Tokyo, Japan) with a helical scan option. All scans were performed at suspended end-inspiratory volume and under synchronization with an electrocar-diogram (ECG) to minimize errors derived from cardiac motion. Each CT image consisted of a 512 × 512 matrix (pixel size 0.5 mm) of numerical data in Hounsfi eld units (HUs). No contrast medium was used. A 0.5 mm collima-tion was selected to provide high resolution in the axial direction, yielding 513 slices for adequate coverage of the lung. The size of the resultant voxel, obtained by stack-ing CT images vertically, was 0.125 mm3. The Cartesian coordinates (x, y, z) were defined for stacked CT images with the z-axis defined in the direction of the body axis. Because of this, the positions of all voxels are described in coordinates or positional vectors.

1.2. Airway segmentation. In CT images, the airway lumen appears as a black area (air) surrounded by a white ring (airway wall). A region-growing method [14] was used to extract the airway from CT images. In this meth-od, the region of interest (ROI) is evolved gradually from a seed area by changing the threshold value of HU until a drastic change in the area of ROI occurs. Here we modi-fi ed the region-growing method so that airway connectiv-ity is secured and continuity in the change in the cross-sectional area of the airway lumen is ensured.

The fi rst seed was manually given to a voxel where the trachea was clearly recognized, and the region-growing method was applied to obtain the lumen of trachea. The

obtained region was projected onto the next CT image to give a hint of a seed area in the next CT image. Given the seed area, an airway region was expanded by the region-growing method. Note that to segment as many airways as possible and to minimize the loss of airways at this stage, we changed the threshold in accordance with HU values of the seed area projected from the previous im-ages. This is exclusive to the current algorithm, since threshold is unchanged in most studies. Although the pro-cedure noted above was for extracting the trachea lumen in the next section, given the lumen data in the initial sec-tion, the same procedure was applied for other airways. The segmentation of the airway was first performed or-thogonally to a transverse section (neck to stomach). Sub-sequently, the airway tree was segmented orthogonally to coronal and sagittal sections. This process was repeated in the order of segmentation orthogonally to transverse sections, coronal sections, and sagittal sections until no more voxels were obtained. Because the segmentation is carried in three directions, we call this method multidi-rectional segmentation. On the other hand, if the segmen-tation is carried only once in one direction, for instance, orthogonally to transverse sections, we call it monodirec-tional segmentation.

1.3. Skeletonization. To gain topological insight into the airways, we skeletonized the airway model. In brief, airways segmented from CT images are trimmed from the outer airway surfaces while airway connectivity is maintained. There are two possible criteria to judge the connection of voxels: one is to consider surface contact between the surfaces of two voxels (face-to-face contact), and the other is to consider not only surface contact, but also edge contact between them (edge-to-edge contact). Here we adopted only the former criteria (face-to-face contact). After skeletonization, branching points where a parent airway bifurcates into two daughter airways were defi ned at the center of voxels that have three face-to-face contacts with skeletonized branches. The 3D skeletoniza-tion algorithm is sensitive to surface irregularities of the model, which result in pseudobranches. In the present study, they were pruned manually.

1.4. Labeling of the generation number and pulmonary segment in the airway model. The global generation num-ber Gg, based on Weibel’s naming convention [14], was assigned to each branch. The global generation number Gg corresponds to the number of branches that should be in that generation as far as all airways bifurcate from the trachea, and there is no termination of airway branches. In this convention, generation zero is given to the trachea (Gg = 0), one is given to its daughter branches, two to their daughters, and so on.

A pulmonary segment (S1–S10), to which each branch belongs, was identified from its positional vectors. To each branch, pulmonary-segment-based generation num-ber Gp was also defi ned. This number Gp is similar to the

495

Morphometric Analysis of the Human Airway Tree


global generation number Gg, except that zero is defi ned at the root airway in each pulmonary segment. The naming convention from the root branch to its daughter branches is the same as that of the global generation number.

1.5. Quantification. The volume V, length L, and diam-eter D of each airway branch were quantifi ed. Volume V was calculated from the number of voxels forming each airway branch. The length L was approximated as the distance between the terminal branching points. The diameter of airway D was estimated under the assump-tion that the airway was truly cylindrical. The branching angle between the parent airway and its daughter airway was estimated by

(1)where Tp is a directional vector of the parent airway, Td is that of the daughter airway, a dot means an inner product, and | | represents the norm or length of a vector. Here the directional vector of an airway branch was approximated from a line segment that connects a bifurcating point at the upstream side (parent airway side) with that at the downstream side.

2. Segmentation and quantifi cation of the airway tree from CT images

An adult male volunteer with no history of cardiac dis-ease was examined in this study. A written informed con-sent was obtained from the volunteer to participate in the study with the approval of Sendai Open Hospital. Figure 1 illustrates frontal and lateral views of a 3-dimensional air-way model reconstructed from the CT images obtained. Here each airway is color-coded based on its generation number, showing that the global anatomy of the airway tree was well captured. For this subject, it was possible to segment 428 airway branches up to and including the 14th generation in terms of the global generation number; this was almost 2.4 times as many as in the recent report

by Palágyi et al. [11].Figure 2 shows a plot of the number of airways ex-

tracted for each generation. In this fi gure, the bars on the left represent the number of airways extracted by the fi rst segmentation where the airways were segmented once along the body axis (monodirectional segmentation), and those on the right represent the airways gained by mul-tidirectional segmentation. The number of airways ex-tracted increased markedly by using the multidirectional segmentation. The number was maximal at the 7th gen-eration and decreased with further generations. Because a bifurcating process can be assumed for the airway, we successfully segmented all airways through the 5th gen-eration, but missed some from the 6th generation onward. In the following context, all data presented are from this subject.

Table 1 summarizes the number of airway branches

Fig. 1. Frontal and lateral views of an airway model labeled with the global generation number.

Fig. 2. Number of airways extracted for each generation by monodirectional segmentation along the body axis and by multidirectional segmentation.

θ = ⋅⎛

⎝⎜⎞

⎠⎟−cos 1 T T

T Tp d

p d

496

M. NAKAMURA et al.


N, the maximum pulmonary-segment-based generation number Gp

max, the minimum diameter Dmin, and the mean branching angle θ in each pulmonary segment. Although the number of airways obtained for each pulmonary seg-ment varied from 2 to 40, the minimum diameter was almost the same (0.99–1.77 mm). This does not mean that the minimum diameter of terminal airways in all pulmo-nary segments was actually the value of Dmin. It means that the airways with diameters of less than Dmin were un-able to be segmented because of the limited spatial reso-lution of current CT images. The average branching angle θ ranged from 30° to 45°, and no significant differences were found between pulmonary segments. These angle data were almost the same as those reported in Horsfi eld and Cumming [15].

The diameters obtained based on the present results were compared with the widely accepted standard data reported previously by Weibel [13], who performed morphometric analysis of the airway from a resin cast of a human cadaver (Fig. 3). Note that the subject of Weibel [13] is not the same as the one used in this study. Furthermore, CTs in this study were obtained at the end-inspiration phase, and the resin cast was made at 3/4 maximal inf lation [13]. In Fig. 3, the diameter at each generation was normalized with that of the trachea. Be-cause explanations on the corrected data presented in Fig. 3 are given in the next section, we here focus on the rela-tion of the data obtained with the segmentation algorithm (original data) to Weibel’s data. The results obtained were in good agreement with those of Weibel [13] up to generation 4, whereas the diameters of generation 5 and later were all larger than those reported previously. This overestimation was obviously attributable to a loss of airways in later generations caused by the limited spatial resolution of CT imaging. Because the spatial resolution of CT was 0.5 mm, resolving the airways with diameters

smaller than this value was impossible. This would also be clear from Table 1. Conversely, it was possible to resolve up to the 4th generation with good accuracy.

3. Correcting the measured diameters of the airway tree based on HU

As described in Section 1.1, the airway was segmented based on the HU of voxels. HU is a representation of the amount of the X-ray beam absorbed by the tissues at each individual point in the body. In theory, the HU of a material varies according to its density of the tissue, with denser tissues having higher values: HU values for

Table 1. Summary of measured data in each pulmonary segment (PS).

Left lung Right lung

PS Gpmax N Dmin θ PS Gp

max N Dmin θ

S1 + 2 6 29 1.33 28.7 S1 3 12 1.77 29.6 S3 7 26 1.67 35.0 S2 4 12 1.65 34.7 S4 2 4 1.49 40.5 S3 7 38 1.43 32.3 S5 5 22 1.61 31.2 S4 4 16 1.51 31.1 S6 5 26 0.99 40.3 S5 8 28 1.12 30.1 S7 – – – – S6 5 22 1.29 37.1 S8 6 26 1.41 40.9 S7 1 2 1.38 42.4 S9 3 8 1.36 39.4 S8 7 24 1.29 46.8 S10 6 35 1.33 35.0 S9 5 22 1.44 26.3 S10 7 40 1.15 41.0

Gpmax, maximum pulmonary-segment-based generation number; N, number of airways; Dmin, minimum diameter (mm);

θ, mean branching angle (deg). S1 + 2, apicoposterior segment, apical segment; S2, posterior segment; S3, anterior segment; S4, superior lingular segment, lateral segment; S5, inferior lingular segment, medial segment; S6, superior segment; S7, medial basal segment; S8, anterior basal segment; S9, lateral basal segment; S10, posterior basal segment.

Fig. 3. Bar plots of the airway diameters (mean ± SD). Data show the originally measured (gray) and corrected (white) airway diameters. The airway diameters are compared to those reported by Weibel [13]. Note that the diameter at each generation was normalized relative to that of the tra-chea.

497

Morphometric Analysis of the Human Airway Tree


was used for estimating parameters. The results of re-gression analysis are shown by the black line in Fig. 5. In this fitting, α , β, and HUair were estimated as 118.3, –0.779, and –996, respectively. The diameter D of each airway was then corrected by

(3)if the diameter was smaller than 4 mm and HUm was larger than HUair.

Corrected diameters were superimposed onto Fig. 3 so as to compare with Weibel [13]. As seen in the fi gure, the corrected diameters were closer to those reported by Wei-bel [13], showing improvement especially for generations 4–8 and demonstrating the potency of this algorithm. It should be noteworthy, however, that diameters for later than generation 8 were still larger than Weibel [13]. It is

Fig. 5. Plot of the median of H. HUm plotted against the di-ameter D of each airway. Circles represent the airway that consisted of more than 64 voxels, whose length was larger than 4 mm, and whose HUm and the mean of HU both lay between –1,040 and –960. The line was fi tted for circles.

air and water are –1,000 and 0, respectively. However, because of the limitations of spatial resolution, HU of the pixel of interest is affected by that of its neighboring pixels, which results in the contamination of HU, espe-cially in regions of varying composition, such as near the airway wall. In the present study, a threshold value of HU for segmenting airways from CT images was adjusted ac-cording to circumstantial HU to maximize the extraction of as many airways as possible. As a consequence, we anticipate that the airway lumen will be overevaluated, leading to overestimation of the airway diameter.

Figure 4 shows normalized histograms of HU of the airway at 0 generation (trachea) and an airway at the 10th generation, in which the number of voxels for each HU level is divided by the total number of voxels compris-ing the airway of interest. For the trachea, the histogram is quite sharp, and the median of HU, HUm, is approxi-mately –972. HUm at the trachea does not correspond to that of the air, which is probably a result of moisture in the air. On the other hand, for the airway at the 10th gen-eration, the histogram is skewed slightly to the right with no single sharp peak, and HUm is –933. Figure 5 shows the measured diameter D plotted against the median of the HU of each airway, HUm. The fi gure shows that HUm became more diverse and larger as the airway diameter became smaller. This was because the proportion of vox-els whose HU was affected by the airway wall increased as the airway became narrower.

The measured diameters could be corrected by cor-recting the error of HUm. To determine the relationship between the median of airway HUm and the diameter D, regression analysis was performed using

(2)where α and β are constants, and HUair is the HU of air. Fitting was performed for the airways that consisted of more than 64 voxels, whose length was larger than 4 mm, and whose HUm and the mean of HU both lay between –1,040 and –960. The Levenberg-Marquardt algorithm

0 2 4 6 8 10 12 14 16 18-1050

-1000

-950

-900

-850

Diameter (mm)

HU

m

′ = −⎛⎝⎜

⎞⎠⎟

D1

βlog

HU HUm air

Fig. 4. Normalized histograms of HU of the airway at (a) 0 generation and (b) the 10th generation.

ba

α

HU exp HUm air= ( ) + βDα

498

M. NAKAMURA et al.


attributed to imperfect segmentations of airways in those generations, as seen in Fig. 2, because of insufficient resolutions to resolve small airways. Because airways that were not segmented were small ones, diameters in later generations tended to be larger than Weibel [13]. Another reason, though minor, is a difference in lung state when diameters were quantifi ed; we took CTs at the end-inspiration phase, though Weibel [13] made the resin cast at 3/4 maximal infl ation. Future advancement in CT technology will facilitate higher resolutions of peripheral airways and thereby contribute to increasing the accuracy in morphometric analyses of lung.

4. Summary and conclusionA morphometric analysis of a human airway tree

was carried out. The results indicated that it was possible to extract up to and including the 14th generation of the airway tree. The airway diameters obtained here were in good agreement with those in Weibel [13] through genera-tion 4. The errors in diameter were corrected by assessing the relationship between the diameter and median value of HU intensity of each airway; consequently, the diameters through generation 8 agreed well with the reported values. Based on these results, we conclude that the use of HU-based correction algorithm combined with rough segmen-tation can be another way to improve data accuracy in the morphometric analysis of airways from CTs.

This work was supported in part by the Global COE Program “in silico medicine” at Osaka University. It was also supported by the Computa-tional Science Research Program, Integrated Simulation of Living Matter Group, Japan. The authors wish to acknowledge Professor Takami Yama-guchi for his useful comments on this work.

REFERENCES

1. Sauret V, Goatman KA, Fleming JS, Bailey AG. Semi-automated tabulation of the 3D topology and morphology of branching networks using CT: application to the airway tree. Phys Med Biol. 1999;44:1625-38.

2. Sauret V, Halson PM, Brown IW, Fleming JS, Bailey AG. Study of the three-dimensional geometry of the central conducting airways in man using computed tomographic (CT) images. J Anat. 2002;200:123-34.

3. Sonka M, Park W, Hoffman EA. Rule-based detection of intrathoracic airway trees. IEEE Trans Med Imaging. 1996;15:314-26.

4. Park W, Hoffman EA, Sonka M. Segmentation of intrathoracic airway trees: a fuzzy logic approach. IEEE Trans Med Imaging. 1998;17:489-97.

5. Palágyi K, Tschirren J, Sonka M. Quantitative analysis of intrathoracic airway trees: methods and validation. Inf Process Med Imaging. 2003;18:222-33.

6. Hoffman EA, Reinhardt JM, Sonka M, Simon BA, Guo J, Saba O, Chon D, Samrah S, Shikata H, Tschirren J, Palágyi K, Beck KC, McLennan G.

Characterization of the interstitial lung diseases via density-based and texture-based analysis of computed tomography images of lung structure and

function. Acad Radiol. 2003;10:1104-18. 7. Suter M, Tschirren J, Reinhardt J, Sonka M, Hoffman E, Higgins W,

McLennan G. Evaluation of the human airway with multi-detector x-ray-computed tomography and optical imaging. Physiol Meas. 2004;25:837-47.

8. Tschirren J, Hoffman EA, McLennan G, Sonka M. Intrathoracic airway trees: segmentation and airway morphology analysis from low-dose CT scans. IEEE Trans Med Imaging. 2005;24:1529-39.

9. Tschirren J, Hoffman EA, McLennan G, Sonka M. Segmentation and quantitative analysis of intrathoracic airway trees from computed tomography images. Proc Am Thorac Soc. 2005;2:484-7, 503-4.

10. Tschirren J, McLennan G, Palágyi K, Hoffman EA, Sonka M. Matching and anatomical labeling of human airway tree. IEEE Trans Med Imaging.

2005;24:1540-7. 11. Palágyi K, Tschirren J, Hoffman EA, Sonka M. Quantitative analysis of

pulmonary airway tree structures. Comput Biol Med. 2006;36:974-96. 12. Wiemker R, Rogalla P, Blaffert T, Sifri D, Hay O, Shah E, Truyen R, Fleiter T. Aspects of computer-aided detection (CAD) and volumetry of pulmonary

nodules using multislice CT. Br J Radiol. 2005;78:S46-56. 13. Weibel ER. Morphometry of the human lung. New York: Springer-Verlag; 1963. 14. Kiraly AP, Higgins WE, McLennan G, Hoffman EA, Reinhardt JM. Three-

dimensional human airway segmentation methods for clinical virtual bronchoscopy. Acad Radiol. 2002;9:1153-68.

15. Horsfield K, Cumming G. Angles of branching and diameters of branches in the human bronchial tree. Bull Math Biophys. 1967;29:245-59.

automated segmentation and morphometric analysis of the ......data with computational models. as...

Documents