HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS		QUICK SEARCH:   [advanced] Author: Keyword(s): Year:  Vol:  Page:

This Article Abstract Figures Only Full Text (PDF) All Versions of this Article: 2283020505v1 228/3/871    most recent Submit a response Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Chabat, F. Articles by Hansell, D. M. Search for Related Content PubMed PubMed Citation Articles by Chabat, F. Articles by Hansell, D. M.

Obstructive Lung Diseases: Texture Classification for Differentiation at CT¹

¹ From the Department of Visual Information Processing (F.C., G.Z.Y.) and Division of Investigative Sciences (D.M.H.), Imperial College of Science, Technology and Medicine, Royal Brompton Hospital, Sydney St, London SW3 6NP, England. Received April 30, 2002; revision requested July 10; final revision received December 19; accepted January 13, 2003. F.C. supported by Imatron, San Francisco, Calif. Address correspondence to D.M.H. (e-mail: d.hansell@rbh.nthames.nhs.uk).

	ABSTRACT

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 
An automated technique for differentiation between a variety of obstructive lung diseases on the basis of textural analysis of thin-section computed tomographic (CT) images is described. From four regions of interest on each image, local texture information was extracted and represented by a 13-dimensional vector that contained statistical moments of the CT attenuation distribution, acquisition-length parameters, and co-occurrence descriptors. A supervised Bayesian classifier was used for texture feature segmentation. The technique was tested with a new cohort of subjects (n = 33, 660 regions of interest) with a similar spectrum of diseases. The proposed technique discriminates well between patterns of obstructive lung disease on the basis of parenchymal texture alone.

© RSNA, 2003

Index terms: Bronchiolitis obliterans, 60.219 • Computed tomography (CT), thin-section, 60.12115 • Computers, diagnostic aid, 60.12115 • Emphysema, 60.7512, 60.7513 • Lung, CT, 60.12115

	INTRODUCTION

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 
Thin-section computed tomography (CT) is an accurate imaging technique for the detection of various obstructive lung diseases, including centrilobular emphysema and constrictive bronchiolitis. Features on thin-section CT images can be subtle, particularly in the early stages of disease, and diagnosis is subject to interobserver variation. The main image characteristic used for the detection of obstructive lung diseases is the presence of areas of abnormally low attenuation in the lung parenchyma, which, in the case of emphysematous destruction of the lung parenchyma, can be detected automatically by means of attenuation masking (1,2). However, areas of decreased attenuation are a feature of other obstructive lung diseases; thus, identification does not always permit a confident diagnosis. To refine the differential diagnosis of obstructive lung disease, it is necessary to take into account the textural appearance of lung parenchyma with abnormally low attenuation. The aim of this study was to describe and test an automated method for the differentiation of centrilobular emphysema, panlobular emphysema, constrictive obliterative bronchiolitis, and normal lung tissue on the basis of texture features.

	Materials and Methods

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 
Subjects
The samples for supervised learning were obtained from a set of CT images obtained in healthy subjects (n = 11) and patients with panlobular emphysema (n = 7), centrilobular emphysema (n = 11), and constrictive obliterative bronchiolitis (n = 15). Images were acquired with a standard thin-section CT protocol (1.5-mm beam collimation, sharp kernel reconstruction algorithm) (model C-150-L; Imatron, San Francisco, Calif). The healthy subjects underwent CT as part of a separate clinical trial, which was approved by the institutional review board, and all subjects gave informed consent. The remaining studies were undertaken for clinical indications and were consecutive. Retrospective use of such data does not require the approval of our institutional board review or patient informed consent.

The visual characteristics for each of the four classes of images are illustrated in Figure 1. Panlobular emphysema results in areas of uniform hypoattenuated lung, as shown in Figure 1a, whereas centrilobular emphysema produces small focal, approximately circular, areas of lung destruction (3) that superficially resemble cysts (ie, air-containing spaces sometimes with a thin definable wall), as shown in a case of severe centrilobular emphysema in Figure 1b. The conspicuity of the pattern is dependent on the severity of the disease. For a mild case, as shown in Figure 1c, the characteristic textural appearance is visible but less obvious. The air trapping and underperfusion caused by constrictive obliterative bronchiolitis results in homogeneously hypoattenuated lung (4), as shown in Figure 1d. For comparison, Figure 1e shows the CT appearance of normal lung parenchyma. Figure 1 also shows enlarged views of areas of lung that typify these diseases. The technique was tested on ROIs from images (obtained with the same CT machine and scanning protocol) of a different group of individuals that comprised healthy subjects (n = 9) (different individuals from the same cohort used for the "learning set") and patients with panlobular emphysema (n = 8), centrilobular emphysema (n = 9), and constrictive obliterative bronchiolitis (n = 7).

View larger version (76K): [in this window] [in a new window] [Download PPT slide]   Figure 1a. Transverse thin-section CT scans of the chest (window level, -500 HU; width, 1,000 HU). On each image, the circle highlights a region of interest (ROI) that is typical of a particular condition. An enlarged view of the ROI is shown on the left side of each image. (a) Panlobular emphysema: Destruction of the lung parenchyma results in areas of homogeneously decreased attenuation. (b) Severe centrilobular emphysema: The CT appearance of the disease superficially resembles that of air cysts. (c) Mild centrilobular emphysema: The features of centrilobular emphysema are visible but less obvious. (d) Constrictive obliterative bronchiolitis results in homogeneously decreased lung attenuation. (e) Normal lung tissue: The mean CT attenuation value of the lung parenchyma is higher than that in cases of obstructive lung disease.

View larger version (76K): [in this window] [in a new window] [Download PPT slide]   Figure 1b. Transverse thin-section CT scans of the chest (window level, -500 HU; width, 1,000 HU). On each image, the circle highlights a region of interest (ROI) that is typical of a particular condition. An enlarged view of the ROI is shown on the left side of each image. (a) Panlobular emphysema: Destruction of the lung parenchyma results in areas of homogeneously decreased attenuation. (b) Severe centrilobular emphysema: The CT appearance of the disease superficially resembles that of air cysts. (c) Mild centrilobular emphysema: The features of centrilobular emphysema are visible but less obvious. (d) Constrictive obliterative bronchiolitis results in homogeneously decreased lung attenuation. (e) Normal lung tissue: The mean CT attenuation value of the lung parenchyma is higher than that in cases of obstructive lung disease.

View larger version (73K): [in this window] [in a new window] [Download PPT slide]   Figure 1c. Transverse thin-section CT scans of the chest (window level, -500 HU; width, 1,000 HU). On each image, the circle highlights a region of interest (ROI) that is typical of a particular condition. An enlarged view of the ROI is shown on the left side of each image. (a) Panlobular emphysema: Destruction of the lung parenchyma results in areas of homogeneously decreased attenuation. (b) Severe centrilobular emphysema: The CT appearance of the disease superficially resembles that of air cysts. (c) Mild centrilobular emphysema: The features of centrilobular emphysema are visible but less obvious. (d) Constrictive obliterative bronchiolitis results in homogeneously decreased lung attenuation. (e) Normal lung tissue: The mean CT attenuation value of the lung parenchyma is higher than that in cases of obstructive lung disease.

View larger version (71K): [in this window] [in a new window] [Download PPT slide]   Figure 1d. Transverse thin-section CT scans of the chest (window level, -500 HU; width, 1,000 HU). On each image, the circle highlights a region of interest (ROI) that is typical of a particular condition. An enlarged view of the ROI is shown on the left side of each image. (a) Panlobular emphysema: Destruction of the lung parenchyma results in areas of homogeneously decreased attenuation. (b) Severe centrilobular emphysema: The CT appearance of the disease superficially resembles that of air cysts. (c) Mild centrilobular emphysema: The features of centrilobular emphysema are visible but less obvious. (d) Constrictive obliterative bronchiolitis results in homogeneously decreased lung attenuation. (e) Normal lung tissue: The mean CT attenuation value of the lung parenchyma is higher than that in cases of obstructive lung disease.

View larger version (72K): [in this window] [in a new window] [Download PPT slide]   Figure 1e. Transverse thin-section CT scans of the chest (window level, -500 HU; width, 1,000 HU). On each image, the circle highlights a region of interest (ROI) that is typical of a particular condition. An enlarged view of the ROI is shown on the left side of each image. (a) Panlobular emphysema: Destruction of the lung parenchyma results in areas of homogeneously decreased attenuation. (b) Severe centrilobular emphysema: The CT appearance of the disease superficially resembles that of air cysts. (c) Mild centrilobular emphysema: The features of centrilobular emphysema are visible but less obvious. (d) Constrictive obliterative bronchiolitis results in homogeneously decreased lung attenuation. (e) Normal lung tissue: The mean CT attenuation value of the lung parenchyma is higher than that in cases of obstructive lung disease.

For each ROI, a statistical descriptor was derived in the form of an N-dimensional vector v. Each vector v contained the values of n = 13 textural features chosen to describe the CT attenuation characteristics in the ROI. These features included nth-order statistics of the distribution of the CT values: mean, SD, skewness, and kurtosis.

Other features that describe the spatial dependence of gray-scale distributions were derived from the set of co-occurrence matrices computed at each ROI (6). Each ROI was approximated on P = 16-level gray scale. A set of 20 matrices {C_d,l}^1d4,1l5 was then derived, which represent the co-occurrence of gray levels along four directions d, at distances ranging from l = 1 pixel to l = 5 pixels. They characterize the spatial relationships of gray levels in textural patterns in a way that is invariant with monotonic gray-level transformations. To reduce the dimensionality of the feature vector v, five scalar measurements were extracted from each co-occurrence matrix: energy, entropy, maximum, contrast, and homogeneity. From a matrix C = (C_i,j)_1i4,1j4, the measurements are computed as follows (6):

where = 1 and ß = 1; and

where i and j in these equations represent the row and column indexes of the co-occurrence matrices. The mean value of these five parameters, over the 20 co-occurrence matrices, was incorporated in the feature vector v that described an ROI.

Since co-occurrence matrices do not capture the shape aspects of the gray-level primitives, acquisition-length parameters were also computed at each ROI to provide another set of features to be included in v. A primitive is a maximum contiguous set of constant gray-level pixels located in a line (7). For each ROI approximated on a P = 16-level gray scale, the number B(a, r) represents the total number of primitives of the length r and gray level a. Textural description can be summarized by deriving short primitive emphasis (spe), long primitive emphasis (lpe), gray-level uniformity (glu), and primitive length uniformity (plu). The short primitive emphasis value measures the predominance of short primitives in a textural pattern. A high short primitive emphasis value denotes a pattern that consists mainly of short lines with a constant gray level. The short primitive emphasis value can be computed with the following formula, which gives higher weight to the shorter primitives:

where r_max denotes the maximum primitive length in the ROI, and B_tot is the total number of acquisitions. Thus,

Similarly, long primitive emphasis measures the predominance of long primitives in a textural pattern, and its value is given by

Gray-level uniformity measures gray-level dispersion of the primitives. A high gray-level uniformity value denotes a textural pattern where primitives belong to a small number of gray levels, as in a checkerboard, for instance. Gray-level uniformity is computed as follows:

Primitive length uniformity measures the similarity in length of the primitives. A high primitive length uniformity value denotes a pattern where all primitives have a roughly equal length. Primitive length uniformity is computed as follows:

In summary, each ROI was characterized by a 13-dimensional vector v, which contained the values given in Table 1. The validity of these descriptors for capturing textural information has been demonstrated in a wide range of computer vision applications (8). The use of fractal dimensions has also been suggested (9), but we did not use the method because of its computational cost.

0k3

where µ_k represents the mean of vectors in class k;

and K_k is the covariance matrix of vectors in class k;

Each distribution is unimodal, exhibiting one peak at µ_k, and decays in all directions, according to the quadratic form

The mean vectors µ_k and the symmetric positive definite covariance matrices K_k completely define the distributions _k and could be derived from the sets of samples {Sk}_0k3 with Equations (12) and (13). The inverse and the determinant of each matrix K_k were obtained through LU decomposition (13). Following the assumption that the distribution of vectors in each class is normal, a Bayes classifier could be implemented, with the k class-specific probabilities being equal to the distribution _k(v). Thus, for all k,

It is important to stress that the a priori probability for each class could not be computed from the sets of samples {Sk}_0k3. In this article, we assumed that all a priori probabilities were equal. Thus, for all k,

Once the vectors µ_k and the matrices K_k had been defined with the training sets {Sk}_0k3, each new testing samples v could be classified as belonging to class k₀, with the minimum error rate decision from the Bayes decision rule (14). Thus, we decide for k₀ if

where

In the case of equal a priori probabilities for all classes, as shown in Equation (16), Equation (18) can be simplified as

The a posteriori probability prob(k₀|v) gave a measure of the confidence of the classifier in assigning the label k₀ to a sample v. To model the lack of confidence of the classifier, samples could be rejected and remain unclassified if their a posteriori probability was less than a predefined threshold p_threshold. Thus, we reject v if

The value of p_threshold could be selected to maximize sensitivity and specificity of the classifier used as a diagnostic test, while retaining a sufficient number of reclassified samples. Standard statistical tests of diagnostic accuracy (sensitivity, specificity, and predictive value) were used to evaluate the technique. Since it was not feasible to obtain histopathologic confirmation of the dominant disease process in each of the ROIs, the diagnosis assigned by the experienced radiologist for each ROI (not for the "whole patient") was regarded as a standard of reference for the purpose of this study.

	Results

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 
Without a set probability threshold p_threshold, all 660 ROIs of the testing set were classified, irrespective of the confidence of the classifier. A total of 398 ROIs were then labeled correctly, giving an overall sensitivity of 60.3% and overall specificity of 86.7%. With a set value of p_threshold = p₁ such that half the samples (330 ROIs) were confidently reclassified and half the samples were rejected, 243 ROIs were correctly classified, which gave overall sensitivity of 73.6% and overall specificity of 91.2%. Sensitivity, specificity, and positive predictive value for each of the four classes are given in Table 2. With a threshold p_threshold = p₂ such that a fourth of samples are reclassified, 132 of 165 ROIs were labeled correctly, giving overall sensitivity and specificity of 80.0% and 93.3%. Figure 2 shows the sensitivity and specificity for several values of p_threshold. Globally, the setting of a higher value of p_threshold improves the performance of the algorithm but leaves a larger number of samples unclassified.

View larger version (19K): [in this window] [in a new window] [Download PPT slide]   Figure 2. Line graph shows relationship between the threshold pthreshold on a posteriori probabilities and the overall sensitivity and specificity of the classifier. As pthreshold increases, and only samples with increasing confidence are classified, the overall sensitivity and specificity of the technique improve. However, a larger proportion of samples are then rejected and remain unclassified. A trade-off can be obtained with a value of pthreshold such that half the samples are reclassified confidently, for instance. Corresponding overall sensitivity and overall specificity are 73.6% and 91.2%.

As an illustration, Figure 3 shows the way the proposed Bayesian classifier can be applied not only to user-selected ROIs but also to the whole parenchyma for textural segmentation (5). Normal lung tissue is shown in Figure 3a (left), and the result of classification is shown in Figure 3a (right). All the samples classified confidently, with a posteriori probabilities greater than p₁, were labeled as normal. Signs of centrilobular emphysema are shown in Figure 3b (left), and the segmentation in Figure 3b (right) shows that most samples were classified as centrilobular emphysema, with some areas of homogeneous lung labeled as panlobular emphysema. Figure 3c (left) shows constrictive obliterative bronchiolitis, and its segmentation is given in Figure 3c (right). An example of panlobular emphysema is shown in Figure 3d (left), and the corresponding segmentation is shown in Figure 3d (right).

View larger version (63K): [in this window] [in a new window] [Download PPT slide]   Figure 3a. Application of the classifier to segmentation of lung parenchyma. At pixels where the confidence of the classifier is low (smaller than the p1 threshold), no label is assigned. (a) Left: Transverse CT scan obtained in a healthy subject (window level, -800 HU; width, 1,000 HU). Right: Transverse CT scan obtained with automated classification. Samples classified confidently are labeled as normal. (b) Left: Transverse CT scan obtained in a patient with signs of centrilobular emphysema. Right: In transverse CT scan, most of the classified samples are labeled as centrilobular emphysema. Some areas of lung with homogeneously decreased attenuation are classified as constrictive obliterative bronchiolitis. (c) Left: Transverse CT scan in a patient with constrictive obliterative bronchiolitis. Right: In transverse CT scan, most of the parenchyma is labeled as constrictive obliterative bronchiolitis. Some areas are classified as normal. Areas of lung with increased attenuation adjacent to areas of decreased attenuation make the classifier identify the texture of centrilobular emphysema in the vicinity of the major bronchi. (d) Left: Transverse CT scan obtained in a patient with panlobular emphysema. Right: In transverse CT scan, most of the parenchyma is labeled as panlobular emphysema, with some areas classified as constrictive obliterative bronchiolitis.

View larger version (71K): [in this window] [in a new window] [Download PPT slide]   Figure 3b. Application of the classifier to segmentation of lung parenchyma. At pixels where the confidence of the classifier is low (smaller than the p1 threshold), no label is assigned. (a) Left: Transverse CT scan obtained in a healthy subject (window level, -800 HU; width, 1,000 HU). Right: Transverse CT scan obtained with automated classification. Samples classified confidently are labeled as normal. (b) Left: Transverse CT scan obtained in a patient with signs of centrilobular emphysema. Right: In transverse CT scan, most of the classified samples are labeled as centrilobular emphysema. Some areas of lung with homogeneously decreased attenuation are classified as constrictive obliterative bronchiolitis. (c) Left: Transverse CT scan in a patient with constrictive obliterative bronchiolitis. Right: In transverse CT scan, most of the parenchyma is labeled as constrictive obliterative bronchiolitis. Some areas are classified as normal. Areas of lung with increased attenuation adjacent to areas of decreased attenuation make the classifier identify the texture of centrilobular emphysema in the vicinity of the major bronchi. (d) Left: Transverse CT scan obtained in a patient with panlobular emphysema. Right: In transverse CT scan, most of the parenchyma is labeled as panlobular emphysema, with some areas classified as constrictive obliterative bronchiolitis.

View larger version (57K): [in this window] [in a new window] [Download PPT slide]   Figure 3c. Application of the classifier to segmentation of lung parenchyma. At pixels where the confidence of the classifier is low (smaller than the p1 threshold), no label is assigned. (a) Left: Transverse CT scan obtained in a healthy subject (window level, -800 HU; width, 1,000 HU). Right: Transverse CT scan obtained with automated classification. Samples classified confidently are labeled as normal. (b) Left: Transverse CT scan obtained in a patient with signs of centrilobular emphysema. Right: In transverse CT scan, most of the classified samples are labeled as centrilobular emphysema. Some areas of lung with homogeneously decreased attenuation are classified as constrictive obliterative bronchiolitis. (c) Left: Transverse CT scan in a patient with constrictive obliterative bronchiolitis. Right: In transverse CT scan, most of the parenchyma is labeled as constrictive obliterative bronchiolitis. Some areas are classified as normal. Areas of lung with increased attenuation adjacent to areas of decreased attenuation make the classifier identify the texture of centrilobular emphysema in the vicinity of the major bronchi. (d) Left: Transverse CT scan obtained in a patient with panlobular emphysema. Right: In transverse CT scan, most of the parenchyma is labeled as panlobular emphysema, with some areas classified as constrictive obliterative bronchiolitis.

View larger version (54K): [in this window] [in a new window] [Download PPT slide]   Figure 3d. Application of the classifier to segmentation of lung parenchyma. At pixels where the confidence of the classifier is low (smaller than the p1 threshold), no label is assigned. (a) Left: Transverse CT scan obtained in a healthy subject (window level, -800 HU; width, 1,000 HU). Right: Transverse CT scan obtained with automated classification. Samples classified confidently are labeled as normal. (b) Left: Transverse CT scan obtained in a patient with signs of centrilobular emphysema. Right: In transverse CT scan, most of the classified samples are labeled as centrilobular emphysema. Some areas of lung with homogeneously decreased attenuation are classified as constrictive obliterative bronchiolitis. (c) Left: Transverse CT scan in a patient with constrictive obliterative bronchiolitis. Right: In transverse CT scan, most of the parenchyma is labeled as constrictive obliterative bronchiolitis. Some areas are classified as normal. Areas of lung with increased attenuation adjacent to areas of decreased attenuation make the classifier identify the texture of centrilobular emphysema in the vicinity of the major bronchi. (d) Left: Transverse CT scan obtained in a patient with panlobular emphysema. Right: In transverse CT scan, most of the parenchyma is labeled as panlobular emphysema, with some areas classified as constrictive obliterative bronchiolitis.

	Discussion

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

In the present study, training and testing of the classifier were based on the visual classification made by an experienced radiologist. The reliability of subjective visual assessment for this classification is open to question. Other methods, however, such as pulmonary function tests, do not reliably distinguish between the various pathologic causes of obstructive lung diseases. Biopsy is not performed routinely in patients with obstructive pulmonary disease, either because patients have such severe disease that surgery is precluded or because they have mild disease that does not warrant an invasive procedure. Furthermore, there are regional differences in the predominant disease; in patients with emphysema, areas of centrilobular and panlobular disease may coexist in different parts of the lungs. Despite its limitations, CT evaluation by an experienced radiologist is a reasonable standard of reference. Within this context, the proposed classifier achieved high sensitivity and specificity. Also, results of evaluation of the classifier with the training set demonstrated that the proposed technique allows correct encapsulation and discrimination of the textural information contained in the training set. Irrespective of the diagnostic reliability of subjective assessment, the method used in the current study classifies ROIs in a way that is consistent with that of the experienced observer who provided the training data; nevertheless, the ability of the algorithm to successfully match textures chosen by one observer falls short of fully competent diagnostic performance. The assumption that the distribution of feature vectors in each class is normal can also be questioned. However, performance comparison with a minimum-distance classifier, which makes no assumption about the model of the probability densities, showed that the proposed method is less sensitive to sampling noise.

In this study, the largest number of cases of misclassification resulted from confusion between cases of panlobular emphysema and constrictive obliterative bronchiolitis. This is not surprising, given the similar visual appearances at CT of the lung parenchyma for these two conditions. Nevertheless, the general discriminating value of the proposed classifier was demonstrated, with high overall sensitivity and specificity.

The size of the ROIs (radius = 22 pixels) was established in collaboration with a radiologist, who assigned ROIs that contained CT characteristics of the respective diseases. An example of a textural classifier applied to CT images relies on smaller ROIs (5 x 5 pixels [19]). In that situation, however, the purpose of texture-based image analysis was to perform full anatomic segmentation, and the size of the ROIs needed to match the typical size of anatomic structures such as major pulmonary vessels or bronchial lumina. Anatomic segmentation had already been performed on the images analyzed by our classifier, and the ROIs encompassed only the lung parenchyma.

The thresholding of a posteriori probabilities allows tuning of the sensitivity and specificity of the classifier according to the relative cost of false-positive and false-negative findings. A probabilistic model is well suited for integration in a fully automated decision-support system. Automatic selection of ROIs necessarily produces samples that belong to none of the four categories that the classifier was trained for, and it is then important that the low confidence of the classifier can be measured. Bayesian probabilities offer a good paradigm for measuring such uncertainties.

We have demonstrated that textural distinction between several diseases that cause decreased attenuation of the lung parenchyma is feasible. A description of the spatial distribution of CT values with statistical moments, co-occurrence matrices, and acquisition-length parameters provides enough information for a pattern classifier. With the assumption that the distribution of feature vectors for each class is normal, the Bayesian decision rule allows implementation of a classifier with high sensitivity and specificity. On the basis of a suitable threshold on the a posteriori probabilities, classification of samples characteristic of the CT appearance of three obstructive lung diseases and normal lung tissue can be achieved with sensitivity and specificity of 73.6% and 91.2%. The accuracy of the method is good, which suggests its value and that is should be included as one of the main CT feature extractors for the automated detection of obstructive lung diseases.

	FOOTNOTES

 
Abbreviation: ROI = region of interest

Author contributions: Guarantor of integrity of entire study, G.Z.Y.; study concepts and design, F.C., G.Z.Y., D.M.H.; literature research, F.C.; clinical studies, D.M.H., F.C.; experimental studies, G.Z.Y., F.C.; data acquisition and analysis/interpretation, F.C.; statistical analysis, F.C.; manuscript preparation, F.C.; manuscript definition of intellectual content and editing, F.C., D.M.H., G.Z.Y.; manuscript revision/review, D.M.H., G.Z.Y.; manuscript final version approval, D.M.H.

	REFERENCES

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion REFERENCES

 

1. Kinsella M, Müller NL, Abboud RT, Morrison NJ, DyBuncio A. Quantitation of emphysema by computed tomography using a "density mask" program and correlation with pulmonary function tests. Chest 1990; 97:315-321.[Abstract/Free Full Text]
2. Sakai N, Mishima M, Nishimura K, Itoh H, Kuno K. An automated method to assess the distribution of low attenuation areas on chest CT scans in chronic pulmonary emphysema patients. Chest 1994; 106:1319-1325.[Abstract/Free Full Text]
3. Stern EJ, Frank MS. CT of the lung in patients with pulmonary emphysema: diagnosis, quantification, and correlation with pathologic and physiologic findings. AJR Am J Roentgenol 1994; 162:791-798.[Abstract/Free Full Text]
4. Hansell DM, Rubens M, Padley SPG, et al. Obliterative bronchiolitis: individual CT signs of small airways disease and functional correlation. Radiology 1997; 203:721-726.[Abstract/Free Full Text]
5. Chabat F, Hansell DM, Yang GZ. Gradient correction and classification of CT lung images for the automated quantification of mosaic attenuation pattern. J Comput Assist Tomogr 2000; 24:437-447.[CrossRef][Medline]
6. Haralick RM. Statistical and structural approaches to texture. Proc IEEE 1979; 67:786-804.
7. Sonka M, Hlavac V, Boyle R. Image processing, analysis and machine vision London, England: Chapman & Hall, 1993.
8. Haralick RM, Shapiro LG. Computer and robot vision Reading, Mass: Addison-Wesley, 1992.
9. Shimizu K, Johkoh T, Ikezoe J, et al. Fractal analysis for classification of ground-glass opacities on high-resolution CT: an in vitro study. J Comput Assist Tomogr 1997; 21:955-961.[CrossRef][Medline]
10. Kittler J, Hatef M, Duin RPW, et al. On combining classifiers. IEEE Trans Pattern Analysis Machine Intelligence 1998; 20:226-239.[CrossRef]
11. Srivastava MS, Carter EM. An introduction to applied multivariate statistics New York, NY: North Holland, 1983; 25-37.
12. Schurmann J. Pattern classification: a unified view of statistical and neural approaches New York, NY: Wiley, 1996.
13. Gourlay AR. Computational methods for matrix eigenproblems London, England: Wiley, 1973.
14. Schalkoff R. Pattern recognition: statistical, structural and neural approaches New York, NY: Wiley, 1992.
15. Buteau M, Makram-Ebeid S. A computer vision approach to emphysema detection and quantification In: Proceedings of the 18th International Conference of the IEEE Engineering in Medicine and Biology Society. Amsterdam, the Netherlands: IEEE Engineering in Medicine and Biology Society, 1996; 1089-1090.
16. Bergin CJ, Müller NL, Miller RR. CT in the qualitative assessment of emphysema. J Thorac Imaging 1986; 1:94-103.[Medline]
17. Uppaluri R, Mitsa T, Sonka M, et al. Quantification of pulmonary emphysema from lung computed tomography images. Am J Respir Crit Care Med 1997; 156:248-254.[Abstract/Free Full Text]
18. Katsuragawa S, Doi K, MacMahon H, et al. Classification of normal and abnormal lungs with interstitial diseases by rule-based method and artificial neural networks. J Digit Imaging 1997; 10:108-114.[Medline]
19. Heitmann KR, Kauczor HU, Mildenberger P, et al. Automatic detection of ground glass opacities on lung HRCT using multiple neural networks. Eur Radiol 1997; 7:1464-1472.
20. Delorme S, Keller-Reichenbecher MA, Zuna I, et al. Quantitative assessment of high-resolution computed tomography findings by computer-assisted texture-based image analysis. Invest Radiol 1997; 32:566-574.[CrossRef][Medline]

This article has been cited by other articles:

				A. Madani, A. Van Muylem, V. de Maertelaer, J. Zanen, and P. A. Gevenois Pulmonary Emphysema: Size Distribution of Emphysematous Spaces on Multidetector CT Images--Comparison with Macroscopic and Microscopic Morphometry Radiology, September 1, 2008; 248(3): 1036 - 1041. [Abstract] [Full Text] [PDF]

				A. C. Best, J. Meng, A. M. Lynch, C. M. Bozic, D. Miller, G. K. Grunwald, and D. A. Lynch Idiopathic Pulmonary Fibrosis: Physiologic Tests, Quantitative CT Indexes, and CT Visual Scores as Predictors of Mortality Radiology, March 1, 2008; 246(3): 935 - 940. [Abstract] [Full Text] [PDF]

				S. I. Rennard Chronic Obstructive Pulmonary Disease: Linking Outcomes and Pathobiology of Disease Modification Proceedings of the ATS, May 1, 2006; 3(3): 276 - 280. [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]

				C. Beigelman-Aubry, C. Hill, A. Guibal, J. Savatovsky, and P. A. Grenier Multi-Detector Row CT and Postprocessing Techniques in the Assessment of Diffuse Lung Disease RadioGraphics, November 1, 2005; 25(6): 1639 - 1652. [Abstract] [Full Text] [PDF]

This Article Abstract Figures Only Full Text (PDF) All Versions of this Article: 2283020505v1 228/3/871    most recent Submit a response Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Chabat, F. Articles by Hansell, D. M. Search for Related Content PubMed PubMed Citation Articles by Chabat, F. Articles by Hansell, D. M.