1b)
Division of Physiologic Imaging
Department of Radiology
University of Iowa College of Medicine
Iowa City, IA 52242
A critically important component in the development of new methods
for treatment of pulmonary diseases is the development of sensitive
techniques for assessing alterations in regional lung structure and function.
We describe an automated method for
segmentation of airway trees from three-dimensional sets of CT images.
The method is based on a combination of conventional
three-dimensional seeded region growing
that is used to identify large airways,
knowledge-based
two-dimensional segmentation of individual CT slices to identify
probable locations of smaller diameter airways, and
merging of airway regions across the three-dimensional set of
slices resulting in a tree-like airway structure.
The preliminary validation of the method was done in eighty
3 mm thick CT sections from
two 40 slice data sets of a canine thorax scanned with lungs held at
1.5~kPa and 2.5 kPa.
The method's performance was compared with that of
the conventional three-dimensional region growing method.
The knowledge-based approach to identification of potential
airways in individual image slices substantially outperforms
the conventional method and promises to be applicable to
in vivo pulmonary CT images.
A critically important component in the development of new methods for treatment of pulmonary diseases is the development of sensitive methods for assessing alterations in regional lung structure and function. Global pulmonary function tests are often too insensitive to allow for assessment of small, incremental improvements. Likewise, the conventional A-P chest roentgenogram suffers from the same limitations. Computer tomography (CT) and particularly high resolution CT (HRCT) eliminates superposition of structures and has advanced the ability of the radiologist to assess regional abnormalities 1. Semi-quantitative scoring systems have been developed to tabulate the visual assessment of the CT image data sets and such systems have proven to be useful in assigning measures of disease severity to an individual patient 2. Pulmonary complications of cystic fibrosis are the major source of debilitating effects and ultimately the major cause of death 3. Objective measures of disease are critical to the assessment of medical therapy for cystic fibrosis and other pulmonary diseases. Since comparison of small changes in intrathoracic airways resulting over time from treatment like the gene therapy can only be reliable if a quantitative scoring system is developed, automated or semi-automated detection of airway trees from HRCT image data is of paramount importance. Furthermore, a method developed to track airway structure over multiple lung volumes and across pharmacologic interventions will allow for advances in our understanding of a broad based set of physiologic and patho-physiologic conditions.
In this paper, we describe an automated method for segmentation of airway trees from three-dimensional sets of CT images. The method is based on a combination of the following approaches: 1) conventional three-dimensional seeded region growing used to identify large airways; 2) knowledge-based two-dimensional segmentation of individual CT slices to identify probable locations of smaller diameter airways; and 3) merging of airway regions found in 1) and 2) across the three-dimensional set of slices resulting in a tree-like airway structure. The method has two major stages. In the first stage, potential airway regions are identified in each CT slice. In the second stage, these regions are either accepted as airways or rejected using knowledge about the 3D airway connectivity and expected airway tree properties. In the following paragraphs, we will concentrate on methods to identify potential airways in individual slices.
The main strategy of our automated airway tree segmentation method is a reliable detection of larger branches, knowledge-based detection of potential smaller diameter airways, and identification of additional tree branches using a priori anatomical knowledge about airway tree structure. Our method for three-dimensional airway tree detection consists of several steps
1. Primary airway tree definition as that portion of the tree which
consists of major tree branches.
2. Preprocessing of individual CT image slices.
3. Airway and vessel detection in individual image slices.
4. Airway tree construction using a knowledge-based
tree-growing strategy.
Experience shows that the primary airways fall within a tight gray value range in CT images and differ significantly from the airway walls, thus exhibiting high contrast borders. Gray-level based techniques using three-dimensional seeded region growing 4, 5, 6, 7 work well in major pulmonary tree branches. Using 3-D connectivity, a contiguous volume connected to the manually identified location in the mainstem bronchus is extracted and represents the primary tree. The gray level threshold used for region growing must be quite conservative to guarantee that no non-tree structure will become a part of the primary tree. Experience shows that gray level thresholds from a wide interval satisfy this requirements.
1a)
1b)
Figure 1: HRCT image of the canine lung.
were selected to segment airways, background, and vessels.
As shown here, no
thresholds can be selected to yield acceptable segmentation.
2a)
2b)
2c)
2d)
Figure 2: Preprocessing of individual CT image slices
and rule-based detection of airways and vessels in CT images
(see Fig. 1 for original).
After the lungs are segmented from the HRCT images using either semi-automated threshold-based approach or automated mask-based approach such as described in 8, potential locations of airways are identified using a knowledge-based procedure. The goal of the preprocessing step is to identify all possible locations of airways and vessels in individual CT slices. Identification of the airways in individual CT slices starts with image preprocessing. The image is zoomed 2x (Fig. 1a) to make the small airways large enough to be enhanced using scale-sensitive 13x13 hat transform 9. Since the kernel size was designed to enhance small and medium diameter airways and vessels, large diameter airways and vessels are not well represented in the hat-transformed image (Fig. 2a). Large diameter airways and vessels are identified using simple gray level threshold criteria and are added to the hat-transformed images. The preprocessed image is segmented using edge-based region growing 10, 11, in which the gray level of each region reflects the gray level of the region growing seed point (Fig. 2b). Therefore, regions with potential correspondence to airways tend to be dark and vessel regions bright. As such, regions with intermediate gray levels are considered background. Only intermediate gray levels are modified
gray = 128 IF (min_background < gray < max_background)Adjacent regions are merged if their gray levels differ by less than a pre-selected value. The preprocessing results in a substantially simplified image suitable for knowledge-based analysis (Fig. 2c.). Note, that similar results cannot be achieved by direct application of gray-level based segmentation no matter how carefully the airway and vessel thresholds are selected (Fig. 1b). In contrary, the segmentation shown in Fig. 2c is quite insensitive to the selection of thresholds.
In CT images, vessels appear generally brighter than the
lung tissue and are thus relatively easy to identify. Airways usually
run parallel with vessels, and if large enough, they are black since
the air is well distinguished from the airway wall and adjacent
vessels. Unfortunately, smaller airways often have gray levels
quite similar to those of the lung tissue due to partial volume
effects. Therefore, it is more difficult to identify peripheral airways than
to detect vessels. This forms the rationale
for our airway detection strategy. We utilize the knowledge about the
pairwise appearance of vessels and airways, identify vessels first,
and look for airways only in proximity to the detected vessels.
Regions which were not previously marked
as background during the preprocessing
stage are either
accepted as potential vessels (and assigned a vessel label),
accepted as potential airways (and assigned an airway label),
or rejected (and assigned a background label).
Additionally, each potential airway and vessel region is accompanied
with a confidence rating based on a priori knowledge about
pulmonary anatomy.
The label assignment and confidence rating is performed in region adjacency graphs using a rule-based system. In the region adjacency graph, regions correspond to graph nodes and adjacency is represented by arcs between nodes 9. The region adjacency graph also carries the information about original CT gray-level, shape, and size properties of each image region 12, 13. Using the rule-based labeling approach, the vessel regions are identified first and used to guide the airway identification. Vessels are always bright in CT lung images. Therefore, each potential vessel region is assigned a vessel label if it has above-threshold gray level properties in the original CT image. Each vessel label is accompanied with a two-grade confidence rating. Using the identified vessels for guidance, airways are identified as dark regions (with below-threshold gray level properties), which are adjacent to a region previously labeled as vessel. Dark regions not adjacent to a region labeled vessel are labeled background. Similarly, each region adjacent to a region labeled vessel, which is larger than pre-specified size is assigned a background label because it most likely does not represent an airway. Each airway is accompanied with a 4-grade confidence rating based on the adjacency to high or low confidence vessels and the region's relative darkness with respect to its neighbors. Airways with highest confidences detected in Fig.1a are shown in Fig. 2d. The currently used set of rules is now given. The following variables represent parameters of the method: max_vessel_size, max_airway_size, max_airway_gray, min_vessel_gray, big_size; the names are self-explanatory. Three inter-regional relations are used in the rules: is-adjacent-to, is-brightest-of-all-neighbors, is-darkest-of-all-neighbors.
region = background IF ((region_size > max_vessel_size)
OR (region_size > max_airway_size)
OR (max_airway_gray < region_brightness < min_vessel_gray))
region = vessel IF (region_gray >= min_vessel_gray)
region = airway IF ((region_gray > max_airway_gray)
AND (region <= is-adjacent-to vessel)
AND (region NOT is-adjacent-to big_vessel))
vessel(confidence) = 1.0 IF (vessel is-brightest-of-all-neighbors)
vessel(confidence) = 0.5 IF (vessel NOT is-brightest-of-all-neighbors)
airway(confidence) = 1.0 IF ((airway is-adjacent-to(vessel(confidence) == 1))
AND (airway is-darkest-of-all-neighbors))
airway(confidence) = 0.8 IF ((airway is-adjacent-to (vessel(confidence) == 0.5))
AND (airway is-darkest-of-all-neighbors))
airway(confidence) = 0.6 IF ((airway is-adjacent-to (vessel(confidence) == 1))
AND (airway NOT is-darkest-of-all-neighbors))
airway(confidence) = 0.4 IF ((airway is-adjacent-to (vessel(confidence) == 0.5))
AND (airway NOT is-darkest-of-all-neighbors))
The presented set of rules is considered preliminary and will be modified and fine-tuned with the increased knowledge about the pulmonary tree.
After the primary tree is merged with the smaller airways identified
in individual slices, the three-dimensionally connected object containing
the primary tree is called the transitory tree. The primary
tree represents the reliable part of the transitory tree.
Since potential smaller airways are used for transitory tree
construction, some parts of the transitory tree may not correspond
to the true airway tree and must be removed, and some parts of the
true airway tree may be missing and must be added. Assigned airway
region confidence ratings are used in the remove/add decision process.
" After filling possible holes inside the transitory airway tree, 3-D skeletonization 14 is used to identify a tree skeleton. Structure of the skeleton is stored in a data structure in which branch points correspond to tree nodes and airway branches correspond to arcs between nodes. Each tree node carries information about its parent node, child nodes, paths between the node and its parent, an its x, y, z coordinates thus making any airway tree traversal possible. If loops or geometrically improbable shapes are detected in the tree data structure, the corresponding skeleton parts are removed from the transitory tree. Using the tree data structure, the transitory tree is explored in a systematic manner to search for missing tree branches. Tree-like structures of appropriate shape that were not connected with the transitory tree are searched in proximity to the transitory tree. If found, the transitory tree data structure is updated and the new subtree added. Using a priori knowledge and airway confidence ratings, each branch is accompanied by a confidence rating. Branches with very low confidences are deleted. All tree merges and splits are realized in the tree data structure. When adding or deleting subtrees, the following anatomical knowledge is utilized: airways do not loop; airways do not change their directions abruptly between branch points; and an abrupt change of direction suggests the existence of a branch point. After no additional branch can be added to the transitory tree, the tree is considered complete. The tree construction stage is discussed in detail in 15.
Validation of computer-determined airway trees is an important part of our work. In this paper, we report preliminary visual validation of the first airway tree detection stage -- the performance of the method to reliably identify airways in individual CT image slices. By design, the airway detection in individual slices is supposed to generate potential airways, making full validation based on comparison of results with manually defined independent standard infeasible at this stage of development. Assessment of correct topology and accurate geometry of the entire airway trees represents the second validation stage and is reported in 15.
The preliminary validation of the method was done in eighty 3 mm thick CT sections from two 40 slice data sets of a canine thorax scanned with lungs held at 1.5 kPa and 2.5 kPa. Image data was acquired from a supine, anesthetized 18 kg dog scanned via an Electron Beam CT (EBCT) scanner (Imatron Corp.; South San Francisco, CA). Prior to the onset of scanning, the dog was hyperventilated so as to reduce alveolar carbon dioxide and thus the drive to breathe. The dog remained apneic with lungs held at the preselected lung volume throughout the scanning period via a servo control system built into the custom designed CWE 9000 respirator (CWE; Ardmore, PA). Images were reconstructed into a 256x256 matrix so that with an 18 cm field of view, in-plane pixel dimensions were 0.73 mm. Image analysis was done in 8-bit gray-scale resolution.
The results of the automated airway detection method were visually
compared to
manually identified positions of airways in
the HRCT slices that served as an independent standard. To manually
identify airway locations, interactive analysis options of the
Volumetric Image Display and Analysis (VIDA) software system
To assess three-dimensional properties of the detected potential airways, the three-dimensional airway image sets were rendered and displayed as a movie showing a rotating airway tree. To compare the performance of our new knowledge-based airway detection method with that of the conventional method, three-dimensional seeded region growing 4, 5, 6 as described in 7 was used for comparison.
Visual validation of the detected potential airways in 80 individual slices from two EBCT volumes showed good correspondence between visually-identified and computer-detected airways. As mentioned earlier, potential airways by design specify airway locations with four levels of confidence. Final airways are identified using information about three-dimensional airway tree continuity 15 and airway confidence rating. Fig. 3 shows example of an original image slice, airways identified using conventional three-dimensional seeded region growing, and airways detected using our knowledge-based approach.
Figure 3: EBCT image of a canine lung.
To demonstrate the three-dimensional properties of the detected
airways, the three-dimensional set of data with identified
potential airways
was rendered and is presented in three views in
Fig. 4. For comparison, the airway tree
detectable using the conventional method is also shown.
Note that although far from perfect,
our airway tree provides substantially more information about
the airway tree structure. Considering the fact that the presented
data will be further processed to connect small gaps and remove non-connectable
segments, the achieved results are very promising.
(242k)
Figure 4: Three-dimensional rendering of airways detected using the conventional three-dimensional seeded region growing approach (left) and our automated knowledge-based airway detection method (right) in three different views. The three-dimensional airway data serve as an input for further processing to remove disconnected segments and to connect small gaps.
Objective measures of disease are critical to the assessment of medical therapy for cystic fibrosis and other pulmonary diseases. The chest radiograph has been the cornerstone of imaging in patients with cystic fibrosis 16, 17. Unfortunately, descriptive abnormalities on the chest radiographs are often vague (line shadows, mottled shadows, ring shadows, large shadows) 17. In addition, inter-observer agreement among readers is variable depending upon previous training and clinical experience. HRCT evaluation of the lung fields in cystic fibrosis patients is superior to the chest radiograph 18, 19. Bhalla et al. 2 have shown that the chest radiograph is unable to detect the presence of bronchiecstasis (irreversible dilatation of the bronchial tree) in 45%of segments where it is demonstrated on HRCT. Even more important, the plain chest X-ray misses 92%of the mucous plugging that is present in HRCT. This is important since massive mucous plugging accounts for 95%of deaths in patients with cystic fibrosis. While there is excellent agreement as to the presence of airway abnormality on HRCT, it is more difficult to obtain agreement as to severity of the abnormality. Therefore, it is of extreme importance to develop an automated or semi-automated method for reliable quantitative analysis of airway dimensions in the entire bronchial tree. The discussion will focus on two areas: 1) the relationship of our approach to other airway tree detection methods; and 2) experimental design issues.
No automated or semi-automated method for airway tree detection has been published to date which has been applied to HRCT images of the in vivo lung. There are several major limitations of HRCT which make the extraction of the entire airway trees from CT image sets difficult 20, 21: 1) In-plane image resolution is insufficient to visualize airways with a diameter smaller than 1 mm and to quantify changes in airway dimensions in airways smaller than 1.5 mm; 2) Greater slice thickness dimensions relative to in-plane pixel dimension cause some discontinuity of bronchi between slices; 3) The three-dimensional image data set does not consist of perfectly contiguous slices due to the variability of pulmonary volume at each CT section. From the mentioned limitations, the first two will be eventually resolved by progress in imaging technology. The third one may be minimized by computer-controlled image acquisition by using scanner triggering at a user-selected lung volume.
Three-dimensional reconstruction of the airway tree has been recently performed from contiguous scanning of thin-section CT data obtained on air-inflated fixed-lung specimens 22, 23. Airway tree segmentation was achieved by thresholding of the CT data. Due to partial volume effect only a few branches were reconstructed. In 7, three-dimensional seeded region growing was used on data from an isolated lobe of a dog lung that was processed under seven static pressures ranging from 0.3 kPa (inspiration) to 2.0 kPa (expiration). After extraction of the airway tree, the central axis of each branch was obtained, filtered, and bifurcation points were detected. Physiological measurements such as airway tree segment lengths, branching angles, and cross-sectional areas of the branches were obtained. Although quite suitable for analysis of the excised lungs, we found the method insensitive for analysis of our in vivo data as was shown in Fig. 3, Fig. 4 using the basically identical method.
As a result of in vivo imaging limitations, the airway tree is often disconnected in several locations and gray levels of airways change substantially with decreased airway diameter due to partial volume effects, limited resolution between scans, and motion artifacts. Therefore, no airway segmentation method based solely on three-dimensional connectivity and absolute gray-level parameters of airways may be successful in vivo. Our knowledge-based method was designed to overcome some of the limitations of CT image acquisition. The knowledge-based approach to identification of potential airways in individual image slices outperforms conventional methods. Followed by the knowledge-based strategy of merging tree elements to build a reliable airway tree, the method promises to be applicable to in vivo images assuming thin slice image acquisition and minimization of motion artifacts by utilizing a constant-lung-volume-triggered image acquisition.
Since the independent standard definition was based on visual analysis of HRCT images, the independent standard is necessarily imperfect due to well known image acquisition limitations and limitations of visual analysis. However, in the non-existence of better independent standards for in vivo images, the independent standard was quite appropriate for the preliminary validation.
A limitation of our experimental setup results from using only canine EBCT data sets. The lung-volume-control-triggering device is not yet available for patients. This data set represented a particularly difficult challenge for our detection algorithm in that the partial volume effects were most likely the worst case scenario. The new release of the Electron Beam CT scanner will have 1.5 mm slice thickness, and other volumetric X-ray CT scanning modalities would be expected to be equal to or better than this in regards to thin slice, small field of view, high resolution capabilities. Furthermore, the 15--18 kg dog thoraces which we are currently using as a test data base would be at the small end of the spectra of patients who would be cooperative enough to undergo such a study without the need for anesthesia. Therefore, larger number of identified airways and more airway tree branch generations can be expected to be identified in adult human HRCT data sets. (Fig. 4)
In judging the achieved results, it should be remembered that our detected airways only represent an intermediate stage of processing. The missing airway segments may be added and extra segments removed during the subsequent airway tree construction which uses the knowledge-based tree-growing strategy. It was shown that most of the airways were detected in the images. The method may be fine-tuned to either increase sensitivity and detect less obvious airways (and increase the number of falsely identified airway locations) or to decrease the sensitivity and possibly miss some airways. We assume that falsely detected airways will not have the desired three-dimensional connectivity (as suggested from 3D rendering of detected airways) and as such will be removed during the final tree growing stage. Therefore, higher sensitivity of the method was preferred.
The presented validation in individual image slices is not sufficient on its own. Therefore, the second validation stage was designed to test the method's overall airway tree detection performance. For objective validation, computer generated and physical phantoms were developed. The second validation stage in phantoms is described in 15.
The geometry and structure of the pulmonary airways can only be quantitated correctly in three-dimensional space. Accurate quantitation of the airways from high resolution CT image data is expected to contribute significantly to the evaluation of gene therapy for cystic fibrosis. Our method holds the promise to successfully track the full airway tree over time to evaluate changes on a regional basis via our ability to find the same location within the tree across studies. By defining the data structure through which specific lung locations can be tracked and by identifying the skeleton defining airway geometry, the utility of multidimensional imaging in the study of regional pulmonary geometry is substantially enlarged. Correspondence between airway structure across volumes remains an open problem.
This work was supported in part by NIH RO1-HL-42672, the 1993/94 Pilot and Feasibility Project of the Cystic Fibrosis Research and Development Program, and the 1993/94 Carver Scientific Research Initiative Grants Program.
Papers |
DPI Homepage |
VIDA |
NLM |
Contact Us |
Search
