| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1 The Woolcock Institute of Medical Research, and 2 The Cooperative Research Centre for Asthma, and 3 University of Sydney, and 4 The Key Centre for Microscopy and Microanalysis, University of Sydney, Sydney, and 5 Dept of Radiology, Royal Prince Alfred Hospital, Camperdown, and 6 Dept of Respiratory Medicine, Royal North Shore Hospital, St. Leonards, Australia.
CORRESPONDENCE: G. G. King, The Woolcock Institute of Medical Research, University of Sydney, Sydney 2006, Australia. Fax: 61 299066391. E-mail: ggk{at}woolcock.org.au
Keywords: Computer-assisted image processing, imaging, phantoms, validation studies
Received: February 3, 2005
Accepted June 30, 2006
 |
ABSTRACT |
|---|
 TOP ABSTRACT METHODSRESULTSDISCUSSIONREFERENCES |
|---|
Inflation-fixed porcine lung cubes were scanned by HRCT and micro-CT. A total of 59 lumen area (Ai), 30 wall area (Aaw) and 11 lumen volume (Vi) measurements were made. Ai was measured from the cut surface of 11 airways by morphometry. Airways in scanned images were matched using branching points. After calibration, the errors of Ai, Aaw and Vi HRCT measurements were determined.
The current authors found a systematic, size-dependent underestimation of Ai and overestimation of Aaw from HRCT measurements. This was used to calibrate an HRCT measurement algorithm. The 95% limits of agreement of subsequent measurements were ±3.2 mm2 for Ai, ±4.3 mm2 for Aaw, and ±11.2 mm3 for Vi with no systematic error. Morphometric measurements agreed with micro-CT (±2.5 mm2) without systematic error.
In conclusion, micro-computed tomography image data from inflation-fixed airways can be used as calibration standards for three-dimensional lumen volume measurements from high-resolution computed tomography, while morphometry is acceptable for two-dimensional measurements. The image dataset could be used to validate other developmental three-dimensional segmentation algorithms.
Due to multidetector technology, volumetric high-resolution computed tomography (HRCT) image data are routinely acquired, and have been used for quantitative image analysis of airways and lung tissue in many studies 1–9. Direct measurements of airway structures can be made from HRCT image data, making it useful for studying asthma, cystic fibrosis and chronic obstructive pulmonary disease. Airway dimensions have been measured from HRCT image data using customised computer-analysis algorithms that have been validated using airway phantoms of known dimensions constructed from nonanimal materials, such as plexiglass and sweet potato 4, 8, 10–12. In these studies, planimetric measurements were made of a cut surface of the phantom. It was therefore reasonably assumed that the dimensions would be completely uniform along the length of the airway phantom and that orientation in the z-axis of the scanning plane could be accurately controlled.
Validation of airway measurement algorithms advanced further with the use of explanted animal lungs as calibration standards. Porcine lungs wet fixed with formalin 13 have been used and the current authors have previously used formalin inflation-fixed porcine lungs 8 for validation. The use of liquid-fixed resected human lung tissue for validation has also recently been described 13. Explanted lung preparations are more suitable than artificial phantoms for validation because they closely approximate the range of tissue densities and dimensions, and the complex structure of lung tissues in vivo. This includes features such as mucosal folding and irregularities in airways, which are not present in artificial phantoms. This difference compared with artificial material phantoms is, in turn, associated with different estimates of measurement errors for HRCT 8. HRCT measurements are affected by volume averaging, particularly in the scanner's z-axis. This is due to the complexity of the structures, range of tissue densities, inability to control airway orientation in the z-axis during scanning, and the fact that the z-axis is the longest dimension of the voxel (fig. 1
). To calibrate three-dimensional (3D) HRCT measurements of airways, the current authors required a 3D standard of known dimensions so that the effects of volume averaging would be accounted for in the calibration process, while also allowing measurement of the precision and accuracy of the HRCT method. As morphometry is two-dimensional (2D), there is by definition no volume averaging in the z-axis to affect this measurement.
|
). The primary aim of the present study was to validate micro-CT against the current calibration standard of morphometry of the cut surface of the lung and calliper measurement of cylindrical rods. The secondary aim was to use measurements of lumen area (Ai) and wall area (Aaw) in inflation-fixed porcine airways from micro-CT images to calibrate an HRCT airway segmentation algorithm. The final aim was to assess the precision and accuracy of 3D measurements of airway volume (Vi) made using the calibrated segmentation algorithm.
|
|
METHODS |
|---|
|
TOPABSTRACTMETHODSRESULTSDISCUSSIONREFERENCES |
|---|
).
|
). Fixation time was 
2 h, depending on the size of the specimen.
|
Digital imaging of the lung
Immediately prior to being wrapped and coated with paraffin, the cut surfaces of the lung cubes were photographed through a dissecting microscope with a 3.34-megapixel digital camera (Nikon Coolpix 995 with MDC2 Relay Lens; Nikon, Tokyo, Japan) and images were saved as 1024x1024-pixel bitmap files. After the lung cubes were wrapped and coated in paraffin they were scanned in both the HRCT (GE 4-slice Lightspeed; GE Medical Systems, Minneapolis, MN, USA) and SkyScan 1072 micro-CT desktop systems (SkyScan, Aartselaar, Belgium). HRCT settings were 100 kV and 100 mAs at a slice thickness of 1.25 mm and a pitch of one in helical mode. Images were reconstructed using the GE high spatial frequency "Bone" algorithm at a field of view (FOV) of 18 cm, yielding a voxel size of 0.35x0.35x1.25 mm. Images were archived in DICOM format for analyses later. Micro-CT settings were 100 kV and 100 mAs at a slice thickness of 19 µm and a FOV of 19 mm, yielding a voxel size of 19x19x19µm. Each cube was scanned every 0.23° as it rotated around its vertical axis until it had rotated 180°, at an exposure time of 0.9 s per scan. Images were reconstructed using the ConeRec software (SkyScan). Reconstructed image data were archived in bitmap format for analyses later.
Image analysis
A single operator (J.R. Dame Carroll) measured Ai and Aaw of airways visible in cross-section in the digital photographic bitmap images of the cut surface of the lung, manually tracing the inner and outer edges of the airway wall with the computer mouse using ImageJ software (National Institute of Mental Health, Bethesda, MD, USA). Each airway was traced twice to assess for measurement repeatability, and the average of the two tracings was used for analysis.
Ai and Aaw measurements were made directly from the HRCT image data (DICOM format) using an in-house, custom-written software program. The HRCT lung images were displayed on-screen and the airway of interest was identified manually. The centroid of the airway was automatically identified, and 20 radial spikes were created running out from it. The inner and outer edges of the airway wall were defined as the point along the radial spikes at which the greatest rate of change in density occurred, i.e. the transition from air to soft tissue.
Ai and Aaw were measured from micro-CT bitmap images using the same method as for morphometry, tracing the inner and outer edges of the airway wall using the ImageJ software. The lumen and wall areas were calculated as the product of the pixel size (19x19 µm2) and the number of voxels within the traced border. A single operator (J.R. Dame Carroll) traced each airway border twice, and the average of the two measurements was used for analysis. For comparisons of micro-CT and morphometry, all airways that were visible in both the micro-CT FOV and the digital photographs of the cut surface of the lung specimens were measured. This resulted in a range of airway sizes comparable to those measured by the current authors in a previous HRCT study of asthmatic and normal airways 16.
For comparisons between HRCT and micro-CT, airways were matched using airway branch points as anatomical markers. Since the HRCT slice thickness of 1.25 mm was >60 times thicker than the micro-CT slice thickness of 19 µm, every 30th micro-CT image was used to give an effective slice thickness of 570 µm. To calibrate Ai measurements from HRCT, the average of three consecutive Ai and Aaw measurements from micro-CT images which covered a thickness of 1.14 mm, was compared with each corresponding HRCT slice of thickness 1.25 mm. Micro-CT and HRCT measurements of Vi were calculated as the sum of all lumen voxels in the airway. Voxel dimensions for micro-CT were 19x19x570 µm, since only every 30th image was used (30 slicesx19µm = 570µm). The voxel dimensions for HRCT were 0.35x0.35x1.25 mm, since slice thickness was 1.25 mm.
Statistical analysis
All data are shown as mean±95% confidence interval, unless otherwise specified.
Intra-observer variation
Intra-observer variations in 72 calliper measurements of the rods, in Ai and 
Ai measurements from morphometry in 29 airways, and in Ai measurements from micro-CT in 30 airways, were determined by the methods of Bland and Altman 17, 18 and expressed as repeatability, being the 95% limits of agreement.
Agreement between imaging methods
The degree of agreement of the Ai and Aaw measurements made by the semi-automated HRCT algorithm and the micro-CT technique was calculated by the method of Bland and Altman 17, 18 and expressed as the 95% limits of agreement, being t0.05xSD of the differences between micro-CT and HRCT measurements, where t0.05 is the critical t-value corresponding to the sample size at the 0.05 level of significance. Agreement between micro-CT measurements and calliper measurements of the artificial material rods was similarly calculated.
Calibration of HRCT measurements
The relationship between Ai and Aaw measurements from micro-CT and HRCT was used as a calibration factor for the HRCT measurements and the subsequent 95% limits of agreement were recalculated as above.
|
RESULTS |
|---|
|
TOPABSTRACTMETHODSRESULTSDISCUSSIONREFERENCES |
|---|
Morphometry versus micro-CT of airways
There was good agreement between Ai measurements of 11 airways from the digital photographs of the cut lung surface compared with measurements from micro-CT (fig. 5). Within the range of airway sizes the current authors were able to study (0.9–36.4 mm2), the mean difference was 0.14±0.6 mm2 (p = 0.69), independent of airway size; the 95% limits of agreement were ±2.19 mm2. The intra-operator (J.R. Dame Carroll) repeatabilities of Ai and Ai measurements from morphometry were ±0.555 mm2 or 3.1% and ±0.066 mm or 1.6%, respectively. The intra-operator repeatability of Ai measurements from micro-CT was ±0.557 mm2 or 1.9%.
|
). From a plot of micro-CT Ai and HRCT Ai (data not shown; r2 = 0.96, p<0.0001), the difference between the Ai measured from both methods was shown to fit the following equation:
|
After this equation, or calibration factor, was applied to all HRCT Ai measurements, the 95% limits of agreement of Ai measurement were ±3.2 mm2 (fig. 6c
). Percentage errors (fig. 6b and d) were, as expected, larger for smaller airways. The percentage error was greatest in airways with Ai <10 mm2 (internal diameter <3.5 mm).
A total of 30 airway wall area measurements from micro-CT and HRCT were compared (mean Ai 29.4±9.6 mm2, idealised diameter 6.5±1.0 mm). Airways with Aaw <10 mm2 could not be segmented for wall area reliably with the current authors’ segmentation algorithm, and were therefore excluded from the analysis. The excluded airways had a corresponding Ai <5 mm2, and an idealised diameter of <2.5 mm. The overestimation of Aaw when measured using HRCT was size dependent and linear (fig. 7a). From a plot of micro-CT Aaw and HRCT Aaw (data not shown; r2 = 0.97; p<0.0001), the difference in Aaw was found to fit the following equation:
|
After calibration of the HRCT Aaw measurements, the 95% limits of agreement of Aaw measurements were ±4.3 mm2 (fig. 7c). Percentage errors were again larger for smaller airways (fig. 7b and d). The percentage error was greatest in airways with an Aaw <10 mm2.
Airway volume was measured using the calibrated segmentation algorithm for all airways that had at least three consecutive HRCT images in which segmentation could be done. Airway lumen volume was calculated as the product of voxel volume (0.35x0.35x1.25 mm) and the number of voxels within the lumen from all contiguous slices along the airway volume. Volume measurements from HRCT images from 11 airways using the calibrated measurements agreed closely with micro-CT Vi measurements with no systematic error or size dependence (fig. 8a). The 95% limits of agreement were ±11.2 mm3. Again, when errors were expressed as a percentage of airway volume, the relative errors were larger for smaller airways (fig. 8b).
|
|
DISCUSSION |
|---|
|
TOPABSTRACTMETHODSRESULTSDISCUSSIONREFERENCES |
|---|
There are several possible sources of errors in airway measurements from HRCT image data. These include volume averaging (which can be exacerbated by more acute angles of airway orientation relative to the scanning plane); the scanner’s point spread functions and reconstruction algorithms; and the lumen segmentation methods 19. It is possible to measure the angles of orientation of airways from volumetric HRCT data to correct for volume averaging 20, potentially reducing variation in HRCT measurements. The present authors were unable to do this with the current iteration of their analysis algorithm, but it needs to be addressed in future studies. It is likely that these sources of error may differ between manufacturers and models of HRCT machines. Therefore, it may be useful to examine the variations in CT image data arising from differences in clinical CT scanners, as well as those arising from differences in reconstruction algorithms. The magnitude of the variation is not known, nor is it known whether it is clinically significant, although it is common for serial chest CT scans for monitoring disease to be carried out using different scanners.
The manual outlining of airway lumen borders on micro-CT image data might have been an additional source of error, but the present authors think this is likely to be very small. This is because the very small errors in measurements of diameter obtained from the cylindrical rods are an order of magnitude smaller than the errors introduced through HRCT image acquisition and analysis. Measurements of Ai, Aaw and Vi made from micro-CT were unaffected by window settings because of the great spatial resolution and the great contrast at the lumen interface. It is interesting that the 95% limits of agreement between micro-CT and surface morphometry (±2.5 mm2) were similar to the agreement between micro-CT and HRCT (±3.2 mm2). The disagreement between micro-CT and surface morphometry may be due to the reconstruction plane of micro-CT being slightly out of alignment with the cut surface of the lung which was photographed. The implication is that similar errors are likely to have occurred when comparing micro-CT with HRCT and this also suggests that the comparison method could be improved in the future by using a 3D registration process to remove this potential source of error. It should also be noted that in some cases, the airways tended to taper towards the surface of the lung cubes, with the airway reaching its largest diameter toward the centre of the lung cube. These airways were not included in the comparison between micro-CT and morphometry, because they were not visible on the cut surface of the lung.
Volume averaging is the factor most likely to account for the bulk of the errors in Ai measurements. The effects of volume averaging on apparent Ai are likely to be greatest in the z-axis in HRCT, because the z-dimension of the HRCT voxel (1.25 mm) is almost four times as long as the x- and y-dimensions (0.35 mm). This is relevant because airways are usually not parallel to the z-axis in human HRCT studies, nor were they in the current study. As a result, the effects of volume averaging and, hence, the magnitude of the underestimation of lumen area are greater as airway size decreases and the angle of orientation increases (fig. 1) 8, 19, 21. The use of micro-CT for calibration allowed the current authors to account for the effects of volume averaging, which would not have been possible with 2D measurements. Furthermore, when spiral CT is used, rather than conventional axial HRCT, there is a blurring of structures in the z-axis due to the continuous movement of patients through the rotating tube. Finally, Ai, Aaw and Vi were measured from HRCT image data in Hounsfield Units, a measure of tissue density 19. They were, therefore, unaffected by window level or width settings, which can affect visual analyses of photographed images 11, 22.
The current authors found that the lower limit of airway size for which lumen measurements could be made with accuracy, in terms of percentage errors, was 10 mm2. This corresponds to an idealised internal diameter of 3.5 mm in the current data set, i.e. fourth-generation or segmental airways. Airway wall area measurements were accurate in airways with wall areas >10 mm2. This corresponds to airways with internal diameters of 2.5 mm, i.e. fifth-generation or sub-segmental airways. The implication is that greater numbers of airways need to be measured when looking at any percentage changes in Ai or Aaw associated with interventions, for airways smaller than those specified previously. However, there is greater power if absolute changes in either Ai or Aaw are being measured. It is essential to be able to correct or calibrate the measurements made from HRCT, and to know the accuracy of the subsequent calibrated measurements so that studies that are designed to look for changes in airway dimensions are powered appropriately. For instance, the current authors found that even after calibration, smaller airways continued to have large errors associated with them when expressed as a percentage of airway size. Consequently, a much larger number of small airways would need to be studied in patients to detect a real effect of an intervention on smaller airway dimensions than in larger airways where the error as a percentage of total area is much smaller.
The current authors used one airway segmentation method for HRCT, although there are many airway analysis algorithms, all of which are based on slightly different segmentation criteria 1, 3–8, 21. In the present study, the point at which the density change over two pixels was maximal along the radial spike was used, while others have used the more common "full width at half-max" method. The former method is likely to result in a greater underestimation of the lumen, but it was chosen for the current study because it appeared to segment the lumen reliably even in small airways. Another possible factor is the number of radial spikes used in the segmentation. The ideal number may vary depending on the edge-detection method and is worthy of further study. Such systematic errors relating to a particular methodology of finding the airway lumen edge can be corrected by comparison with calibration standards generated by studies, such as the present one. The precision of different segmentation methods can then be meaningfully compared. The data set of HRCT airway images could be used in the future to compare, and possibly standardise, measurements made using the different segmentation algorithms. Given the increasing likelihood that computerised segmentation will be used routinely in clinical practice to measure airway dimensions, documentation of the differences in Ai and Aaw measurements associated with different segmentation methods needs to be done using calibration standards in follow-up studies.
The formalin steam fixation method did not significantly alter lung density, which makes this validation method likely to be more accurate and relevant to in vivo lung imaging than validation with artificial material phantoms. Lung densities pre- and post-fixation were 0.34 g·mL-1 and 0.38 g·mL-1, respectively. The tissue density of the formalin-fixed pig lung was compared with human in vivo lung as measured by HRCT and found to be similar. The lung cubes were very delicate and needed to be handled extremely gently as they were moist and not completely stiff. A very sharp and thin knife was used to slice the lung, but a slight distortion of the parenchyma at the site of slicing was noticed in the lung cubes. Since all of the analysis was performed after slicing, the effect of the small distortions at the sliced edge was constant across imaging techniques and not considered an important issue. Wet fixation (a method commonly used for excised lungs, in which formalin solution is pumped into the lung or lobe) was considered unsuitable for this validation because it results in fluid-filled alveoli, thereby altering mean lung density, as well as introducing additional fluid into the airway lining. Dry fixation was similarly considered to be unsuitable because it causes desiccation of the lung. Therefore, the calibration standard described in the present study gives a more "real life" measure of the artefacts in airway appearance and measurement that would occur in HRCT scans of human subjects. Another strength of the present study is that the micro-CT, morphometry and HRCT were performed after fixation, which enabled the current authors to determine measurement errors that were associated with image analysis, since there could be no errors due to fixation artefacts.
HRCT provides the unique opportunity to monitor the response of airways to treatment or other stimuli by making direct measurements on airways in vivo over time. It is essential that airway measurements are accurate and that the accuracy of the measurements is independent of airway size, since an airway that is reduced in size owing to moderate or severe increases in smooth muscle tone, inflammation or wall thickening, may experience large changes in lumen size if the inflammation and bronchoconstriction resolve. If the accuracy of the measurements is largely size dependent, then the magnitude of the change in airway dimensions will be masked. It is widely accepted that HRCT measurements of lumen area tend to underestimate actual lumen area in a size-dependent fashion such that the absolute error increases as airway size increases, and HRCT measurements of airway wall tend to overestimate actual airway wall dimensions in the same size-dependent fashion 8. Therefore, if a patient was scanned during an acute exacerbation of asthma while the airways were inflamed and constricted, then scanned again after recovery and treatment with anti-inflammatory medication, the errors associated with the two scans of the same airway would differ greatly. This process also works in reverse, such that the underestimation of airway lumen size before bronchoconstriction will be larger than the underestimation error associated with the same airway measured after constriction, so the absolute change in lumen area would itself be underestimated.
In the present study, a software standard, not a hardware standard, has been developed. A data set of micro-CT and HRCT images has been created. This can be provided to other researchers to be analysed using their own segmentation software. This is not a hardware comparison tool for CT scanners. In order to make this method applicable to other scanners, a better, longer-lasting method for preserving the lung tissue samples after they are fixed needs to be developed so that the lung cubes can be sent to different research sites to be scanned by different scanners. The paraffin wax used to seal the cubes in the present study is air permeable, and the current authors found that after a number of weeks the tissue had shrunk noticeably when the cubes were scanned again.
In summary, the current authors have used micro-computed tomography as the standard to calibrate and then measure the accuracy of a computerised segmentation algorithm for measuring lumen area, wall area and lumen volume using high-resolution computed tomography. The importance of this work is in providing a three-dimensional standard for calibration, which allows more accurate in vivo measurements to be made of airway dimensions from using high-resolution computed tomography. This will allow clinicians and researchers to assess airway pathophysiology more effectively. The micro-computed tomography method is more technically complex and only available in large academic centres, but clearly its value is as a three-dimensional calibration standard which allows comparison of different models of CT scanner, reconstruction methods and segmentation algorithms. The current authors’ segmentation algorithm, which utilised a radial-spike method, was shown to underestimate lumen area and overestimate wall area in a size-dependent manner, which is in keeping with results of previous studies by the current authors and others. Having an unbiased three-dimensional method will allow researchers to obtain important data on airway lengths, lumen areas and wall areas that are useful, for instance, in modelling studies and in longitudinal treatment studies. The close agreement between morphometry and micro-computed tomography measurements suggests that the two methods are comparable for use as two-dimensional standards. The ultimate goal is that a standard calibration method can be used between research groups to address differences in airway measurements due to different hardware, scanning methodology and reconstruction, and segmentation algorithms.
|
REFERENCES |
|---|
|
TOPABSTRACTMETHODSRESULTSDISCUSSIONREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  B. Thompson, G. King, and R. Harding Commentary on "The role of the large airways on smooth muscle contraction in asthma" J Appl Physiol, October 1, 2007; 103(4): 1465 - 1465. [Full Text] [PDF]
|
|
|
|
 |
|
 T. E. Robinson Computed Tomography Scanning Techniques for the Evaluation of Cystic Fibrosis Lung Disease Proceedings of the ATS, August 1, 2007; 4(4): 310 - 315. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
