Airway Tree Structure
The first part of this study deals with the methodology to extract morphometric information from HRCT images of the human lung using the commercial software PW2. The software uses validated algorithms to generate data through streamlined, automated procedures. However, there are several limitations to the quantity of the information provided without any user intervention; that is, how many airways are segmented and identified. The results of the segmentation, skeletonization, and identification of individual airways are prone to a number of artifacts due to either decreased image quality (low signal-to-noise ratio, motion) or algorithmic restrictions (multiple segmentations, leakages, and skeletonization issues), defects which are usually intensified in pathological lungs. Effects such as these can be observed in Figs. 2b–d. It has to be mentioned that a newer version of the software, which is already commercially available, is reported to obtain results deeper in the tree with much less user intervention (http://www.vidadiagnostics.com/).
The delineation of the airway tree seems to be robust down to Generation 5, especially in the lower lobes. This is shallower than the theoretically optimal resolution expected by modern HRCTs. A possible reason for this is the relatively high asymmetry at certain parts of the tree, where the minor daughter size is so much smaller than the major one that no unique segmentation is achievable. The result is often a discontinuation of airway tree segmentation, even though the distal subtree might contain several high-quality airway samples. Similarly, in regions with complicated bronchial connectivity, such as the RUL of Fig. 6a or the right subsuperior segment (RB*, Fig. 6b), the skeleton resulting from the algorithmic routine may be unrealistic. This effect can be seen in Table 2, where the trifurcations and small “intermediate” airways are given per generation and per lobe. Under these conditions, the skeletonization algorithm regularly fails to delineate the airways correctly. Finally, the left lung is especially prone to motion artifacts due to proximity of the heart; these artifacts often look like low attenuation regions or “shadows” in the parenchyma. However, the expansion of the airway tree beyond this generation is possible and, indeed, required for the acquisition of additional information on airway-tree morphology. Therefore, manual intervention by the user for the optimization of the 3D reconstruction is a significant part of the process.
As mentioned earlier, since the process is semiautomatic the manual segmentation of airways might not be accurate enough. This is especially true for the asthmatic lung, where airway segmentation is less successful due to some bronchial disease characteristics. One example is the bronchial stenosis shown in Fig. 3. In some cases the lumen might not be visible at all for entire tree sections. These “invisible” subtrees are detectable intuitively, by locating subtrees that are not connected to any central airways. It is a subject for investigation whether these images represent permanent (or semipermanent) airway closures due to remodeling and inflammation of the bronchial wall (Wiggs et al., 1992; Pepe et al., 2005; Aysola et al., 2008; Zanini et al., 2010; Al-Muhsen et al., 2011), mucous plugging, or smooth muscle contraction.
Other characteristics of asthma that can affect airway segmentation include “spurs” at bifurcation points, where the lumen is compressed by increased WT, increased radius of curvature of the airways (probably due to the extra elasticity of the smooth muscle), and higher signal-to-noise ratios in the images, possibly due to extra attenuation from mucous secretions. Because of all these effects, the extraction of information from the asthmatic images requires intuition, attention, and training on the part of the user. For that reason, a necessity exists for better algorithmic processes that eliminate user intervention to penetrate deep into the tree and quantify features such as the spurs or the levels of lumen caliber reduction. It is important to note that the 3D representation adds an important aspect to the investigation of clinically important defects, as many of the stenoses observed in our study were difficult to observe in the 2D HRCT slices.
A measure of the successful segmentation of the bronchi can be given by measuring the actual versus predicted number of airways (Table 2). These results are within the expected range of the results published by Salito et al., 2011 for images captured at residual volume (Generation 5: healthy 33% and emphysema 11%) and images acquired at total lung capacity (TLC) (Generation 6: healthy 67% and emphysema 22%). However, Montaudon et al., 2007b managed to segment more than 100% of Generation 5 airways (healthy lungs), with this percentage dropping to 84% at Generation 6 and 14% at Generation 8.
It is unclear why this discrepancy in airway detection might exist between the different techniques for similar healthy groups. The most likely reason is that the difference in breathing regimes affected the size of the bronchi, making it easier to segment out individual airways in healthy lungs at TLC. Other possible factors are algorithmic or scanning device differences. It is worth mentioning the fact that segmentation penetrates deeper in the lower lobes suggesting the possibility of positive correlation between the gravity angle (Sauret et al., 2002) and airway size.
An important finding of this work is that significant differences in airway tree structure from the theoretical bifurcating tree structure were identified (Table 2). In many cases, successive bifurcations are connected by small “intermediate” airways that may or may not be visible in the 3D object. The “intermediates” are important to the tree structure for several reasons, the most important of which is the effect they have on air flow and particle deposition, since bifurcations and trifurcations have different flow profiles (Choi et al., 2010; Gemci et al., 2008; van Ertbruggen et al., 2005). It is important to note that there is no clear delineating distinction between a trifurcation and a short child airway. An implication to tree modeling is the imbalance they cause to the number of generations in a particular subtree, causing errors of a statistical nature. Therefore, some postprocessing of the data is always necessary in order to determine the “real” connectivity and, thus, the generation of all bronchi.
A broad range of studies exists investigating the airway remodeling in different diseases (Niimi et al., 2000; Gono et al., 2003; de Jong et al., 2005; Kotaru et al., 2005; Pepe et al., 2005; Olivier et al., 2006; Aysola et al., 2008; Hoshino et al., 2010; Al-Muhsen et al., 2011). However, it is unclear if there is a higher probability for polyfurcations in different diseases. Generally speaking, there is a lack of studies in the literature correlating how diseases with genetic backgrounds affect the structure and connectivity in the airway tree or if there is a causality effect between a disease and particular types of connectivity structures. Some recent reports suggest that there is a connection between cellular dysfunction in prenatal stages of lung development and lung structure in the diseased lung postnatally, a fact that remains to be further investigated and quantified (Morrisey and Hogan, 2010). The data presented in this study fall inside the categories described by Ghaye et al., 2001, with the exception of the RM Lobe trifurcations, which has not been previously reported. However, not enough samples were available to enable any definitive conclusions.