Airway segmentation and analysis for the study of mouse models of lung disease using micro-
CT
This article has been downloaded from IOPscience. Please scroll down to see the full text article.
2009 Phys. Med. Biol. 54 7009
(http://iopscience.iop.org/0031-9155/54/22/017)
Download details:
IP Address: 159.237.12.65
The article was downloaded on 05/10/2010 at 11:27
Please note that terms and conditions apply.
View the table of contents for this issue, or go to the journal homepage for more
Home Search Collections Journals About Contact us My IOPscience
IOP PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY
Phys. Med. Biol. 54 (2009) 7009–7024 doi:10.1088/0031-9155/54/22/017
Airway segmentation and analysis for the study of
mouse models of lung disease using micro-CT
X Artaechevarria
1
,DP
´
erez-Mart
´
ın
1
,MCeresa
1
, G de Biurrun
2
,
D Blanco
2
, L M Montuenga
2
, B van Ginneken
3
, C Ortiz-de-Solorzano
1
and A Mu
˜
noz-Barrutia
1
1
Cancer Imaging Laboratory, Center for Applied Medical Research, 31008 Pamplona, Spain
2
Biomarkers Laboratory, Center for Applied Medical Research, University of Navarra, 31008
Pamplona, Spain
3
Image Sciences Institute, 3584CX Utrecht, The Netherlands
E-mail: xabiarta@unav.es
Received 13 July 2009, in final form 14 October 2009
Published 4 November 2009
Online at stacks.iop.org/PMB/54/7009
Abstract
Animal models of lung disease are gaining importance in understanding the
underlying mechanisms of diseases such as emphysema and lung cancer.
Micro-CT allows in vivo imaging of these models, thus permitting the
study of the progression of the disease or the effect of therapeutic drugs in
longitudinal studies. Automated analysis of micro-CT images can be helpful
to understand the physiology of diseased lungs, especially when combined with
measurements of respiratory system input impedance. In this work, we present
a fast and robust murine airway segmentation and reconstruction algorithm.
The algorithm is based on a propagating fast marching wavefront that, as it
grows, divides the tree into segments. We devised a number of specific rules
to guarantee that the front propagates only inside the airways and to avoid
leaking into the parenchyma. The algorithm was tested on normal mice, a
mouse model of chronic inflammation and a mouse model of emphysema. A
comparison with manual segmentations of two independent observers shows
that the specificity and sensitivity values of our method are comparable to the
inter-observer variability, and radius measurements of the mainstem bronchi
reveal significant differences between healthy and diseased mice. Combining
measurements of the automatically segmented airways with the parameters of
the constant phase model provides extra information on how disease affects
lung function.
1. Introduction
High-resolution micro-computed tomography (micro-CT) is an excellent tool to study rodent
models of lung disease because of the inherent x-ray absorption contrast between tissue and air
0031-9155/09/227009+16$30.00 © 2009 Institute of Physics and Engineering in Medicine Printed in the UK 7009
7010 X Artaechevarria et al
that exists in the lungs (Ford et al 2007, Johnson 2007). This technology has been successfully
used to study different disease models such as lung cancer and emphysema (De Clerck et al
2004, Postnov et al 2005, Froese et al 2007). Furthermore, its applications are likely to grow
with the advent and improvement of flat-panel detectors.
Latest research on lung cancer points toward common initiation mechanisms in lung
cancer, chronic obstructive pulmonary disease (COPD) and inflammatory lung disease (de
Torres et al 2007, Houghton et al 2008, Cassidy et al 2007). In this scenario, the analysis of
biologically relevant small animal models will be critical to gain knowledge on the specifics of
each disease and the relationships existing between them. The widespread use of micro-CT in
lung disease-related animal studies raises the need for precise and robust automated analysis
and quantification tools for micro-CT images. Standardized automated techniques are thus
required to allow for interlaboratory comparisons and reproducible validation of the results.
Segmenting the airways is required by most automatic lung analysis algorithms, especially
in mice micro-CT images, due to the relatively large volume occupied by the airways. It is well
known that the functionality of central airways and lung peripheral areas varies considerably,
and the morphological effects of a great number of lung pathologies are different in the
airways and the parenchyma. Separating airways from the rest of the lung is thus of the
utmost importance when analyzing a disease (such as emphysema, chronic inflammation or
cancer) which affects lung density differentially. Moreover, the analysis of segmented airways
may be interesting in itself to study airway-specific morphological changes such as stenosis,
bronchiectasis, etc. Finally, the airways can be used as a reference for image registration in
follow-up studies or for atlas-based segmentation of the lungs, lobes and pulmonary segments.
To the best of our knowledge, only two airway segmentation methods in micro-CT images
have been reported that are of use in mice. Chaturvedi and Lee (2005) segmented silicon casts
of excised mice lungs using an interactive region growing algorithm. More relevant to the
present work, Shi et al (2007) reported a fully automatic algorithm to segment in vivo images
of healthy mice, acquired with an iso-pressure breath hold protocol (Namati et al 2006). This
algorithm works by searching airway lumens in 2D transversal slices that are then used to
reconstruct the full 3D tree. The main drawback of this method is that it is based on the search
of candidates in 2D, which may lead to problems due to the high variability of the shape of
the airway lumens as seen in transversal slices. Moreover, reported computation times were
fairly high (about 30 min for each scan).
Due to scarce previous work on mouse airway segmentation, it is worth reviewing the most
relevant approaches that exist for segmenting the human airways in CT images. Schlathoelter
et al (2002) introduced an algorithm for simultaneous segmentation and reconstruction of the
airways. This algorithm was based on a propagating front that divides the tree into branches
during segmentation. The framework was extended and generalized in a subsequent work by
B
¨
ulow et al (2004). More recently, van Ginneken et al (2008) also used the framework for
human airway segmentation with a multi-threshold approach to increase robustness. Kiraly
et al (2002) compared two different methods: an adaptive region growing algorithm and an
algorithm that combined region growing and mathematical morphology. Their conclusion
was that the region growing method was faster than the hybrid method, but also slightly less
accurate. Aykac et al (2003) used a two-step approach. In a first step, candidate airways were
identified on transversal slices using grayscale morphological reconstruction. In a second step,
valid candidates were connected to build a 3D airway tree. Fetita et al (2004) also employed
mathematical morphology to obtain a first approximation of the airways. The morphological
operator worked in 3D and was specifically devised for this purpose. An energy-minimizing
reconstruction algorithm was used to build the final airway tree. The work by Tschirren et al
(2005
) was based on fuzzy connectivity. They made use of small adaptive regions of interest
Airway segmentation in mouse models of lung disease 7011
around the already segmented airway areas. Thus, the algorithm adapted to local image
characteristics, leaks were detected early and the computing time was reduced. In a recent
work, Graham et al (2008) present a method that can be summarized in three main steps. In the
first step, a conservative segmentation of the major airways is obtained via 3D region growing
on heavily smoothed data. Then, possible branch segments are detected and connected to
each other by nonlinear filtering and surface interpolation. The final segmentation is obtained
using a global graph partitioning algorithm, which connects the valid branch segments to the
major airways.
The main reason that prevents a direct translation of these algorithms to mice micro-CT
data is the low signal-to-noise ratio (SNR) of the images, caused by the required small voxel
size and the limitation on radiation dose imposed by the in vivo studies. In particular, following
the simple model assumptions from Ford et al (2003), a reduction of one order of magnitude
invoxelsize(0.5mminCTto50μm in micro-CT) implies an increase of two orders of
magnitude in the variance of the linear attenuation coefficient, if the rest of parameters remain
unchanged.
Taking this into account, we decided to develop the flexible segmentation and
reconstruction framework first reported by Schlathoelter et al (2002). This framework has
multiple advantages. First, it allows for simultaneous segmentation and reconstruction of tree-
like structures. Second, the topological and morphological information from the segmented
tree can be used to guide the segmentation of the remaining branches. Third, its modular
configuration allows for easy introduction of application-specific segmentation rules. In fact,
one of the major contributions of our work is the use of new features when compared to those
in the previously reported applications of the framework, to adapt to the low SNR and the
special morphology of the mice airways.
To complete the morphological information provided by the image-based measurements,
we use respiratory system input impedance measurements and the constant phase model
parameters, which have been widely used to assess respiratory mechanics in multiple animal
models (Hantos et al 1992, Collins et al 2003, Tomioka et al 2002). The combination of micro-
CT imaging and constant phase model parameters has also been used to analyze animal models
of lung disease before. In particular, Lundblad et al (2007) qualitatively analyzed ex vivo micro-
CT images in a mouse model of allergical inflammation and combined it with measurements
of tissue elastance. In this work, we propose to use quantitative airway measurements since
this approach is likely to provide valuable information to better understand the morphology
and function of lungs affected by diseases such as emphysema and inflammation.
The rest of the article is structured as follows. In section 2, we briefly summarize the
airway segmentation framework and give details about the new features we have introduced.
In section 3, the image acquisition protocol, which includes the respiratory system input
impedance measurement, and the airway segmentation validation experiments are detailed.
Results are presented in section 4. A final discussion (section 5) concludes the article.
2. Methods
2.1. Prefiltering
Micro-CT images contain high levels of noise. This justifies the need for a filtering step
before the analysis. To this end, we used a 3D grayscale closing by reconstruction filter with
a spherical structuring element of radius 1 voxel (Vincent 1993). The 2D version of this filter
was reported to yield positive results in the previous work by Shi et al (2007). This filter
increases the contrast of the darkest regions of the image while preserving the shape of the
7012 X Artaechevarria et al
Figure 1. Block diagram of general tree segmentation and reconstruction framework.
structures. The radius was selected because it represented a good trade-off between noise
removal and contour information preservation.
2.2. Airway tree segmentation and reconstruction
The adopted framework has been explained in detail in previous works by Schlathoelter
et al (2002) and B
¨
ulow et al (2004). A block diagram summarizing the main execution flow is
shown in figure 1. A key concept is the segment. A segment is a set of contiguous points that
has been segmented by a growing wavefront without bifurcations. The growing wavefront
is initialized in one seed point and grows into neighboring voxels that fulfill certain voxel
acceptance criteria. After every wavefront propagation step, several conditions are checked to
prevent leaks. When the wavefront bifurcates, new segments are initialized and added to the
rest of pending segments in a segment queue. This queue is a first-in, first-out (FIFO) list, thus
ensuring that branches from upper airways are processed first. When a segment is finished, its
correctness is checked in a segment evaluation step.
It was necessary to develop new voxel acceptance, propagation evaluation and segment
evaluation criteria, due to the particularities of our segmentation task. Details about these new
features are given in this subsection.
