Segmentation and Virtual Exploration of Tracheo-Bronchial Trees
Dirk Mayer , Dirk Bartz , Sebastian Ley , Steffi Thust , Claus Peter Heussel , Hans-Ulrich Kauczor , and Wolfgang Straßer
Department of Radiology, Johannes-Gutenberg-University Mainz, Langenbeckstr. 1, D-55131 Mainz, Germany Email:
[email protected]
WSI/GRIS, University of T¨ubingen, Sand 14, D-72076 T¨ubingen, Germany, Email: bartz,strasser @gris.uni-tuebingen.de
Abstract. The tracheo-bronchial tree as part of the lung is part of one of the most important organs of the human body. Inhaled air is distributed to the alveolus where oxygen and carbondioxide exchange between air and blood takes place. In this paper, we introduce the virtual endoscopy system VIVENDI to perform virtual inspections of tracheo-bronchial trees based on their segmentation and of the complementing blood vessels. It is based on a hybrid segmentation pipeline which enables the segmentation of vascular and tracheo-bronchial structures down to the 7th generation of the bronchi. Keywords: Tracheo-bronchial tree, hybrid segmentation, CT, virtual endoscopy. 1. Purpose The tracheo-bronchial tree as part of the lung is part of one of the most important organs of the human body. Inhaled air is distributed to the alveolus where oxygen and carbon-dioxide exchange between air and blood takes place. The tracheo-bronchial tree is complemented by a system of pulmonary venous and arterial blood vessels which transports the blood to and from the heart into the lungs. Several pathologies can jeopardize a sufficient lung function. Among them are tumors, pulmonary embolism, collapse of the lungs (atelectasis), pneumonia, emphysema, asthma, and many more. For a proper diagnosis and treatment, the respective pathologies need to be identified and in some cases quantified. Furthermore, lung surgery requires a pre-operative planning of the intervention. Currently, this is done using a combination of computed tomography (CT) as a tool for morphological imaging of the whole lung parenmchyma, and bronchoscopy as an interventional tool for inspection of the central airways and deriving tissue samples. Due to the recent technical development improving resolution and scan velocity, CT might be an promising alternative to bronchoscopy. Even because it images smaller airways and distally to obstructions, which is a limit in real bronchoscopy. In this contribution, we introduce the virtual endoscopy system VIVENDI to perform virtual inspections of tracheo-bronchial trees based on their segmentation and of the complementing blood vessels. It provides an interactive visualization of both structures down to the 7th generation of the bronchi. This is enabled by a hybrid segmentation technique which overcomes
leakage problems of most previous approaches. 1.1. Related Work Virtual bronchoscopy has been proposed since 1994 [9]. It used helical CT as imaging modality and standard 3D region growing as segmentation technique [8,4,7]. However, with these techniques, only the upper bronchi could be segmented [8]. More recently, Kirarly et al. [1] use a combination of a adaptive 3D region growing, 2D mathematical morphology and an optional 2D median filter for increasing the robustness of the algorithm. Ability and effects of the methods are discussed on results of segmented multislice images. Grenier et al. [5] gives an overview of existing visualization techniques depending on different airway diseases. 2. Methods The segmentation and visualization is based on CT data acquired by a Siemens Somatom Volume Zoom multi-slice CT scanner (convolution kernel 50, collimation of 1.25mm and increment of 1mm). It typically generates a stack of 250-300 images with a matrix of 512 x 512 pixels resolution and a spacing which varies from sub-millimeter (i.e., 0.6mm - 0.7mm) for the pixel distance and 1mm for the slice distance. Based on this dataset, we segment the tracheobronchial and the blood vessel tree of the lungs using the segmentation system SegoMeTex. Subsequently, we reconstruct the inner surface of these structures and generate other datastructures needed for virtual endoscopy. Final, we explore the dataset by a virtual bronchoscopy procedure using the VIVENDI system. 2.1. Segmentation The segmentation procedure of SegoMeTex consists of three stages (Fig. 1). In the first stage, the trachea and central bronchi are segmented using standard 3D region growing methods. Partial volume effects and limited resolution of the CT scan (which essentially cause this effect)
(a)
(b)
(c)
Figure 1. Results of three segmentation stages: (a) Region growing, (b) 2D wave propagation, (c) 2D template matching
render this method as not sufficient for further generations of the bronchi. Therefore a 2D wave propagation is initiated to segmented the main branches. Finally, a 2D template matching procedure is used to segmented small lumen, which might be only a voxel large. 2.2. Masked 3D Region Growing We can make certain assumptions for volume data from CT; all voxels of an intesity value below -950 HU are definitely air, thus part of an airway. Every voxel of a value above -775 will be non-airway, since it is the density value of a bronchial wall or other tissue. Every voxel in this uncertain area between this isovalue interval (from -950 HU to - 775 HU) can possibly belong to airway. The double threshold 3D region growing algorithm extracts all voxels which are definitely air, starting in the trachea. To prevent the leaking into the parenchyma of the lungs in smaller airways (e.g. in empyhsema), we use a masking technique from texture analysis. If the average gray value of a 3x3 voxel cube centered at the current voxel is within a sensitive range, we consider this voxel as being part of the airway. Usually the bronchial tree up to the 5th generation can be segmented with this method. 2.3. 2D Wave Propagation Starting from segmented voxels of the previous step, 2D wave propagation tries to reconstruct bronchi walls in the 2D plane. It starts at the boundary voxels of the airway voxels from 3D region growing and propagates waves to detect the walls of the bronchi. Voxels in the uncertain areas are classified by fuzzy logic rules which takes the density value, the local N4 neighborhood (in 2D) gradient, and some knowledge of already classified voxels from the previous wave into account. Furthermore, several rules are monitoring the shape of the bronchi tree to prevent leaking into the non-airways parts of the lungs. With these rules, a number of variables is tested, such as the number of segments in a wave (must be less than three), the number of steps in the recursive bronchi segmentations, the number of voxels and the diameter of the current bronchi. To follow a bronchus through several slices, virtual waves are propagated in neighboring slices. If one of these virtual waves is similar to the wave propagation in the current slice, an actual recursive wave propagation in the neighboring slice is initiated. 2.4. 2D Template Matching The previous two approaches will leak into the surrounding area, if the airways become too small to be picked up. To prevent this leaking, we apply a 2D template matching technique which evaluates the candidate area below templates with the isovalue category ”uncertain” (between -770 HU and -950 HU). Various templates are generated by a 2D seeding that starts from the boundary voxels of the previous segmentations (Fig. 2). Variations of the isovalue are considered, as long as the number of selected boundary voxels is below the critical limit (i.e., 35 voxels), since it can be assumed that they did not leak out. The templates are moved over the potential locations. Based on a fuzzy logic rules driven evaluation, which takes into account the average density value of the template area and the average contrast to the surrounding voxels in the N8 neighborhood (within a single slice), they are combined to establish an improved representation. 2.5. Virtual Endoscopy Based on the generated segmentation (see Fig. 3c), we reconstruct the iso-surface of the segmentation using the standard Marching Cubes algorithm (Lorensen, Cline 1987). However, the
(a)
(b)
(c)
(d)
(e)
(f)
Figure 2. Template matching: (a) shows a peripheral airway. 2D seeding is started on pixels of the category definitely airway. The result is marked blue (b). From there the seeding area templates are formed and tested on different locations around the seeding. The seeding is repeated on the category airway and template (marked by a pattern, (c) to (e)). (f) shows the best result after the classification.
direct reconstruction of a segmentation can result in a bumpy appearance, caused by interpolation artifacts due to high density differences. Therefore, we add one layer of voxels to the segmentation which reduces this effect. The application of more continues distance fields can further reduce this effect. However, it might also reduce reconstructed segmentation details. For the virtual endoscopic exploration, we use the reconstructed models (i.e., tracheo-bronchial tree and pulmonary blood vessel tree) and render those independently. Usually, we assume a viewpoint inside a bronchi, hence we render its surface semi-transparent and the pulmonary vessels opaque [3]. However, we can also assume a viewpoint inside of the blood vessels. Consequently, the transparency parameters would be switched. Independent from the chosen viewpoint location, the enclosing structure is rendered, while the outside structure should be rendered opaque. Note that also other structures (i.e., a tumor) can be added to the scene. In this case, the rendering parameters need to be chosen accordingly. For the actual virtual bronchoscopy, we start in the trachea (Fig. 3a) and traverse the tracheobronchial tree (Fig. 3b) to the target region of interest. Interesting structures or pathologies can be documented for further evaluation.
(a)
(b)
(c)
Figure 3. Virtual endoscopy: (a) View in trachea down to main bi-furcation, (b) lower airways segmented by 2D template matching. The diamond shape interpolation artefacts are due to high zoom factor in this close-up. (c) Final segmentation result. The colors code the different segmentation steps: region growing (blue), 2D wave propagation (red), 2D template matching (yellow).
3. Results The segmentation method was successfully tested on 22 patients, who were separated into three groups of normal subjects (n=8), emphysema (n=5), and lung diseases with increased density, such as pneumonia (n=9). The results of the segmentation were compared to segmentations using only the threshold/region growing method and assessed by an experienced chest radiologist, who also identified false positive segmentation results [6]. The presented hybrid segmentation method identified bronchi up to the sixth generation with a sensitivity of more than 58%, and a positive predictive value (PPV, similar to specificity: correctly found branches relative to all segmented branches) of more than 90%. Beyond the sixth generation, the sensitivity drops below 30%, while the PPV maintains its high level. For emphysema patients, the PPV was slightly reduced to due the damage of the alveolus which increased the leakings into false positive branches. For pneumonia patients, the sensitivity (and to a smaller extent the PPV) are significantly reduced compared to the other datasets. This is due to the increased density in the airways which increases the difficulties of tracking the inner surface tremendously. After the segmentation, we explored a subset of the datasets interactively (more than 10fps on a standard PC) using the virtual endoscopy software [2]. The exploration exposed a high quality reconstruction, even of small structures throughout the dataset. 4. Conclusion In this contribution, we discussed SegoMeTex, a hybrid segmentation system which was combined with VIVENDI, a virtual endoscopy system. Based on a high-quality segmentation,
meaningful features could be visualized. Whereas the segmentation provided good results for a large variety of patients, it encountered problems identifying all airways with pneumonia patients. In general, virtual bronchoscopy is a valuable tool for the localization and measurement of stenosis for treatment planning. However, mild stenosis, submucosal infiltation, and superficial spreading tumors cannot be identified. Acknowledgement This work was supported by the European Commission (IST-1999-14004: ”COPHIT”, the DAAD/British Council, and by DFG Project CatTrain. We would like to thank Jan Fischer and Anxo del Rio of the WSI/GRIS for modifications of the implementation of VIVENDI. REFERENCES 1. A. Kiraly and W. Higgins and G. McLennan and E. Hoffman and J. Reinhardt. ThreeDimensional Human Airway Segmentation Methods for Clinical Virtual Bronchoscopy. Academic Radiology, 9(10):1153–1168, 2002. 2. D. Bartz and M. Skalej. VIVENDI - A Virtual Ventricle Endoscopy System for Virtual Medicine. In Proc. of Symposium on Visualization, pages 155–166,324, 1999. ¨ G¨urvit, D. Freudenstein, and M. Skalej. Interactive and Multi3. D. Bartz, W. Straßer, O. modal Visualization for Neuroendoscopic Interventions. In Proc. of Symposium on Visualization, pages 157–164, 2001. 4. G. Ferretti, D. Vining, J. Knoplioch, and M. Coulomb. Tracheobronchial Tree: ThreeDimensional Spiral CT with Bronchoscopic Perspective. Journal of Computer Assisted Tomography, 20(5):777–781, 1996. 5. P. Grenier, C. Beigelman-Aubry, C. Fetita, F. Preteux, M. Brauner, and S. Lenoir S. New Frontiers in CT Imaging of Airway Disease. European Radiology, 12(5):1022–1044, 2002. 6. D. Mayer, S. Ley, B. Brook, S. Thust, C. Heussel, and H. Kauczor. 3D-Segmentierung des menschlichen Tracheobronchialbaums aus CT-Bilddaten. In Proc. of Workshop Bildverarbeitung f¨ur die Medizin, pages 333–337, 2003. 7. J. Rodenwaldt, L. Kopka, R. Roedel, A. Margas, and E. Grabbe. 3D Virtual Endoscopy of the Upper Airways: Optimization of the Scan Parameters in a Cadaver Phantom and Clinical Assessment. Journal of Computer Assisted Tomography, 21(3):405–411, 1997. 8. R. Summers, D. Feng, S. Holland, M. Sneller, and J. Shelhamer. Virtual Bronchoscopy: Segmentation Method for Real-Time Display. Radiology, 200:857–862, 1996. 9. D. Vining, R. Shifrin, E. Haponik, K. Liu, and R. Choplin. Virtual Bronchoscopy (abstract). In Radiology, volume 193(P), page 261, 1994.
