INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Accurate and reliable methods for objective quantitative assessment of alterations in the lung are necessary for diagnosis and treatment of pulmonary diseases. Although chest radiographs have played an important role in determining the presence and extent of lung disease, it has been shown that radiograph-based examination cannot reliably distinguish between pulmonary emphysema and other causes of overinflation of the lung, except when bullae are present (1). It has also been shown that subjects with normal chest radiographs may have underlying pathologic evidence of interstitial lung diseases (2, 3). It is now well established that computed tomography (CT) is superior to chest radiographs in characterization of pulmonary diseases (4, 5). CT has the ability to display anatomy free from superimposing structures and provides greater visual clarity.

Previous studies utilizing CT to assess interstitial lung disease can be categorized as visual assessment studies and computer-based analyses. In the visual assessment studies (6), one or more observers were asked to assign a numerical grade to the severity of disease. These grades were then either correlated with pathology scores, with bronchoalveolar lavage, or with pulmonary function test (PFT) results. A wide range of interobserver and intraobserver correlations (r values) were reported, ranging from 0.33 to 0.96 for interobserver and between 0.37 and 0.96 for intraobserver agreements, indicating substantial variation in visual assessment scores. Correct global diagnosis of parenchymal lung disease can be made 40 to 70% of the time, with experienced observers agreeing with each other for global diagnosis 76 to 85% of the time (14, 15). Discrepancies within and between observers and the lack of quantitative measures make outcomes analyses difficult.

In a move toward digital analysis of CT studies, two main disease evaluation metrics have evolved. One is the mean lung density (MLD) analysis, where it has been shown that MLD is lower in emphysematous subjects than in normal subjects (16), and higher in subjects with idiopathic pulmonary fibrosis (IPF) (17) or sarcoidosis when compared with normal subjects (18).

The second measure is the lowest fifth percentile of the density histogram (HIST) or some other but similar form of histogram analysis (19). It has been shown in emphysematous subjects that the HIST was shifted towards lower densities, and correlated with pathology or PFTs. Similarly, the histograms of the IPF subjects were shown to shift towards higher densities. In the evolution of these previously reported quantitative studies the CT slice thickness has progressively decreased from 10 mm towards thinner sections of 3 mm or less (HRCT). Various scanning protocols have been used, especially experimenting with the lung at different lung volumes, in an effort to enhance changes in lung density that are ascribed to a disease process. Finally, there have been increasing attempts to take advantage of the large quantity of digital data that is collected by the CT scanning process (24), much of which is not appreciated in the hard copy image.

These approaches of quantifying digital data are objective. However, very few of these studies used normal control subjects. Further, these studies generally examined only a single parameter such as mean lung density, and did so in a single disease process. There are disadvantages in using only lung density information. The measurement of attenuation is highly dependent upon every slight change in lung volume and can vary with beam-hardening effects, scatter, and drifts in scanner calibration. Also the lowest fifth percentile of the histogram is contributed to by both an increase in lower densities (possible emphysema) as well as by the distribution of higher densities (possible fibrosis). These values are therefore difficult to interpret in the presence of mixed disease.

Approaches to providing quantitative assessment of HRCT based on density alone have been both insensitive and nonspecific in the detection of pulmonary parenchymal disease. No method has so far emerged that can be utilized to study multiple forms of parenchymal lung disease. In our previous work (25), we presented a complex texture-based method called the adaptive multiple feature method (AMFM) for studying normal and emphysematous lung parenchyma. In this study, we demonstrated the utility of the AMFM approach by objectively characterizing and discriminating between multiple parenchymal pathologies.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Subjects

Twenty normal subjects and 13 with emphysema, 19 with IPF, and 20 with sarcoidosis were used in this study. The normal subjects were volunteers who had no history of pulmonary disease, were life-long nonsmokers, had normal lung function, and were taking no medication. The subjects with emphysema were chosen from among the subjects undergoing lung volume reduction surgery for emphysema. These subjects all had established severe emphysema, and the HRCT scans were performed as part of their preoperative evaluation. The subjects with IPF and sarcoidosis were selected to have a range of lung function abnormalities from the Specialized Center of Research (SCOR) program in interstitial and occupational lung disease. All of those patients had well-reviewed clinical, pulmonary function, radiologic, and pathologic features of the disease. Four slices from each multislice scan were included in the analysis. By an arbitrary convention, two of these slices were from the region at the carina and the remaining two were from the region halfway between the carina and the base of the lung.

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 (DL_CO) 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 (26). The predicted normal values used were those of Morris and coworkers (27) for spirometry, Goldman and Becklake (28) for lung volumes, and Van Ganse and colleagues (29) for diffusing capacity.

Image Data Acquisition

HRCT 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 subjects with emphysema were acquired in the supine position. A standardized protocol was used assuring uniformity of slice thickness across all subjects. The field of view was 300 to 400 mm and collimation was 3 mm. The images were reconstructed to 512 × 512 pixels. HRCT reconstruction kernel was used. The grey level resolution of the images was 11 bit.

Image Data Analysis

The previously developed AMFM (25) for examining the lung parenchyma from HRCT scans was used. The AMFM is a texture-based method that combines statistical texture measures and a fractal measure. In total, 17 measures of texture were used. Included in these measures were the grey level distribution measures, run-length measures, co-occurrence matrix measures, and a geometric fractal dimension (GFD). The grey level distribution measures were the mean (MEAN), variance (VAR), skewness (SKEW), kurtosis (KURT), and grey level entropy (GREYENT). The run length measures were the short-run emphasis (SRE), long-run emphasis (LRE), grey level nonuniformity (GLN), run length nonuniformity (RLN), and run percentage (RP). The co-occurrence matrix measures were the angular second moment (ASM), correlation (CORR), contrast (CON), entropy (ENT), inertia (INER), and inverse difference moment (IDM). A feature selection program based on the divergence measure was used to select an optimal subset of features to best discriminate the tissues under consideration. A database of HRCT slices of known classification, called the training set, was used to train a Bayesian classifier. A test set of HRCT slices, distinct from the training set, was used to compute the accuracy of the method in tissue characterization. The training and the test sets are detailed below.

HRCT images obtained from normal subjects and those with emphysema, IPF, and sarcoidosis were compared in a global analysis. This analysis was performed using the whole single image slice of the lung field as the region of interest (ROI) to extract the texture features. Two subject groups, three subject groups, and all four of the subject groups were compared. In this global analysis, the performance of the AMFM was compared against the performance of the MLD and HIST methods. Both of these methods were applied to exactly the same data sets as the AMFM, using the criteria for these methods that have been previously published.

Regional analysis was then performed allowing for comparisons of normal and diseased lungs using the AMFM. Each lung in the HRCT slice was divided, by arbitrary convention, into six equal regions, anterior to posterior. The regions were initially reviewed by a trained pulmonologist experienced in HRCT scan assessment of parenchymal lung disease, and those CT regions that had obvious movement artifact, or significant fissure lines, were removed from the analysis. Normal areas from diseased subjects were also not included in the analysis. Emphysema, IPF, and sarcoidosis regions were individually compared with normal regions. In each of the six regions, the computer classifier was trained using alternate regions from normal and diseased lungs. The classifier was then asked to identify the remaining regions from the test set as being normal or diseased. This process was repeated for all six regions.

Statistical Analysis

Regression analysis was performed to determine the correlation between the PFT values of each subject group and the AMFM, MLD, and HIST methods. Each subject group was plotted separately. The MLD and HIST methods were analyzed against each PFT value by using a single texture feature. The AMFM was analyzed against each PFT value by multiple regression analysis using the optimal features chosen for the four group comparison. The feature values for these analyses were obtained by averaging the feature values computed by the AMFM, MLD, and HIST methods in the global analysis over all four slices. Regression analysis was performed using Stat View 4.1 (Abacus Concepts, Inc., Berkeley, CA). Significance of the correlation was assessed using ANOVA using the same package.

In the global analysis, the sensitivities, specificities (30), and overall accuracies for each subject group were calculated. The sensitivity of a subject group was defined as the percentage of samples of that group correctly classified as such. The specificity of a subject group was defined as the percentage of samples not belonging to a group correctly classified as such. The overall accuracy was defined as the percentage of correctly classified samples of all subject groups under consideration. For the four group comparison, the positive predictive and negative predictive values were also computed. McNemar's ² test (31) was used to test the significance of the differences in performance of the AMFM as compared with the MLD and HIST methods.

In the regional analysis, the sensitivity and specificity of each subject group, and the overall accuracy of all subject groups combined, were computed for each of the six regions.

In comparisons involving only two subject groups at a time, it should be noted that the sensitivity of one group is the specificity of the other group and vice versa. Hence, in such discrimination tasks, sensitivity and specificity with respect to only one subject group were reported.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Pulmonary Function Tests

The pulmonary function test results for each group are shown in Table 1. There was some significant correlations (p < 0.05) between the lung function test results and the AMFM, MLD, and HIST parameters.

With the AMFM. TLC actual correlated in the normal and sarcoidosis groups. TLC% predicted correlated in the IPF and sarcoidosis groups. FEV₁% predicted correlated in the IPF group. Pa_O2 actual correlated in the emphysema and IPF groups. DL_CO actual and DL_CO% predicted correlated in the IPF and sarcoidosis groups. RV actual correlated in the normal, IPF, and sarcoidosis groups.

With MLD. TLC actual correlated in the sarcoidosis group. TLC% predicted and FEV₁ actual correlated in the IPF and sarcoidosis groups. FEV₁% predicted correlated in the IPF group. FEV₁/FVC and Pa_O2 actual correlated in the emphysema and normal groups, respectively. DL_CO actual and DL_CO% predicted correlated in the IPF and sarcoidosis groups. RV actual correlated in the sarcoidosis group.

With HIST. TLC actual correlated in the sarcoidosis group.

Global Analysis

In the global analysis, the whole lung slice was used as the region of interest. Two-group, three-group, and finally, four-group comparisons were performed in the four subject groups: 80 normal, 52 emphysema, 76 IPF, and 80 sarcoidosis slices were available, among which half the samples from each group were used for training the classifier and the remaining were used for testing purposes.

In the two-group comparisons (Table 2), the AMFM successfully discriminated emphysema versus normal with an overall accuracy of 100% compared with an accuracy of 91% by the MLD and 99% by the HIST method. IPF versus normal were discriminated with an overall accuracy of 99% by the AMFM compared with overall accuracies of 87 and 71% using the MLD and HIST methods, respectively. In sarcoidosis versus normal groups, the AMFM performed the classification with an overall accuracy of 91%, whereas the overall accuracies using the MLD and HIST methods were 73 and 55%, respectively. In the emphysema versus IPF, emphysema versus sarcoidosis, and IPF versus sarcoidosis groups, the AMFM either was better than the MLD and HIST methods or was comparable. In the emphysema-normal discrimination, the AMFM performed significantly better (p > 0.01) than the MLD method and in the IPF versus normal and sarcoidosis versus normal groups, the AMFM performed significantly better (p > 0.01) than both the MLD and HIST methods. The individual group sensitivities, specificities, as well as the optimal sets of features used by the AMFM for these comparisons are tabulated in Table 2.

In the three-group comparisons (Table 3), the AMFM again performed substantially better than the MLD and HIST methods. In the normal versus emphysema versus IPF groups, the AMFM had an overall accuracy of 100% compared with an overall accuracy of 85 and 76% by the MLD and HIST methods. In the normal versus IPF versus sarcoidosis groups, the overall accuracies were 83, 56, and 47% by the AMFM, MLD, and HIST, respectively. The overall accuracies in the normal versus emphysema versus sarcoidosis groups were 94, 74, and 65% using the AMFM, MLD, and HIST methods, respectively. Finally, emphysema versus IPF versus sarcoidosis discrimination overall accuracies were 77, 70, and 73% using the AMFM, MLD, and HIST methods, respectively. In all the three-group comparisons except the emphysema versus IPF versus sarcoidosis group, the AMFM performed significantly better (p > 0.01) than both the MLD and HIST methods. The individual group sensitivities, specificities, as well as the optimal features for these classifications are given in Table 3.

Combining all four-subject groups, the overall accuracy of the AMFM was 81% compared with 60 and 56% by the MLD and HIST methods. Here also, the AMFM performed significantly better (p > 0.01) than the MLD and HIST methods. The sensitivities, specificities, positive predictive, and negative predictive values, and the optimal combinations of features for the four subject group classification chosen by the AMFM are shown in Table 4.

Regional Analysis

The normal regions were compared individually with emphysema, IPF, and sarcoidosis regions, using the AMFM; 10 to 15% of the normal, IPF, and sarcoidosis regions and about 35% of the emphysema regions were removed for reasons specified in the METHODS section, most commonly because of movement artefact, and 106 normal, 104 emphysema, 76 IPF, and 160 sarcoidosis regions were then available for analysis. Among the regions that were retained, half were used for training the classifier and the remaining were used for testing.

For the normal versus emphysema regional analysis, over all six regions, the average sensitivity and specificity for emphysema detection was 87 ± 8 and 99 ± 2%, respectively. The average overall accuracy was 95 ± 2%. The sensitivities, specificities, and the optimal features used for each region are shown in Table 5.

For the normal versus IPF regional analysis, over all six regions, the average sensitivity and specificity for IPF detection were 82 ± 5 and 89 ± 6%, respectively. The average overall accuracy was 86 ± 4%. The sensitivities, specificities, and the optimal features used for each region are shown in Table 6.

For the normal versus sarcoidosis regional analysis, over all six regions, the average sensitivity and specificity for sarcoidosis detection were 64 ± 8 and 90 ± 7%, respectively. The average overall accuracy was 77 ± 6%. The sensitivities, specificities, and the optimal features used for each region are shown in Table 7.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

This step is a further step in our goal of developing a generalizable, automated method that can provide for an objective and sensitive method for quantitating pulmonary disease, based upon CT scan information. In our previous study, we demonstrated the utility of this method in discriminating the texture differences that result from the effects of gravity on normal lung, as well as the texture differences between normal and emphysematous lung. In the present study, we have continued to develop the AMFM method, and have extended observations by including two other common parenchymal lung diseases, IPF and sarcoidosis.

The MLD and HIST features have been the only two computer-based quantitative CT measures so far proposed to identify emphysema, and MLD has been the only feature to quantify IPF and sarcoidosis.

Emphysema has traditionally been characterized as a form of chronic obstructive pulmonary disorder. However, emphysema can often occur without airway obstruction (32). Growing evidence exists that emphysema occurs with pulmonary fibrosis (33). So far, the MLD and HIST methods have been used to study emphysema, IPF, or sarcoidosis only on an individual disease basis. Although the MLD and HIST methods have shown some success in identifying individual diseases, the utility of these methods in discriminating multiple forms of lung disease, in a single comparison, has not been tested. The concurrent classification of complex parenchymal pathology is necessary to allow for the assessment of different lung pathologies in the same patient. We have compared the AMFM with the MLD and HIST methods in various conditions to test the robustness of all three methods, and have applied each method to the same data set with this study, to allow for direct comparisons.

Global analysis was performed using two-subject groups at a time, three-subject groups at a time, and finally, all four- subject groups. In two-subject group discriminations, emphysema-normal subjects were identified by the AMFM, MLD, and HIST methods successfully with greater than 90% overall accuracy. The subjects with emphysema included in this study were severe cases. Hence, it is probable that this aided all three methods in performing the discrimination task. Emphysema versus IPF and emphysema versus sarcoidosis groups were also discriminated by all three methods with greater than 95% overall accuracies. Emphysema is a disease causing predominantly a decrease in lung density, whereas IPF and sarcoidosis are predominantly associated with an increase in density. Thus, these discriminations are easier to perform. However, the power of the AMFM in differentiating IPF versus normal and sarcoidosis versus normal is maintained (greater than 90% overall accuracy), whereas the abilities of the MLD and HIST methods diminish in these subject groups, with overall accuracy of the MLD falling to 73% and the overall accuracy of the HIST falling to 55%. IPF versus sarcoidosis is more difficult to discriminate for all three methods. This is intuitive because the basic lung patterns caused by both diseases can often be similar. Even so, the AMFM still performs better than the MLD and HIST in this discrimination. Overall, in the two-subject group discriminations, the AMFM performed 2 to 18% better than the MLD and as much as 36% better than the HIST methods.

In the three-subject group discrimination, again, the AMFM performed substantially better than the MLD and HIST methods. Here, the AMFM performed 7 to 27% better than the MLD and 4 to 36% better than the HIST methods.

In differentiating all four subject groups simultaneously, the AMFM was 81% accurate overall, performing about 20% better than the MLD and above 25% better than the HIST. The four class comparison is a very important one as it is potentially the most useful for screening of HRCT slices. Four class comparisons generally are also much more complex, and they are a greater test of the usefulness of the methodologies as a generalizable tool. Here the four class comparisons demonstrates that the AMFM performs much better than the one-dimensional MLD and HIST. From all these results, it can be seen that the AMFM performed better than the currently available quantitative parenchymal evaluation methods.

After global analysis, regional analysis was performed to further study the AMFM. The purpose of the regional analysis was to ascertain the ability of the AMFM in identifying and differentiating local textures present in the diseased lungs versus the normal lungs. The regional discriminations between each group, emphysema, IPF, sarcoidosis, and normal, were successful. However, an observation to be noted from global as well as regional analysis is that sarcoidosis is harder to discriminate from normal since the nodular pattern caused by sarcoidosis is very subtle and may be similar to the normal lung pattern. In addition, sarcoidosis can mimic IPF in CT scans. The AMFM method can also be applied successfully to data accumulated from other CT scanners.

There were some correlations between the AMFM and the pulmonary function tests, especially for TLC and some indices of gas exchange. The association with TLC is not surprising given the significant influence of lung volume (especially % predicted) in regional lung expansion and thus the reconstructed radiographic attenuation values. The correlation with gas exchange indices (Pa_O2 and DL_CO) is of interest as the AMFM might be measuring structural lung components that relate to function. This will need further assessment.

We conclude from this study that the AMFM can be adapted to discriminate normal lung from a lung with interstitial lung diseases. The success of the AMFM lies in the fact that it uses multiple measures of texture as opposed to single measures by the MLD and HIST methods. Hence, the AMFM is able to suitably draw from the large database of features it has access to, for differentiating the tissues under consideration. The availability of several features that are measuring quite different textural properties contributed positively in enhancing the performance of the AMFM as compared with the MLD and HIST methods. The AMFM can be further enhanced by training it to recognize the basic lung patterns such as honeycombing, ground glass, nodular, etc., rather than global disease categories. By performing such pattern recognition tasks on a highly regionalized basis, it is expected that further improvements can be made in detection of and discrimination between pulmonary parenchymal pathology.