DOI:10.2214/AJR.05.0911
AJR 2007; 188:W276-W280
© American Roentgen Ray Society
Application of an Optical Flow Method to Inspiratory and Expiratory Lung MDCT to Assess Regional Air Trapping: A Feasibility Study
Drew A. Torigian1,
Warren B. Gefter,
John D. Affuso,
Kiarash Emami and
Lawrence Dougherty
1 All authors: Department of Radiology, University of Pennsylvania School of
Medicine and Hospital of the University of Pennsylvania, 3400 Spruce St.,
Philadelphia, PA 19104-4283.
Received May 27, 2005;
accepted after revision August 18, 2006.
Address correspondence to D. A. Torigian
(Drew.Torigian{at}uphs.upenn.edu).
WEB This is a Web exclusive article.
Abstract
OBJECTIVE. We describe the application of an optical flow method to
inspiratory and expiratory high-resolution volumetric lung MDCT for the
assessment of regional air trapping.
CONCLUSION. Qualitative and quantitative assessment of regional air
trapping is feasible using an optical flow method to align volumetric MDCT
data sets.
Keywords: air trapping • airway • chest imaging • functional lung imaging • high-resolution CT • lung disease • MDCT
Introduction
The optical flow method was first developed in the early 1980s to
approximate image motion from sequential time-ordered images
[1,
2] and is particularly suited
for 3D registration of CT images of the lung because of the presence of
high-resolution high-contrast vascular structures that act as fiducial
landmarks for volume alignment. As an application to inspiratory and
expiratory high-resolution volumetric lung MDCT, the optical flow method holds
potential for rapid and reproducible assessment of mechanical lung function.
In this report, we assess the feasibility of using an optical flow method to
accurately register separate inspiratory and expiratory high-resolution CT
data sets so that image subtraction can be performed. The resulting
attenuation difference color maps allow the qualitative and quantitative
assessment of regional pulmonary air trapping.
Materials and Methods
Approval for this retrospective study and a waiver from the Health
Insurance Portability and Accountability Act were obtained from our
institutional review board before study initiation. An MDCT examination of one
patient with multifocal air trapping was retrospectively selected from our
institution's image archives. The study was a high-resolution CT examination
of the chest that was acquired in a fashion similar to that described by
Nishino and Hatabu [3]; axial
images of contiguous 5-mm sections were obtained during inspiration and
expiration on a 16-MDCT scanner (Sensation 16, Siemens Medical Solutions)
using 120 kVp, 100 mAseff, a table speed of 12 mm per rotation, and
a gantry rotation time of 0.5 second. No ECG gating or triggering and no
respiratory gating were implemented during image acquisition. Axial images
were subsequently reconstructed with a 2-mm slice collimation and 1-mm slice
overlap in the axial plane using a lung reconstruction algorithm. The image
volume size was 269 x 346 x 247 using a 30-cm field of view for
both the end-inspiratory and end-expiratory image volumes.
The thin-section inspiratory and expiratory images, which did not show
patient information, were transferred to an offline computer in DICOM format.
The expiratory image volumes were registered with the inspiratory volumes
using the optical flow method on a standard desktop PC (Dimension 450, Dell
Inc.) with dual 3.06-GHz Xeon processors (Intel) and custom software that is
based on the methods developed by Bergen et al.
[4] and adapted for medical
imaging applications by Kumar et al.
[5]. The optical flow method
processing time was less than 10 minutes. Alignment performance was judged by
computing the cross correlation between the inspiratory and expiratory image
volumes before and after registration. Because the original volumes were often
severely misaligned, there were sections that had no counterpart in the
comparison volume. Therefore, only the volumes common to both data sets were
used for correlation.
For the areas of air trapping to be highlighted, attenuation differences
were computed by subtracting the inspiratory volumes from the aligned
expiratory volumes and were subsequently displayed with color maps using
public domain software (ImageJ, National Institutes of Health). Areas of
pulmonary air trapping generally correspond to regions with low attenuation
difference values, whereas areas with less or no air trapping correspond to
regions with higher attenuation difference values.
Color map images showing attenuation differences in the entire right lung
were further quantitatively analyzed using a computer code developed in a
Matlab 7 environment (MathWorks). After images were segmented from the
background using ImageJ software, a histogram of pixel frequency versus pixel
attenuation difference was generated and the corresponding colors were placed
under the curve using the same color scale as that used for the color map
images. Peak, mean, and SD attenuation difference values were calculated for
the histogram. A gaussian curve fit was defined as follows:
where a1 is the peak, b1 is the
mean, and c1 / 
2 is the SD. A gaussian curve was
then performed, and peak, mean, SD, and full-width at half-maximum (FWHM)
values (defined as FWHM = 2 SD 2 1n 2) were calculated. Subsequently,
the area under the curve (AUC) percentages for pixel values less than the
arbitrary threshold attenuation difference values of 0, 50, 100, and 200
Houndsfield units (H) were computed.
Results
Initial end-expiratory (Fig.
1A) and end-inspiratory (Fig.
1B) MDCT image volumes showed multiple regions of hyperlucency in
the lung, consistent with multifocal air trapping. The end-expiratory image
volume was registered (Fig. 1C)
to the end-inspiratory image volume via the optical flow method with a
correlation coefficient of 0.94. Multifocal regions of air trapping in the
right lung were visible on the attenuation difference color map images (Figs.
1D,
1E and
1F). Qualitatively, air
trapping at expiration was seen as regions with low attenuation difference
values, whereas areas of lung with less or no air trapping had higher
attenuation difference values.
View larger version (76K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 1D —74-year-old woman with multifocal air trapping. Quantitative
attenuation difference color maps for images obtained in axial (D),
coronal (E), and sagittal (F) planes reveal multifocal regions
of air trapping throughout right lung as dark blue to black regions. Color
scale ranges from 0 H (dark blue to black) to 366 H
(bright orange to yellow). Note white areas in periphery of
axial and coronal images with corresponding values of more than 366 H are
predominantly due to incomplete segmentation of chest wall.
|
|
View larger version (62K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 1E —74-year-old woman with multifocal air trapping. Quantitative
attenuation difference color maps for images obtained in axial (D),
coronal (E), and sagittal (F) planes reveal multifocal regions
of air trapping throughout right lung as dark blue to black regions. Color
scale ranges from 0 H (dark blue to black) to 366 H
(bright orange to yellow). Note white areas in periphery of
axial and coronal images with corresponding values of more than 366 H are
predominantly due to incomplete segmentation of chest wall.
|
|
View larger version (86K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 1F —74-year-old woman with multifocal air trapping. Quantitative
attenuation difference color maps for images obtained in axial (D),
coronal (E), and sagittal (F) planes reveal multifocal regions
of air trapping throughout right lung as dark blue to black regions. Color
scale ranges from 0 H (dark blue to black) to 366 H
(bright orange to yellow). Note white areas in periphery of
axial and coronal images with corresponding values of more than 366 H are
predominantly due to incomplete segmentation of chest wall.
|
|
Quantitatively, attenuation difference values in the right upper lobe,
right middle lobe, and right lower lobe ranged from 0 to 321, 45 to 359, and 5
to 366 H, respectively, reflecting the varying amounts of air trapping in each
lobe. A histogram of pixel frequency versus pixel attenuation difference
values for the entire lung was successfully generated and revealed a range of
attenuation difference values from -484 to 2,100 H, but the pixel attenuation
difference values were truncated at 1,000 H because values above that were
negligible (Fig. 1G).
Attenuation difference pixel values of the lung not thought to be due to the
effects of misregistration or to incomplete segmentation of the chest wall
ranged from 0 to 366 H, as confirmed by manual region-of-interest measurements
in corresponding portions of pulmonary parenchyma on registered
end-inspiration and end-expiration images (Figs.
1D and
1E).
View larger version (10K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 1G —74-year-old woman with multifocal air trapping. Attenuation
difference histogram of entire right lung shows peak pixel frequency of 5.0%,
mean attenuation difference of 151.4 H, and SD of 141.0 H. Color scale ranges
from 0 H (dark blue to black) to 366 H (bright
orange to yellow), with black for all values less than 0 H and
white for all values greater than 366 H, both of which are thought to be due
to misregistration, incomplete segmentation, or both.
|
|
The attenuation difference histogram had a peak pixel frequency of 5.0%, a
mean attenuation difference of 151.4 H, and an SD of 141.0 H. The gaussian
curve fit defined earlier was also successfully performed
(Fig. 1H), resulting in a peak
pixel frequency of 4.4%, a mean attenuation difference of 117.5 H, an SD of
83.0 H, and a FWHM of 195.5 H. The AUC percentages for pixel values less than
0, 50, 100, and 200 H arbitrary threshold values were 6.0%, 16.1%, 36.7%, and
73.9%, respectively.
View larger version (9K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 1H —74-year-old woman with multifocal air trapping. Gaussian curve fit
(red) of attenuation difference histogram (blue) of entire
right lung has peak pixel frequency of 4.4%, mean attenuation difference of
117.5 H, SD of 83.0 H, and full width at half maximum of 195.5 H. Area under
the curve percentages for pixel values less than 0, 50, 100, and 200 H
arbitrary threshold attenuation difference values were 6.0%, 16.1%, 36.7%, and
73.9%, respectively.
|
|
Discussion
The optical flow method is an image-processing tool that allows one to
identify pixels in separate image volumes on the basis of the pixel
intensities and to align them. Registration of the image volumes is performed
through a gross-scale alignment using translation, rotation, and dilation or
contraction, followed by a fine-scale alignment on a pixel-by-pixel basis.
This tool does not require prior information about the image content,
segmentation, or identification of landmarks and has previously been validated
for registration of tagged MR images and for registration of chest CT images,
with an accuracy of 
95%
[6,
7].
Pulmonary air trapping, a manifestation of obstructive lung disease, may be
associated with a wide variety of diseases that involve the small airways or
pulmonary air spaces and may be associated with pulmonary symptoms and signs
such as cough, wheezing, and dyspnea on exertion
[8,
9]. Some common diseases that
may be associated with air trapping are asthma, emphysema, hypersensitivity
pneumonitis, chronic pulmonary thromboembolic disease, bronchiectasis,
bronchitis, and bronchiolitis, along with a wide variety of interstitial lung
diseases [10,
11].
Lung disease with an obstructive component is often assessed initially
using pulmonary function tests, although these tests are limited in that they
provide a global assessment rather than a regional assessment of lung function
and they may be insensitive to the presence of mild disease
[8]. Regional assessment of the
extent and distribution of pulmonary air trapping is also routinely performed
using end-inspiratory and end-expiratory high-resolution chest CT, although
interpretation of these studies is often largely qualitative or
semiquantitative in nature because a diagnostic radiologist visually assesses
the images [10,
12-15].
Imaging approaches tailored to the quantitative evaluation of obstructive lung
disease have been studied using a variety of scintigraphic, CT, and MRI
methods [11,
16-20].
The highly accurate registration between end-expiratory and end-inspiratory
high-resolution volumetric image data sets through the use of an optical flow
method to generate pulmonary attenuation difference maps and attenuation
difference histograms has the potential to increase the sensitivity, accuracy,
and reproducibility of qualitative CT analysis to determine the extent and
distribution of air trappin
particularly when
air trapping is mild and is unevenly distributed throughout the pulmonary
parenchyma. This potential for improvement in diagnostic capability is
suggested by the results of several other approaches to functional lung
imaging in the study of emphysema and the airways that have recently been
reviewed [11].
Furthermore, the use of an optical flow method allows quantitative
information regarding the volume and distribution of pulmonary air trapping to
be obtained in a reproducible and semiautomated fashion. If one considers air
trapping to be a surrogate marker for the extent and severity of obstructive
lung disease [9,
12], then such quantitative
analysis may be a valuable noninvasive tool in the arsenal of functional lung
imaging approaches with some of the following applications: to aid the
radiologist and clinician in detecting and establishing the cause for
obstructive lung disease in patients; to guide the clinician in the choice of
initial therapy for patients with obstructive lung disease on the basis of
disease severity and estimated likelihood of therapeutic success; to
accurately assess the response of obstructive lung disease during
implementation of a therapeutic regimen; and to facilitate animal- and human
patient-oriented drug development research tailored toward the treatment of
obstructive lung disease.
In our approach to regional qualitative and quantitative assessment of air
trapping, we show the feasibility of an optical flow method to accurately
register two separate end-expiratory and end-inspiratory breath-hold
volumetric MDCT data sets in a human subject. The ability to align these image
data sets enabled us to calculate regional attenuation differences in the
entire lung, which revealed areas of air-trapped lung as regions of decreased
attenuation or as areas with no change in attenuation from end-inspiration to
end-expiration (with corresponding low or zero attenuation difference values).
Furthermore, we show the feasibility of the quantitative assessment of air
trapping through the analysis of attenuation difference histograms. The mean
value of the attenuation difference shown by the histogram is a manifestation
of the combined effects of the normal and air-trapped portions of lung,
whereas the SD and FWHM values reflect the heterogeneity of normal and
air-trapped regions in the total lung. In the case presented, the mean
attenuation difference of 117.5 H is less than that expected in a patient
without air trapping, and the SD and FWHM values were large, consistent with
marked heterogeneity of air trapping throughout the lung, as was verified
visually.
AUC values below arbitrarily designated threshold pixel attenuation
difference values are estimates of the overall percentage of air trapping in
the lung. For example, in the present case, if one assumes that attenuation
difference values of less than a threshold value of 100 H indicate air
trapping, then 36.7% of the total lung has air trapping, whereas if one
chooses a threshold value of 200 H, then 73.9% of the lung has air trapping. A
current limitation is that the optimal threshold value for this determination
is not yet known. Another limitation of our approach is that registration and
segmentation were associated with some error. The AUC of 6.0% for values below
an attenuation difference of 0 H is most likely due to the presence of some
misregistration between the inspiratory and expiratory data sets because
air-trapped lung is not expected to decrease in density from end-inspiration
to end-expiration. Attenuation difference values above 366 H were also thought
to be most likely due to the effects of incomplete segmentation of the chest
wall and less likely due to the effects of misregistration.
Further study will be performed to improve the precision and accuracy of
registration and of segmentation. Moreover, additional studies regarding
technical assessment and clinical utility of an optical flow method and
qualitative and quantitative analyses performed in this study will be required
in larger groups of patients, including studies of correlation between our
approach in the assessment of obstructive lung disease and other tests of air
trapping such as pulmonary function testing. With future software development,
it may also be possible to quantitate attenuation differences on histograms,
AUC values, and other quantitative parameters for each lobe of the lung, which
is another area of future study.
Regional qualitative and quantitative assessment of air trapping is
feasible through the use of an optical flow method to accurately align
inspiratory and expiratory high-resolution volumetric MDCT data sets of the
lungs. Application of an optical flow method in this context may therefore
serve as a potentially valuable tool in the noninvasive evaluation of patients
with obstructive lung disease.
Acknowledgments
We thank John M. Woodburn in the department of radiology at the University
of Pennsylvania School of Medicine for his assistance with Matlab software
programming and data analysis.
References
- Horn BP, Schunck BG. Determining optical flow. Artif
Intel 1981; 17:185
-203[CrossRef]
- Lucas BD, Kanade T. An iterative image registration technique with
an application to stereo vision. In: Proceedings of the Seventh
International Joint Conference on Artificial Intelligence. San
Francisco, CA. Morgan Kaufman: 1981:674
-679
- Nishino M, Hatabu H. Volumetric expiratory high-resolution CT of
the lung. Eur J Radiol 2004;52
: 180-184[CrossRef][Medline]
- Bergen JR, Anadan P, Hanna KJ, Hingorani R. Hierarchical
model-based motion estimation. In: Sandini G, ed. Proceedings of
the Second European Conference on Computer Vision. London, UK:
Springer-Verlag, 1992:237
-252
- Kumar R, Hanna K, Asmuth JC, et al. Detecting lesions in magnetic
resonance breast scans. In: Gerson D, ed. 24th AIPR Workshop on
Tools and Techniques for Modeling and Simulation. Bellingham, WA:
Society of Photo-Optical Instrumentation Engineers, 1996:181
-191
- Dougherty L, Asmuth JC, Blom AS, Axel L, Kumar R. Validation of an
optical flow method for tag displacement estimation. IEEE Trans Med
Imaging 1999; 18:359
-363[CrossRef][Medline]
- Dougherty L, Asmuth JC, Gefter WB. Alignment of CT lung volumes
with an optical flow method. Acad Radiol2003; 10:249
-254[CrossRef][Medline]
- Stulbarg MS, Frank JA. Obstructive pulmonary disease: the
clinician's perspective. Radiol Clin North Am1998; 36:1
-13[CrossRef][Medline]
- Hansell DM. Small airways diseases: detection and insights with
computed tomography. Eur Respir J 2001;17
: 1294-1313[Abstract/Free Full Text]
- Arakawa H, Webb WR. Expiratory high-resolution CT scan.
Radiol Clin North Am 1998;36
: 189-209[CrossRef][Medline]
- Goldin JG. Quantitative CT of emphysema and the airways.
J Thorac Imaging 2004;19
: 235-240[CrossRef][Medline]
- Arakawa H, Niimi H, Kurihara Y, Nakajima Y, Webb WR. Expiratory
high-resolution CT: diagnostic value in diffuse lung diseases.
AJR 2000; 175:1537
-1543[Free Full Text]
- Arakawa H, Webb WR, McCowin M, Katsou G, Lee KN, Seitz RF.
Inhomogeneous lung attenuation at thin-section CT: diagnostic value of
expiratory scans. Radiology 1998;206
: 89-94[Abstract/Free Full Text]
- Ng CS, Desai SR, Rubens MB, Padley SP, Wells AU, Hansell DM. Visual
quantitation and observer variation of signs of small airways disease at
inspiratory and expiratory CT. J Thorac Imaging1999; 14:279
-285[Medline]
- Stern EJ, Frank MS. Small-airway diseases of the lungs: findings at
expiratory CT. AJR 1994;163
: 37-41[Abstract/Free Full Text]
- Ishii M, Fischer MC, Emami K, et al. Hyperpolarized helium-3 MR
imaging of pulmonary function. Radiol Clin North Am2005; 43:235
-246[CrossRef][Medline]
- Altes TA, Rehm PK, Harrell F, Salerno M, Daniel TM, De Lange EE.
Ventilation imaging of the lung: comparison of hyperpolarized helium-3 MR
imaging with Xe-133 scintigraphy. Acad Radiol2004; 11:729
-734[CrossRef][Medline]
- Travaline JM, Maurer AH, Charkes ND, Urbain JL, Furukawa S, Criner
GJ. Quantitation of regional ventilation during the washout phase of lung
scintigraphy: measurement in patients with severe COPD before and after
bilateral lung volume reduction surgery. Chest2000; 118:721
-727[Abstract/Free Full Text]
- Park KJ, Bergin CJ, Clausen JL. Quantitation of emphysema with
three-dimensional CT densitometry: comparison with two-dimensional analysis,
visual emphysema scores, and pulmonary function test results.
Radiology 1999;211
: 541-547[Abstract/Free Full Text]
- Hoffman EA, McLennan G. Assessment of the pulmonary
structure-function relationship and clinical outcomes measures: quantitative
volumetric CT of the lung. Acad Radiol1997; 4:758
-776[CrossRef][Medline]
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
Top
Abstract
Introduction
