|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
 |
ABSTRACT |
|---|
TOP
ABSTRACT
INTRODUCTION
METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
We have developed an objective, reproducible, and automated means for the regional evaluation of the pulmonary parenchyma from computed tomography (CT) scans. This method, known as the Adaptive Multiple Feature Method (AMFM) assesses as many as 22 independent texture features in order to classify a tissue pattern. In this study, the six tissue patterns characterized were: honeycombing, ground glass, bronchovascular, nodular, emphysemalike, and normal. The lung slices were evaluated regionally using 31 × 31 pixel regions of interest. In each region of interest, an optimal subset of texture features was evaluated to determine which of the six patterns the region could be characterized as. The computer output was validated against experienced observers in three settings. In the first two readings, when the observers were blinded to the primary diagnosis of the subject, the average computer versus observer agreement was 44.4 ± 8.7% and 47.3 ± 9.0%, respectively. The average interobserver agreement for the same two readings were 48.8 ± 9.1% and 52.2 ± 10.0%, respectively. In the third reading, when the observers were provided the primary diagnosis, the average computer versus observer agreement was 51.7 ± 2.9% where as the average interobserver agreement was 53.9 ± 6.2%. The kappa statistic of agreement between the regions, for which the majority of the observers agreed on a pattern type, versus the computer was found to be 0.62. For regional tissue characterization, the AMFM is 100% reproducible and performs as well as experienced human observers who have been told the patient diagnosis.
| |
INTRODUCTION |
|---|
|
TOP
ABSTRACT
INTRODUCTION
METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
Parenchymal lung diseases are common and produce disabling and frequently fatal illnesses. The common parenchymal lung diseases include subacute inflammatory disorders such as sarcoidosis and idiopathic pulmonary fibrosis, and disorders such as pulmonary emphysema. With the advent of new interventions for halting or reversing lung disease progression, it has become critical that a method be developed that will allow for the quantitative tracking of these diseases (both globally and within lung regions), and for the detection of these disease processes at their earliest stages.
We have previously developed and demonstrated the use of an Adaptive Multiple Feature Method (AMFM) for the computer-aided characterization of X-ray computed tomography lung slices. The AMFM can be trained to identify different lung tissue types using multiple texture features computed from the CT data. Previously published computer-based methods for evaluating CT images of the lung (1) have used information from only one first-order textural feature such as density and therefore do not take full advantage of the complexity of the lung parenchyma as shown and recorded in digital form by HRCT. The AMFM was first applied for differentiating normal subjects versus emphysematous subjects and the performance of the AMFM was demonstrated to be superior as compared with these previously reported methods (7). In a following study (8, 9), the AMFM was applied for a more complex four-subject group differentiation task involving normal subjects along with subjects with emphysema, idiopathic pulmonary fibrosis (IPF), and sarcoidosis. The AMFM again performed substantially better in this discrimination when compared with other methods. In these studies, the AMFM was trained to recognize broad disease categories as opposed to more disease-independent tissue patterns.
We have further developed the AMFM to automatically identify the normal lung pattern and lung patterns commonly associated with parenchymal lung diseases and have applied this tissue typing on a regional basis. The six patterns, which the AMFM has been currently trained to recognize, are: honeycombing, ground glass, bronchovascular, nodular, emphysemalike, and normal. These patterns are commonly utilized in clinical practice to identify regional abnormalities in the lung based upon CT findings since it is the relative presence of such patterns that contribute to the assessment of disease presence and progression.
| |
METHODS |
|---|
|
TOP
ABSTRACT
INTRODUCTION
METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
Subjects
In this study, 20 normal subjects, 13 with emphysema, 19 with IPF, and 20 with sarcoidosis subjects were assessed. Informed consent was obtained for all studies. The normal subjects were volunteers who had no history of pulmonary disease, had normal lung function, were lifelong nonsmokers, and were receiving no medications. The emphysematus subjects were chosen from among the subjects undergoing lung reduction for emphysema. These subjects all had established severe emphysema, and the CT scans were performed as part of their preoperative evaluation for lung reduction. The normal subjects and those with IPF and sarcoidosis were selected from the database of the Specialized Center of Research Program in interstitial and occupational lung disease at the University of Iowa. These subjects were selected to have a range of lung function abnormalities and had well-documented disease with clinical, physiologic, radiologic, and pathologic features of the disease. Although all subjects with lung pathology had well- characterized disease, regional parenchymal characteristics ranged in appearance from normal to a clearly defined presence of disease.
Pulmonary Function Tests
The pulmonary function tests (PFT) consisted of standard spirometry using the Medical Graphics 1070 system (Medical Graphics, St. Paul, MN) and lung volumes via plethysmography using a Medical Graphics 1085 system. Single-breath diffusing capacity (DLCO) was tested using the Medical Graphics 1070 system. The lung function tests were performed using standard protocols, and the American Thoracic Society guidelines were used to determine acceptability (10). The predicted normal values used were those of Morris and coworkers (11) for spirometry, Goldman and Becklake (12) for lung volumes, and Van Ganse and colleagues (13) for diffusing capacity. The average PFT values for all the subject groups are presented in Reference 9.
Image Data Acquisition
CT scans were acquired with an electron beam CT scanner (Imatron Fastrac C-150 XL; Imatron, South San Francisco, CA) at the University of Iowa Hospitals and Clinics. Scans of the normal subjects and those with IPF and sarcoidosis were acquired in the prone position, and the scans of the emphysematus subjects were acquired in the supine position. The field of view ranged from 300 to 400 mm, dependent upon patient size, and slice collimation was 3 mm. The images were reconstructed to 512 × 512 pixels. The grey level resolution of the images was 11 bit. At the time of these CT studies, 3-mm slice collimation was the thinnest achievable by the EBCT scanner.
Image Data Analysis
The AMFM for examining the lung parenchyma from CT scans was partially developed and reported recently (7). The AMFM is a pattern recognition system that discriminates lung patterns on the basis of their textures. The AMFM has two stages. The first stage involves training the AMFM to recognize different HRCT-derived tissue patterns using a preselected database containing representative tissue types; this is referred to as the training set. The second stage involves a new set of HRCT-derived data to be analyzed; referred to as the test set. A feature selection program called the divergence feature method (14) coupled with correlation analysis is used to rank the features on the basis of their power to discriminate between the tissue patterns. The purpose of this step is to identify which subset of features, when taken in combination, provide the best discrimination between the different patterns. The divergence measure for each feature is computed using the mean and the variance of the feature for each tissue pattern. A higher divergence measure indicates a greater discriminatory power. The correlation coefficient between two features is computed using the sample covariances. If two features are highly correlated, the feature with the lower divergence measure is discarded. With the correlated features removed, the features with the highest divergence measures are chosen for the classification task. The number of features chosen is dependent on the classification accuracy on the training set. The feature selection is performed on the training set prior to classification step on the test set. A Bayesian classifier (15) is trained on the samples in the database using the optimal combination of texture features (from the feature selection program). In the testing phase, the Bayesian classifier identifies each sample in the test set as belonging to one of the available tissue pattern types with a certain recorded degree of confidence. The Bayesian classifier is a nonlinear statistical classifier and is based on maximal likelihood estimates. Using this classifier, a sample is classified as belonging to a certain tissue pattern type if the probability that it belongs to this type is higher than the probability that it belongs to any other tissue pattern type. The training samples are used to evaluate all the required a priori probabilities and statistics, which in turn are used to compute the classification probability of a test sample.
Seventeen texture features have been described (7), among which were statistical texture features and a fractal feature. The statistical features were the grey level distribution features, run-length features, and co-occurrence matrix features. The grey level distribution features were the mean, variance, skewness, kurtosis, and the grey level entropy. These measures describe the occurrence frequency of all the grey levels in the region of interest. The run-length features were the short-run emphasis, the long-run emphasis, grey level nonuniformity, run-length nonuniformity, and run percentage. The run-length features describe the heterogeneity and tonal distribution of the grey levels in the region of interest. The co-occurrence matrix features were the angular second moment, correlation, contrast, entropy, inertia, and inverse difference moment. The co-occurrence matrix measures describe the overall spatial relationships that the grey tones have to one another in the region of interest. The fractal feature used was the geometric fractal dimension (GFD), which indicates the degree of roughness of the texture.
In this present study, adding five more novel fractal features has further enhanced the discrimination power of the AMFM. The pulmonary tree can be modeled as a geometric fractal (geometric self-similarity over different scales) (16, 17) as well as a stochastic fractal (statistical self-similarity over different scales) (18). We combined techniques available for both types of fractals to design unique fractal features (19). To compute the GFD, we used the computations of the stochastic fractal dimensions (SFD) as an intermediate step. Fractional Brownian motion model principles were used to assess local SFDs for each pixel in the region of interest (ROI), which were then mapped to grey-level values. This provided an edge-enhanced image to compute both the GFD and the SFD features. Using both types of fractals the structure as well as the texture of the lungs were characterized through fractal properties. The SFD features used were the SFD mean, SFD variance, SFD skewness, SFD kurtosis, and SFD entropy. The technique is summarized as follows.
The local fractal dimensions (stochastic fractal dimension) of the
region of interest are calculated using fractional Brownian motion
model concepts. A 5 × 5 pixel block is centered at each pixel in the
ROI. The average absolute intensity differences between all pairs of
pixels at different distances are recorded. Then, the average absolute
intensity differences of the pixel pairs at different pixel pair distances
are plotted on a log-log scale and the SFD is estimated from the slope
of the best fit line (SFD = 3 
slope). This SFD is assigned to the
pixel at which the block is centered. Every pixel in the region of interest is assigned a SFD in this fashion. The SFD of each pixel is then
converted to an intensity grey level, finally obtaining a fractal dimension distribution transformed image (20). This SFD image enhances the edges of the original ROI and is used as a preprocessing step to
compute the GFD. The edge enhanced transformed grey-level image is converted into a binary image by simple grey-level thresholding, with at a preset grey level, chosen empirically. The box-counting technique is used to calculate a single GFD for the entire ROI. The box-counting procedure involves superimposing the image with superpixels of increasing size (e). Counting the number of superpixels, N(e),
encountered by the structure for each size of the superpixel and plotting N(e) against (e) on a log-log scale, the GFD is estimated from the
slope according to: N(e) = K(1/e)GFD. Superpixels for sizes 1 to 10 were used.
The additional fractal-based features used in this study are the grey-level distribution features computed on the SFD image, namely, SFD mean, SFD variance, SFD skewness, SFD kurtosis, and SFD entropy. The same formulae as in Reference 21 are used for the SFD feature computations.
The grey-level distribution features and fractal features were computed on the original grey scale data of the sample, whereas the run-length and the co-occurrence matrix features were computed on the image produced after an image preprocessing step, edgmentation, as previously described (22). The features used in this study are tabulated in the Appendix .
Image Data Analysis
Regional analysis in observer-defined regions. Two expert observers were asked to review the CT scans of the 72 subjects and independently outline regions from the HRCT scans typifying each of the six common, yet different, tissue patterns (23). These patterns were honeycombing, ground glass, bronchovascular, nodular, emphysemalike, and normal. Each observer then independently reviewed the other's classification, and only areas where agreement was reached regarding the tissue type were used as part of the training set. Example images with the six patterns outlined by an observer are shown in Figure 1. These areas were divided by computer into 31 × 31 pixel blocks (overlapping 15 × 15), this area being predetermined in a series of experiments to be the appropriate one (data not shown). The size of the pixel block was based on minimizing area while maintaining classification sensitivity. Half of these regions were randomly chosen and entered into the training set. The remaining samples were entered into an initial test set: Test Set One. The AMFM then determined the optimal subset of features, and the classifier was trained to use these to recognize the different six tissue patterns.
|
Image Data Analysis
Regional analysis in entire CT slices. The pattern recognition process was then applied to the analysis of whole lung HRCT slices from normal subjects and from subjects with known emphysema, IPF, and sarcoidosis to determine areas of normal lung and of disease. This HRCT data (12 slices) constituted Test Set Two. A 31 × 31 pixel block (overlapping by 15 × 15) was translated across the image, and the tissue label identified by the computer for each block was assigned to a 15 × 15 block at its center. Assigning a label for each area was performed in two ways. In the first method, a label corresponding to one of the six tissue pattern types was assigned by the AMFM irrespective of the confidence in the label, ie., with no limits on the confidence. In the second method, a seventh label, called indeterminate, was given by the AMFM if the confidence in the assigned label being any one of the six tissue patterns was less than 90.0%. The results of classification were displayed in a color-coded fashion. Both the computer results were compared with readings recorded on a computer by five trained and experienced human observers in a blinded design.
The observers consisted of one experienced chest radiologist, one general radiologist, two pulmonary physicians highly experienced in HRCT studies of parenchymal disorders, and one general pulmonary physician. The observers were trained with regard to the use of the computer software, and then were subsequently asked to evaluate 12 separate HRCT-derived lung slices on two separate occasions 6 wk apart. This helped the observers familiarize themselves with the system and enabled assessment of interobserver and intraobserver variation. In the first two readings, the observers were blinded to the patients' clinical diagnoses as well as to the computer output. Observers who had an intraobserver agreement of greater than 50% were retained for a third reading 6 wk later. In the third reading, the observers were given the primary patient diagnosis and were asked to evaluate six slices (that had been included previously in the 12 slices). The observers still had no knowledge as to the computer output, or as to how their previous classification compared with that of the computer, or with each other.
The observers evaluated each slice by labeling each pixel block (15 × 15) on each chosen CT slice. The labeling was performed using a convenient in-house-developed graphic user interface. The interface allowed the user to view the slice along with an overlaid grid. Each observer was asked to label each block on the grid using a mouse. As in the case of the computer-labeled images, the different labels represented the available choices of lung patterns, i.e., honeycombing, ground glass, bronchovascular, nodular, emphysemalike, and normal. A seventh label, called indeterminate, was also provided as a choice. Unlimited amount of time and unlimited editing was allowed. The observers were given the image at optimal lung window settings (contrast) but were allowed to freely change the window and level settings. The observers were free to magnify the whole CT slice as well as the region of interest they were evaluating.
Statistical Analysis
For the image analysis in observer-defined regions, the sensitivity, specificity, and accuracy (24) on Test Set One were computed. The sensitivity is defined as the percentage of samples belonging to a tissue pattern type correctly classified as such. Specificity is the percentage of samples not belonging to a tissue pattern type correctly classified as such. Accuracy is the percentage of samples belonging to all tissue pattern types correctly classified.
For the regional analysis on entire CT slices, statistical analyses regarding interobserver agreement and computer-observer agreements were computed. The agreements were recorded in terms of the percentage of blocks receiving the same label. The interobserver agreement was reported by computing the agreement between observers, taking two observers at a time. For the third reading, the kappa statistic (25) of agreement between the observers was also reported. The kappa statistic accounts for agreement caused by chance. A "gold standard" was defined as the regions for which at least three of the four human observers agreed as to the tissue pattern label. For these regions, the percent agreement between the majority label given by the observers and the computer as well as the kappa statistic was calculated.
| |
RESULTS |
|---|
|
TOP
ABSTRACT
INTRODUCTION
METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
CT Data Analysis
Regional analysis in observer-defined regions (Test Set One). The classification results for Test Set One are shown in Figure 2. The training set and Test Set One (derived from the 72 subjects) each had 40 honeycombing, 122 ground glass, 74 bronchovascular, 83 nodular, 265 emphysemalike, and 315 normal samples. Sensitivities greater than 82.0% and specificities greater than 96.0% and an overall accuracy of 93.5% were obtained using Test Set One. The optimal subset consisted of 15 features and they are presented here in the decreasing order of their information content: kurtosis, mean, grey-level entropy, contrast, angular second moment, variance, grey-level nonuniformity, geometric fractal dimension, inverse difference moment, short-run emphasis, stochastic fractal dimension variance, stochastic fractal dimension skewness, stochastic fractal dimension entropy, stochastic fractal dimension kurtosis, and correlation. These sensitivities, specificities, and accuracy demonstrated that the AMFM was a successful tissue characterization method for discriminating between different lung patterns.
|
CT Data Analysis
Regional analysis in entire CT slices (Test Set Two). As mentioned previously, Test Set Two comprised 12 slices. The intraobserver agreement between the first two reading sessions ranged from 46.2 to 77.1%. Four of the five observers (who had an intraobserver agreement of > 50.0%) were retained for the third reading session.
The third reading results for the computer and the four observers are shown in Figure 3. Those blocks that were marked indeterminate either by the computer or by any observer are shown in white. Note that the AMFM results are displayed with a 90% confidence limit (allowing for an indeterminate outcome) and with the AMFM restricted to choosing one of the six tissue patterns (no confidence limits). From the six HRCT slices, 1,839 regions were analyzed after removing the indeterminate samples. The average interobserver agreement was 53.9 ± 6.2%. The average agreement between the computer and the observers was 51.7 ± 2.9%. The average kappa statistic for the interobserver agreement was 0.42 ± 0.08. The average kappa statistic for the agreement between the computer and the observers was 0.39 ± 0.03. Among these 1,839 samples, there were 1,175 "gold standard" samples. The agreement between the "gold standard" and the computer was 69.9%. The kappa statistic of agreement between the "gold standard" and the computer was found to be 0.62.
|
In comparing with the computer result obtained after setting a 90.0% confidence limit, 1,509 samples remained after excluding the indeterminate samples. The average interobserver agreement was 56.2 ± 7.1%. The average agreement between the computer and four observers was 55.4 ± 3.9%. The kappa statistic for the agreement between observers was 0.45 ± 0.08. The kappa statistic for the agreement between the computer and the observers was 0.44 ± 0.04. Among the 1,509 samples, there were 1,002 "gold standard" samples. The agreement between the "gold standard" and the computer was 75.0%. The kappa statistic of agreement between the "gold standard" and computer was found to be 0.68.
In the first observer reading session, the average interobserver agreement for the four observers was 48.8 ± 9.1% and the average computer-observer agreement was 44.4 ± 8.7%. In the second observer reading session, the average interobserver agreement for the same four observers was 52.2 ± 10.0% and the average computer-observer agreement was 47.3 ± 9.0%. These statistics were computed on the same six slices used for the third reading session (with the computer output generated with no confidence limits). Thus, with a combination of observer training and primary patient diagnosis, observer agreement with the computer improved from the first session to the third session. The AMFM-based computer reading remained 100.0% consistent over the three sessions. This was validated by comparing the outputs generated by the computer by repeatedly analyzing the same test sample. Examples of the three readings of a HRCT slice of a subject with an emphysema, given by two observers, compared with the computer reading, are given in Figure 4.
|
| |
DISCUSSION |
|---|
|
TOP
ABSTRACT
INTRODUCTION
METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
This study is the first report of an automated comprehensive method for the regional evaluation of the CT scans of the lung parenchyma. The AMFM provides for the concurrent evaluation of complex lung patterns. In this study, the AMFM was applied for the detection of common lung patterns associated with interstitial lung diseases, but the method can be trained to recognize any lung pattern. The percent area of the CT slice presenting with each parenchymal lung pattern can be readily assessed using the AMFM. Such assessments can be utilized to follow the course of disease or to evaluate outcomes in response to therapy.
The AMFM performed as well as a group of trained observers who were given the primary diagnosis. In addition, the AMFM is 100% reproducible. It is of interest to note that the average agreement between the computer and the observers in the third session increased as compared with the average agreement between the computer and the observers in the first two sessions, respectively. This was most likely a combination of added training acquired by these observers in the first two readings, and the impact of the human observers receiving additional non-image-based data regarding patient diagnosis. This phenomenon has been previously well documented and is also seen clearly in Figure 4. Observer 1 was consistent in reading the emphysematous CT slice in all three sessions, but Observer 2 dramatically changed the reading in the third session after being informed of the diagnosis. The AMFM has demonstrated to be especially successful for the detection of normal, emphysemalike, ground glass, and bronchovascular patterns of the lung parenchyma (individual kappa > 0.6). The AMFM is less successful for the evaluation of the honeycombing and the nodular patterns (individual kappa = 0.2). This may be partly related to the analysis being restricted to two rather than to three dimensions. It is also likely that using thinner CT slices would better facilitate the characterization of such patterns.
Although HRCT offers images of the lung with increasingly improved anatomic resolution, visual assessment of such images remains subjective and qualitative. Typically, a correct global diagnosis of parenchymal lung disease can be made 40 to 70% of the time, with two experienced readers agreeing with each other for global diagnosis 76 to 85% of the time (26, 27). Such variation has been confirmed in other studies examining lung pattern type, with an interobserver variation of 81% (kappa of 0.48), and a similar level of intraobserver variation (kappa of 0.37 to 0.78) (28). In this work of Collins and coworkers (28), an objective assessment was generated by manually tracing on a piece of paper the extent of visually recorded disease over each CT slice, cutting the paper to fit the outline, and then weighing the paper. The weight of the paper served as an indicator of disease extent. Various other visual scoring systems have been suggested by Remy-Jardin and colleagues (29), although in their study the reproducibility of the scoring system was not assessed. In general, these studies reported interobserver agreement values slightly higher than the values obtained in our study. This may be attributed to the fact that the task given to our observers was much more complex and tedious.
It should be noted that to maximize sensitivity and specificity, the application of the AMFM requires standardization of imaging parameters whereby the training set data are gathered similarly to the test set subject data. Methodologic considerations include, but are not limited to, slice thickness, reconstruction kernel, field of view, body posture, and lung inflation state (30). The AMFM can be used with CT images obtained from any scanner so long as the training set is obtained under the same conditions as the test set. Finally, particular investigators for particular studies can select the training set examples.
One of the great challenges in caring for patients with various parenchymal lung diseases is to establish reproducible measures for the presence, and extent, of the disorder (31). Such measures should record global as well as regional changes that are characteristic. These consistent measures can then be applied to the initial evaluation of subjects with suspected parenchymal disease, especially in conjunction to research studies involving the cellular, biochemical, and molecular processes (as might be defined by bronchoalveolar lavage or open lung biopsy) in these complex disorders. The measures can also be reapplied to objectively study the global and regional evolution of the diseases over time, with or without therapy. The AMFM allows such objective measures to be made. The AMFM is an objective, repeatable tool for the quantitative characterization of lung tissue in the presence of mixed pathologies. Such a methodology is critical as we progress towards the need for refined phentotyping of lung disease and the associated need for improved objective outcome criteria.
| |
Footnotes |
|---|
Correspondence and requests for reprints should be addressed to Geoffrey McLennan, Department of Internal Medicine, University of Iowa, 200 Hawkins Drive, Iowa City, IA 52242. E-mail: geoffrey-mclennan{at}uiowa.edu
(Received in original form April 16, 1998 and in revised form January 20, 1999).
Acknowledgments: Supported in part by Contract N01-LM-4-3511 US PHS from the National Library of Medicine and a Career Investigator Award from the American Lung Association.
| |
References |
|---|
|
TOP
ABSTRACT
INTRODUCTION
METHODS
RESULTS
DISCUSSION
REFERENCES
|
|---|
1.
Muller, N. L.,
C. A. Staples,
R. R. Miller, and
R. T. Abboud.
1988.
"Density mask": an objective method to quantitate emphysema using computed tomography.
Chest
94:
782-787
2. Kemerink, G. J., H. H. Kruize, R. J. Lamers, and J. M. van Engelshoven. 1997. CT lung densitometry: dependence of CT number histograms on sample volume and consequences for scan protocol comparability. J. Comput. Assist. Tomogr. 21: 948-954 [Medline].
3. Zagers, H., H. A. Vrooman, N. J. Aarts, J. Stolk, L. J. Schultze, Kool, J. H. Dijkman, A. E. Van Voorthuisen, and J. H. Reiber. 1996. Assessment of the progression of emphysema by quantitative analysis of spirometrically gated computed tomography images. Invest. Radiol. 31: 761-767 [Medline].
4. Gilman, M. J., R. G. Laurens, J. W. Somogyi, and E. G. Honig. 1983. CT attenuation values of lung density in sarcoidosis. J. Comput. Assist. Tomogr. 7: 407-410 [Medline].
5. Rienmuller, R. K., J. Behr, W. A. Kalender, M. Schatzl, I. Altmann, M. Merin, and T. Beinert. 1991. Standardized quantitative high resolution CT in lung diseases. J. Comput. Assist. Tomogr. 15: 742-749 [Medline].
6. Beinert, T., J. Behr, F. Mehnert, P. Kohz, M. Seemann, R. Rienmuller, and M. Reiser. 1995. Spirometrically controlled quantitative CT for assessing diffuse parenchymal lung disease. J. Comput. Assist. Tomogr. 19: 924-931 [Medline].
7.
Uppaluri, R.,
T. Mitsa,
M. Sonka,
E. A. Hoffman, and
G. McLennan.
1997.
Quantification of pulmonary emphysema from lung computed tomography images.
Am. J. Respir. Crit. Care Med.
156:
248-254
8. Uppaluri, R., M. Sonka, E. A. Hoffman, T. Mitsa, C. Dayton, G. W. Hunninghake, and G. McLennan. 1997. Quantitative assessment in interstitial lung disease using the Multiple Feature (MF) method (abstract). Am. J. Respir. Crit. Care Med. 155: A322 .
9.
Uppaluri, R.,
E. A. Hoffman,
M. Sonka,
G. W. Hunninghake, and
G. McLennan.
1998.
Interstitial lung disease: a quantitative study using
the Adaptive Multiple Feature Method.
Am. J. Respir. Crit. Care Med.
159:
519-525
10. American Thoracic Society. 1979. ATS statement: snowbird workshop on standardization of spirometry. Am. Rev. Respir. Dis. 119: 831-838 [Medline].
11. Morris, J. F., A. Koski, and L. C. Johnson. 1971. Spirometric standard for healthy nonsmoking adults. Am. Rev. Respir. Dis. 103: 57-67 [Medline].
12. Goldman, H. I., and M. R. Becklake. 1959. Respiratory function tests: Normal values at median altitudes and the prediction of normal results. Am. Rev. Tuberc. 79: 457-467 . [Medline]
13. Van Ganse, W. F., B. G. Ferris, and J. E. Cotes. 1972. Spirometric standard for healthy nonsmoking adults. Am. Rev. Respir. Dis. 105: 30-41 [Medline].
14. Andrews, H. C. 1972. Introduction to Mathematical Techniques in Pattern Recognition. Wiley, New York.
15. Sonka, M., V. Hvalac, and R. Boyle. 1993. Image Processing, Analysis and Machine Vision. Chapman and Hall, London.
16. Mandelbrot, B. B. 1983. The Fractal Geometry of Nature. W. H. Freeman and Co., New York.
17. Weibel, E. R. 1963. Morphometry of Human Lung. Academic Press, New York.
18. Pentland, A. P. 1984. Fractal-based description of natural scenes. IEEE Trans. Pattern Anal. Mach. Intell. PAMI-6:661-675.
19. Uppaluri, R. 1997. Automated analysis of pulmonary parenchyma from CT images. University of Iowa, Iowa City, IA.
20. Chen, C., J. S. Daponte, and M. D. Fox. 1989. Fractal feature analysis and classification in medical imaging. IEEE Trans. Med. Imaging 8: 133-142 .
21. Ferdeghini, E. M., B. Pinamonti, E. Picano, F. Lattanzi, R. Bussani, G. Slavich, A. Benassi, F. Camerini, L. Landini, and A. L'Abbate. 1991. Quantitative texture analysis in echocardiography: application to diagnosis of myocarditis. J. Clin. Ultrasound 19: 263-270 [Medline].
22. Bister, M. 1990. Computer Analysis of Cardiac MR Images. Vrije Universiteit Brussel, Brussels, Belgium.
23. Meziane, M. A.. 1992. High-resolution computed tomography scanning in the assessment of interstitial lung diseases. J. Thorac. Imaging 7: 13-25 [Medline].
24. Dixon, W. J., Jr., and F. J. Massey, Jr. 1983. Introduction to Statistical Analysis. McGraw Hill, New York.
25. Fleiss, J. H.. 1971. Measuring nominal scale agreement among many raters. Psychol. Bull. 76: 378-382 .
26.
Nishimura, K.,
T. Izumi,
M. Kitaichi,
S. Nagai, and
H. Itoh.
1993.
The diagnostic accuracy of high-resolution computed tomography in diffuse
infiltrative lung diseases.
Chest
104:
1149-1155
27. Padley, S. P., D. M. Hansell, C. D. Flower, and P. Jennings. 1991. Comparative accuracy of high resolution computed tomography and chest radiography in the diagnosis of chronic diffuse infiltrative lung disease. Clin. Radiol. 44: 222-226 [Medline].
28. Collins, C. D., A. U. Wells, D. M. Hansell, R. A. Morgan, J. E. MacSweeney, R. M. D. Bois, and M. B. Ruebens. 1994. Observer variation in pattern type and extent of disease in fibrosing alveolitis on thin section computed tomography and chest radiography. Clin. Radiol. 49: 236-240 [Medline].
29.
Remy-Jardin, M.,
J. Remy,
B. Wallaert,
D. Bataille, and
P. Y. Hatron.
1993.
Pulmonary involvement in progressive systemic sclerosis: sequential evaluation with CT, pulmonary function tests, and bronchoalveolar lavage.
Radiology
188:
499-506
30. Gruden, J. F., and G. McGuinness. 1996. Subjective pitfalls in HRCT Interpretation. Crit. Rev. Diagn. Imaging 37: 349-434 [Medline].
31. Hoffman, E. A., and G. McLennan. 1997. Need for assessment of pulmonary structure/function relationships and clinical outcomes measures: quantitative volumetric CT of the lung. Acad. Radiol. 4: 758-776 [Medline].
| |
APPENDIX |
|---|
|
This article has been cited by other articles:
 |
|
 |
|
  C. S. Rogers, W. M. Abraham, K. A. Brogden, J. F. Engelhardt, J. T. Fisher, P. B. McCray Jr., G. McLennan, D. K. Meyerholz, E. Namati, L. S. Ostedgaard, et al. The porcine lung as a potential model for cystic fibrosis Am J Physiol Lung Cell Mol Physiol, August 1, 2008; 295(2): L240 - L263. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 E. A. Hoffman, B. A. Simon, and G. McLennan State of the Art. A Structural and Functional Assessment of the Lung via Multidetector-Row Computed Tomography: Phenotyping Chronic Obstructive Pulmonary Disease Proceedings of the ATS, August 1, 2006; 3(6): 519 - 532. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 K. G. Kim, J. M. Goo, J. H. Kim, H. J. Lee, B. G. Min, K. T. Bae, and J.-G. Im Computer-aided Diagnosis of Localized Ground-Glass Opacity in the Lung at CT: Initial Experience Radiology, November 1, 2005; 237(2): 657 - 661. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J.D. Newell Jr, J.C. Hogg, and G.L. Snider Report of a workshop: quantitative computed tomography scanning in longitudinal studies of emphysema Eur. Respir. J., May 1, 2004; 23(5): 769 - 775. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 K. R. Flaherty, J. A. Mumford, S. Murray, E. A. Kazerooni, B. H. Gross, T. V. Colby, W. D. Travis, A. Flint, G. B. Toews, J. P. Lynch III, et al. Prognostic Implications of Physiologic and Radiographic Changes in Idiopathic Interstitial Pneumonia Am. J. Respir. Crit. Care Med., September 1, 2003; 168(5): 543 - 548. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
M. S. Brown, J. G. Goldin, R. D. Suh, M. F. McNitt-Gray, J. W. Sayre, and D. R. Aberle Lung Micronodules: Automated Method for Detection at Thin-Section CT--Initial Experience Radiology, January 1, 2003; 226(1): 256 - 262. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. G. Venegas and G. G. Galletti Low-pass filtering, a new method of fractal analysis: application to PET images of pulmonary blood flow J Appl Physiol, April 1, 2000; 88(4): 1365 - 1373. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| Proc. Am. Thorac. Soc. | Am. J. Respir. Cell Mol. Biol. |