Note: This copy is for your personal non-commercial use only. To order presentation-ready copies for distribution to your colleagues or clients, contact us at www.rsna.org/rsnarights.
Purpose:
To determine whether the relationship between pulmonary function and computed tomographic (CT) lung attenuation in chronic obstructive pulmonary disease (COPD), which is traditionally described with single univariate and multivariate statistical models, could be more accurately described with a multiple model estimation approach.
Materials and Methods:
The study was approved by the local ethics committee. All participants provided written informed consent. The prediction of the percentage area with CT attenuation values less than 2950 HU at inspiration (%LAA2950insp) and less than 2910 HU at expiration (%LAA2910exp) obtained with single univariate and multivariate models was compared with that obtained with a multiple model estimation approach in 132 patients with COPD.
Results:
At univariate analysis, %LAA2950insp and %LAA2910exp values higher than the mean value of this cohort (19.1% and 22.0%) showed better correlation with percentage of predicted diffusing capacity of lung for carbon monoxide (Dlco%) than with airflow obstruction (forced expiratory volume in 1 second [FEV1]/vital capacity [VC]). Conversely, %LAA2950insp and %LAA2910exp values lower than the mean value were correlated with FEV1/VC but not with Dlco%. Multiple model estimation performed with two multivariate regressions, each selecting the most appropriate functional variables (FEV1/VC for mild parenchymal destruction, Dlco% and functional residual capacity for severe parenchymal destruction), predicted better than single multivariate regression both %LAA2950insp (R2 = 0.75 vs 0.46) and %LAA2910exp (R2 = 0.83 vs 0.63).
Conclusion:
The relationship between pulmonary function data and CT densitometric changes in COPD varies with the level of lung attenuation impairment. The nonlinear profile of this relationship is accurately predicted with a multiple model estimation approach. RSNA, 2015
q
Online supplemental material is available for this article. 1
From the Department of Experimental and Clinical Medicine, University of Florence, Largo Brambilla 3, 50134 Florence, Italy. Received July 28, 2014; revision requested September 18; revision received December 14; accepted January 6, 2015; final version accepted January 15. Address correspondence to M. Pistolesi (e-mail: massimo. [email protected]). RSNA, 2015
q
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
1
Imaging
Matteo Paoletti, PhD Lucia Cestelli, MD Francesca Bigazzi, MD Gianna Camiciottoli, MD Massimo Pistolesi, MD
Original Research n Thoracic
Chronic Obstructive Pulmonary Disease: Pulmonary Function and CT Lung Attenuation Do Not Show Linear Correlation1
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
C
omputed tomography (CT) is currently considered the method of choice for providing accurate in vivo information about the pathologic changes in the lung that occur in chronic obstructive pulmonary disease (COPD) (1,2). Many CT studies have been dedicated to in vivo quantification of the extent and severity of
Advances in Knowledge nn Considering the entire range of CT attenuation variability in chronic obstructive pulmonary disease (COPD), airflow obstruction (forced expiratory volume in 1 second [FEV1]/vital capacity [VC]) shows better correlation with CT attenuation values compatible with milder parenchymal destruction (FEV1/VC vs percentage area with CT attenuation values less than 2950 HU at inspiration [%LAA2950insp] , 19.1%, r = 20.40; FEV1/VC vs percentage area with CT attenuation values less than 2910 HU at expiration [%LAA2910exp] , 22.0%, r = 20.64; FEV1/VC 19.1% vs %LAA2950insp, r = 20.22, and vs %LAA2910exp 22.0%, r = 20.24), whereas the percentage of predicted diffusing capacity of lung for carbon monoxide (Dlco%) shows better correlation with CT attenuation values compatible with more severe parenchymal destruction (Dlco% vs %LAA2950insp , 19.1%, r = 20.17; Dlco% vs %LAA2910exp , 22.0%, r = 20.18; Dlco% vs %LAA2950insp 19.1%, r = 20.51; Dlco% vs %LAA2910exp 22.0%, r = 20.62). nn A statistical approach that considers the nonlinearity of the relationship between pulmonary function and CT attenuation values may predict with considerable accuracy CT-determined emphysema by using pulmonary function measurements (R2 = 0.75 for %LAA2950insp and 0.83 for %LAA2910exp). 2
emphysema (3–9). The percentage of lung area with CT attenuation values below predetermined thresholds is the most commonly adopted parameter with which to quantitatively assess lung parenchymal emphysematous destruction as reflected by reduced CT lung attenuation. Microscopic and macroscopic morphometric correlation studies (7–9) showed that the percentage of lung area with x-ray attenuation values of less than 2950 HU and less than 2910 HU can be used to approximate anatomic emphysema on CT scans obtained in inspiration and expiration, respectively. However, because CT measurements are not always available, it would be helpful to be able to estimate the CT-determined severity of emphysema by using functional measurements. The percentage of lung area with CT attenuation values compatible with emphysema has been shown to be related to functional measurements of airflow obstruction (10–14). However, several morphologic and CT studies (15–17) have shown the intrinsic limitations of spirometry for reflecting the different pathophysiologic mechanisms underlying airflow obstruction in COPD. Detection and quantification of emphysema extent with multivariate analysis was improved by adding diffusing capacity and hyperinflation indexes to spirometric data (18–20). Accurate prediction of the percentage of lung area with reduced x-ray attenuation values at CT by using pulmonary function could be of help to further characterize
Implications for Patient Care nn Prediction of CT-determined emphysema with pulmonary function measurements might be of use in clinical practice and in clinical or pharmacologic studies in which CT is not feasible or cost-effective. nn Single linear relationships between pulmonary function measurements and CT lung attenuation in COPD may underestimate the complexity and heterogeneity of the disease.
Paoletti et al
the lung impairment in either patients with COPD who are undergoing routine clinical evaluation or subjects enrolled in clinical and pharmacologic trials in which CT data acquisition is not routinely performed. To the best of our knowledge, the relationship between lung function and CT attenuation changes in COPD has always been studied by using standard linear methods such as Pearson correlation or multiple linear regression analysis. However, the pathophysiology of emphysema is complex and the net CT lung attenuation in COPD could result from intravoxel summation of reduced x-ray attenuation caused by overinflation and/or parenchymal destruction together with increased x-ray attenuation secondary to inflammatory changes (21). It is unlikely that such pathophysiologic processes will sum to an output that is well described with a single linear function. We performed this study to determine whether the relationship between Published online before print 10.1148/radiol.2015141769 Content code: Radiology 2015; 000:1–8 Abbreviations: COPD = chronic obstructive pulmonary disease Dlco% = percentage of predicted diffusing capacity of lung for carbon monoxide FEV1 = forced expiratory volume in 1 second FRC% = percentage of predicted functional residual capacity FVC = forced VC %LAA = percentage of lung attenuation area %LAA2910exp = percentage area with CT attenuation values less than 2910 HU at expiration %LAA2950insp = percentage area with CT attenuation values less than 2950 HU at inspiration TLC% = percentage of predicted total lung capacity VC = vital capacity Author contributions: Guarantors of integrity of entire study, F.B., M. Pistolesi; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; manuscript final version approval, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, L.C., F.B., G.C., M. Pistolesi; clinical studies, L.C., F.B., G.C., M. Pistolesi; statistical analysis, M. Paoletti; and manuscript editing, L.C., F.B., G.C., M. Pistolesi Conflicts of interest are listed at the end of this article.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
pulmonary function and CT lung attenuation in COPD, which is traditionally described with single univariate and multivariate statistical models, could be more accurately described with a multiple model estimation approach.
Materials and Methods Patients The study was approved by the local ethics committee. All participants provided written informed consent. From January 2012 to December 2013, we recruited through the outpatient clinic 132 nonconsecutive patients with COPD (mean age 6 standard deviation, 66 years 6 8). There were 103 men (mean age, 66.1 years) and 29 women (mean age, 63.3 years). The age difference between men and women was not significant with the t and Welch tests (P = .11 with both tests). All patients were smokers, with a mean smoking exposure of 47.8 pack-years 6 34.5. Patients recruited are a part of those participating in the Clinical Identification of Phenotypes of COPD, or CLIP-COPD, study. COPD was diagnosed according to Global Initiative for Chronic Obstructive Lung Disease criteria (22). We included patients aged 40–85 years with stage I–IV COPD and a smoking history of more than 10 pack-years who showed postbronchodilator nonreversible airflow obstruction and who gave written informed consent to undergo chest CT within 48 hours of pulmonary function evaluation. We excluded 193 subjects within 1 month of an exacerbation or with clinical conditions that could interfere with the assessment of pulmonary function or chest CT quantitative parameters, including asthma, diffuse bronchiectasis, cystic fibrosis, interstitial lung disease, acute heart failure, previous chemotherapy and/or radiation therapy, lung cancer, previous lung surgery, known or suspected pregnancy, and metal objects in the chest. Pulmonary Function Testing All patients underwent complete pulmonary function evaluation with use of a mass-flow sensor and multigas
analyzer (V6200 Autobox Body Plethysmograph; Sensor Medics, Yorba Linda, Calif). Pre- and postbronchodilator spirometric data, static lung volumes, and single-breath diffusing capacity for carbon monoxide were measured according to American Thoracic Society–European Respiratory Society guidelines (23–25).
CT Examination All patients underwent volumetric chest CT at full inspiration and full expiration. Patients had been previously instructed on how to perform the respiratory maneuvers while lying in the CT scanner acquisition bed. No contrast medium was used. All scans were obtained by the same team of diagnostic radiology personnel and with the same CT scanner (Sensation 64; Siemens, Erlangen, Germany). CT parameters were as follows: 120 kVp, 200 mAs (inspiratory), 50 mAs (expiratory), B31f reconstruction kernel, section thickness of 0.75 mm, and section interval of 0.5 mm. Image analysis was performed with a workstation (Pulmonary Workstation Apollo 1.0; VIDA Diagnostics, Coralville, Iowa). The extent of lung impairment was assessed by using the percentage of lung area with x-ray attenuation values of less than 2950 HU at full inspiration (%LAA2950insp) and the percentage of lung area with x-ray attenuation values of less than 2910 HU at full expiration (%LAA2910exp). Interactive quality control of the software performance was done by a bioengineer (M. Paoletti) with 11 years of experience in image analysis and medical statistics. Data Analysis Pearson r correlation coefficients were calculated to obtain a preliminary measurement of the association between pulmonary function and percentage of lung attenuation area (%LAA). A forward multiple linear step-wise regression analysis was then performed by using pulmonary function data, age, body mass index, and pack-years as independent variables to define models estimating %LAA2950insp and %LAA2910exp. Analysis of residuals was
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
Paoletti et al
performed and shown in conditioned box plots. The mean value of %LAA was arbitrarily adopted as a cut-off to compare patients with milder lung impairment (%LAA , mean value) and more severe lung impairment (%LAA . mean value). Separate r coefficients were calculated between pulmonary function indexes and the two %LAA ranges. An iterative learning routine (multiple model estimation) (26) was implemented to identify the optimum models reflecting the %LAA observed data distributions. Specific multivariate regression models were identified to estimate %LAA at different levels of lung attenuation. The %LAA thresholds for defining the two submodels were automatically determined during the learning process to optimize the R2 and the standard estimation error. For each submodel, the optimal set of functional predictors was identified by means of a step-wise selection. The identified models undergo internal validation by means of 10-fold crossvalidation (also known as rotation estimation). Additional details about the validation process and practical examples of the prospective application of the models are provided in Appendix E1 (online). P , .05 was considered indicative of a statistically significant difference. Data analysis and statistical analysis were performed by using software (SPSS/PC WIN 11.5.1 [SPSS, Chicago, Ill] and R, Mathcad, version 2001 [Mathsoft, Cambridge, Mass]), and the iterative learning routine was written by using ANSI C++ programming language.
Results The mean anthropometric data from the 132 patients enrolled in the study, together with their smoking habits, lung function data, %LAA2950insp, and %LAA2910exp, are shown in Table 1. Table 2 shows the correlation between CT quantitative values of %LAA and pulmonary function indexes. With the exception of TLC%, Pearson r values were found to be higher for %LAAthan for %LAA2950insp for all 2910exp pulmonary function variables tested. 3
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
Table 2
Table 1 Patient Characteristics Parameter No. of patients No. of men No. of women Age (y) Body mass index (kg/m2) Pack-years FVC% FEV1% FEV1/ VC FEV1/FVC TLC% RV% RV/TLC FRC% Dlco% %LAA2950insp %LAA2910exp
Value 132 103 29 66 6 8 25.4 6 4.2 47.8 6 34.5 85.8 6 21.7 54.1 6 21.4 44.3 6 12.8 50.6 6 12.7 113.1 6 15.1 155.7 6 47.8 51.5 6 10.9 136.9 6 29.8 69.6 6 22.5 19.16 12.4 22.0 6 15.9
Note.—Except where indicated, data are means 6 standard deviations. Dlco% = percentage of predicted diffusing capacity of lung for carbon monoxide, FEV1 = forced expiratory volume in 1 second, FEV1% = percentage predicted forced expiratory volume in 1 second, FRC% = percentage of predicted functional residual capacity, FVC = forced vital capacity, FVC% = percentage of predicted forced vital capacity, RV = residual volume, RV% = percentage of predicted residual volume, TLC = total lung capacity, TLC% = percentage of predicted total lung capacity, VC = vital capacity.
The highest value (r = 0.74, P , .001) was found for the correlation between FEV1/VC and %LAA2910exp.
Multivariate Regression Analysis Table 3 shows the single multivariate linear regression models in which %LAA2950insp and %LAA2910exp were used as dependent variables and pulmonary function variables, age, body mass index, and pack-years were used as independent variables. The step-wise algorithm selected FEV1/VC, Dlco%, and TLC% as predictors of %LAA2950insp (R2 = 0.46, P , .001) and FEV1/VC, Dlco%, and FRC% as predictors of %LAA2910exp (R2 = 0.63, P , .001). Figure 1 illustrates the relationships between %LAA and the functional predictors of the models selected with the step-wise algorithm. Mean values of %LAA2950insp and %LAA2910exp (19.1% and 22.0%, respectively) are 4
Table 3
Correlation between Pulmonary Function Indexes and CT Lung Attenuation Parameter* FVC% FEV1% FEV1/ VC FEV1/FVC TLC% RV% RV/TLC FRC% Dlco%
Paoletti et al
%LAA2950insp
%LAA2910exp
20.11 (.2) 20.44 (,.001) 20.54 (,.001) 20.53 (,.001) 0.44 (,.001) 0.44 (,.001) 0.22 (.01) 0.52 (,.001) 20.59 (,.001)
20.33 (,.001) 20.67 (,.001) 20.74 (,.001) 20.72 (,.001) 0.41 (,.001) 0.54 (,.001) 0.49 (,.001) 0.56 (,.001) 20.64 (,.001)
Single Multivariate Linear Regression Models to Predict CT Lung Attenuation with Pulmonary Function Indexes Variable and Predictor
Note.—Data are Pearson r correlation values. Numbers in parentheses are P values. * FEV1% = percentage predicted forced expiratory volume in 1 second, FVC% = percentage of predicted forced vital capacity, RV = residual volume, RV% = percentage of predicted residual volume, TLC = total lung capacity.
represented in the diagrams. The relationship of FEV1/VC and Dlco% with inspiratory and expiratory %LAA, as shown in Figure 1, has a nonlinear profile depending on the magnitude of %LAA changes. FEV1/VC is better related with %LAA values below the mean value line, whereas Dlco% is better related with %LAA values above the mean value line. Numerically, the Pearson correlation coefficients between the functional predictors and %LAA in the two separate ranges (higher and lower than the mean value) were calculated. FEV1/VC showed consistent correlation with a %LAA2950insp of less than 19.1% and a %LAA2910exp of less than 22.0% (r = 20.40, P , .001 and r = 20.64, P , .001, respectively). For a %LAA2950insp of at least 19.1% and a %LAA2910exp of at least 22.0%, however, the correlations decreased to lower levels (r = 20.22, P = .02 and r = 20.24, P = .004, respectively). Conversely, Dlco% shows a significant correlation only with a %LAA2950insp of at least 19.1% and a %LAA2910exp of at least 22.0% (r = 20.51, P , .001 and r = 20.62, P , .001, respectively), whereas the correlations become not significant for a %LAA2950inof less than 19.1% and a %LAA2910exp sp of less than 22.0% (r = 20.17, P = .21
%LAA2950insp FEV1/VC Dlco% TLC% Intercept %LAA2910exp FEV1/VC Dlco% FRC% Intercept
Coefficient
20.19 20.23 0.25 14.94 20.60 20.21 0.08 52.07
Note.—The step-wise algorithm selected FEV1/VC, Dlco%, and TLC% as predictors of %LAA2950insp (R2 = 0.46, P , .001) and FEV1/VC, Dlco%, and FRC% as predictors of %LAA2910exp (R2 = 0.63, P , .001).
and r = 20.18, P = .12, respectively). Although with great dispersion, FRC% showed a more stable trend over the entire attenuation range (r = 0.38, P , .001 with %LAA2910exp , 22.0% and r = 0.39, P , .001 with %LAA2910exp 22.0%), whereas TLC% had a weak correlation only with a %LAA2950insp of at least 19.1% (r = 0.32, P , .01) and was not significant with a %LAA2950insp of less than 19.1% (r = 0.17, P = .14). The box-and-whiskers plot in Figure 2 shows the distribution of the standardized residuals of %LAA2950insp and %LAA2910exp with multivariate linear regressions. The standardized residuals represent the differences between the observed and estimated values. The residuals are not randomly dispersed around the horizontal axis, which indicates that the associations are not linear. In particular, %LAA was systematically overestimated (negative residuals) in patients with %LAA lower than its mean value and underestimated (positive residuals) in patients with %LAA higher than its mean value.
Multiple Model Estimation Table 4 shows the characteristics of the multivariate regressions developed by applying the multiple model estimation approach. Two specific
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
Paoletti et al
Figure 1
Figure 1: Relationships between %LAA and the functional predictors of single multivariate linear regression models. Correlations between pulmonary function variables and %LAA show lower dispersion when considering %LAA2910exp with respect to %LAA2950insp. Horizontal lines represent mean values of %LAA2950insp and %LAA2910exp (19.1% and 22.0%, respectively). A nonlinear relationship profile, depending on level of %LAA increase, can be observed for both FEV1/VC and Dlco%.
Figure 2
Figure 2: Box-and-whiskers plots show distribution of standardized residuals obtained after applying multivariate linear regression to model %LAA2950insp and %LAA2910exp by pulmonary function variables in patients with different severity of x-ray lung attenuation at CT. To condition box-and-whiskers plots, cohort was divided into two groups and the cut-off values were selected at %LAA2950insp and %LAA2910exp mean values (19.1% and 22.0%, respectively). Both models systematically overestimated %LAA (negative residuals) in patients with milder reduction of x-ray lung attenuation at CT and underestimated %LAA (positive residuals) in patients with greater reduction of x-ray lung attenuation at CT.
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
multivariate regression models were identified: submodel A, which predicts lower levels of reduced x-ray attenuation (%LAA2950insp 21%, %LAA2910exp 26%), and submodel B, which predicts higher levels of reduced x-ray attenuation (%LAA2950insp . 21%, %LAA2910exp . 26%). The %LAA ranges of each submodel were automatically determined to optimize the R2 and the standard error of the two regressions (the two best-fitting models were selected). Among all functional regressors identified with the step-wise process, only FEV1/ VC was selected to predict both %LAA2950insp of 21% or less and %LAAof 26% or less. Conversely, 2910exp Dlco% was the predictor selected for both %LAA2950insp greater than 21% and %LAA2910exp greater than 26%. FRC% was selected together with Dlco% to predict %LAA2950insp greater than 21%. The predicting performances of the multiple models 5
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
Table 4 Multiple Model Estimation to Predict CT Lung Attenuation with Pulmonary Function Indexes Variable, Submodel, and Predictor %LAA2950insp Submodel A FEV1/VC Intercept Submodel B Dlco% FRC% Intercept %LAA2910exp Submodel A FEV1/VC Intercept Submodel B Dlco% Intercept
Coefficient
20.25 23.81 20.12 0.12 20.88
20.51 38.06 20.42 62.87
Note.—Submodel A: %LAA2950insp 21% and %LAA2910exp 26%. Submodel B: %LAA2950insp . 21% and %LAA2910exp . 26%. The threshold values were automatically selected during the learning process to optimize the R2 and the standard error of the two regressions. The predicting performances of the multiple models were much higher than those of the single models (R2 = 0.75 [P , .001] vs 0.46 [P , .001], respectively, for %LAA2950insp and R2 = 0.83 [P , .001] vs 0.63 [P , .001] for %LAA2910exp).
were much higher than those of the single models (R2 = 0.75 vs 0.46, respectively, for %LAA2950insp and R2 = 0.83 vs 0.63 for %LAA2910exp) (Tables 3, 4), which shows that the twofold models are able to cover almost the entire variability range of the original CT data.
Discussion The main results of this study are as follows. First, COPD pulmonary function measurements are not linearly related to CT lung attenuation over the full range of observed attenuations. Second, a multiple model approach that combines measurements of airflow obstruction (FEV1/VC), overinflation (FRC%), and parenchymal destruction (Dlco%) can accurately predict the inspiratory and expiratory %LAA over a wide range of values. These results are in keeping with the 6
notion that COPD is a heterogeneous condition in which inflammatory and destructive changes that affect airways and parenchyma may combine to cause the lung attenuation changes measured with CT. Similar to previous investigators (18–20), we found that linear multivariate analysis only partially improves the predictive ability of each of the single functional variables of obstruction, overinflation, and parenchymal destruction. We also noticed that the relationships between some functional predictors and %LAA were not linear but varied depending on the degree of the CT densitometric alteration. In particular, despite the overall fair Pearson correlation coefficients, for %LAA values compatible with greater parenchymal destruction a very weak association with FEV1/ VC was evidenced, whereas for %LAA values compatible with lower or absent parenchymal destruction no significant association with Dlco% was demonstrated. Accordingly, if we adopted a standard correlation-based single-model approach, we could reach the misleading conclusion that CT parenchymal destruction varies proportionally to the severity of airflow limitation or to the reduction in diffusing capacity, and vice-versa. In our dataset, this is true only in limited ranges of %LAA. This twofold profile heavily affects the performances of traditional multivariate regression models. The analysis of the residuals in our cohort confirms the notion that forcing a single linear model approach in the presence of nonlinear relationships may produce a systematic under- or overestimation of the predicted parameter (27). The complexity of COPD cannot be expressed with a simple measurement of expiratory airflow obstruction (28–32). Even though the percentage of emphysema has been reported to increase with increasing Global Initiative for Chronic Obstructive Lung Disease stages (14,28), there is a considerable amount of discordance unexplained by the current classification approach (14–17). Recent data obtained in a large cohort of patients with COPD
Paoletti et al
demonstrated that, when correlated with %LAA2950insp, spirometric indexes of obstruction (eg, FEV1 and FEV1/ FVC) fail to determine with accuracy the presence and the severity of emphysema as reflected by quantitative CT evaluation (14,17). In this line of evidence, the results of the present study may be of clinical significance. Our results showed that, for high values of %LAA2950insp, the contribution of FEV1/VC was not significant to predict parenchymal destruction, whereas Dlco% accounted for a large amount of that variation with a small contribution by FRC%. Conversely, Dlco% did not significantly predict lower values of %LAA2950insp. Another interesting finding was that, considering %LAA as a parameter of reference, FEV1/VC instead of FEV1/ FVC was the airflow obstruction index invariably selected by the stepwise algorithms. This may point to the known greater accuracy of FEV1/VC with respect to FEV1/FVC in the assessment of airway obstruction in COPD. Recently, Bafadhel et al (33) showed that diffusing capacity has great accuracy in the identification of CT-detected emphysema. In general, the results of our study may indicate that this relationship could be particularly accurate in those patients who are more severely affected by emphysema and confirm that diffusing capacity has a relevant role in predicting the level of parenchymal destruction as assessed with quantitative CT. The most recent large-scale clinical trials considered the measurement of %LAA2950insp at CT as an accurate assessment of the amount of emphysema (34,35). The relationship between pulmonary function data and CT quantitative parameters observed in this study, if confirmed and prospectively validated in larger samples, may be of use to obtain a fairly accurate evaluation of lung attenuation impairment for clinical and pharmacologic studies in which the use of CT might not be feasible or cost effective. Our study has some limitations. The number of patients should be increased to achieve a more reliable
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
learning process in the identification of the specific predicting models. Another intrinsic limitation is the absence of a reference standard that could help verify the real extent of emphysema, inflammation, and other pathophysiologic processes. However, the values of %LAA2950insp in our cohort could be considered to substantially cover the variability in parenchymal destruction reported in similar larger-scale studies. We also included expiratory %LAA2910exp in our analysis. The main limit of this measurement is the lower accuracy of the parameter itself in representing pulmonary emphysema. In fact, both pathologic alterations due to loss of elastic recoil and remodeling of small airways could be potentially reflected in the CT detection of decreased expiratory lung attenuation (9). A further limitation is that the CT data were acquired, as in the majority of similar studies, without control of the level of lung inflation. This could have increased the variation of the relationships found. In summary, this study demonstrated that the relationship between lung function and CT parenchymal destruction in COPD is nonlinear. In particular, only FEV1/VC was found to show significant correlation with mild parenchymal destruction. Dlco% was, conversely, the variable accounting for most of the variation in the presence of more severe parenchymal destruction. To overcome the observed nonlinearity of the relationship, we identified two specific linear models that better fit the actual data distribution over its full range of variability and that may be used in each patient to predict with an acceptable level of accuracy “CT-determined emphysema” in clinical and pharmacologic studies. Disclosures of Conflicts of Interest: M.P. disclosed no relevant relationships. L.C. disclosed no relevant relationships. F.B. disclosed no relevant relationships. G.C. disclosed no relevant relationships. M.P. disclosed no relevant relationships.
References 1. Lynch DA, Al-Qaisi MA. Quantitative computed tomography in chronic obstruc-
tive pulmonary disease. J Thorac Imaging 2013;28(5):284–290. 2. Litmanovich DE, Hartwick K, Silva M, Bankier AA. Multidetector computed tomographic imaging in chronic obstructive pulmonary disease: emphysema and airways assessment. Radiol Clin North Am 2014;52(1):137–154. 3. Hayhurst MD, MacNee W, Flenley DC, et al. Diagnosis of pulmonary emphysema by computerised tomography. Lancet 1984; 2(8398):320–322. 4. Müller NL, Staples CA, Miller RR, Abboud RT. “Density mask:” an objective method to quantitate emphysema using computed tomography. Chest 1988;94(4):782–787. 5. Knudson RJ, Standen JR, Kaltenborn WT, et al. Expiratory computed tomography for assessment of suspected pulmonary emphysema. Chest 1991;99(6):1357– 1366. 6. Miniati M, Filippi E, Falaschi F, et al. Radiologic evaluation of emphysema in patients with chronic obstructive pulmonary disease: chest radiography versus high resolution computed tomography. Am J Respir Crit Care Med 1995;151(5):1359– 1367. 7. Gevenois PA, de Maertelaer V, De Vuyst P, Zanen J, Yernault JC. Comparison of computed density and macroscopic morphometry in pulmonary emphysema. Am J Respir Crit Care Med 1995;152(2):653– 657.
Paoletti et al
13. Xie X, de Jong PA, Oudkerk M, et al. Morphological measurements in computed tomography correlate with airflow obstruction in chronic obstructive pulmonary disease: systematic review and metaanalysis. Eur Radiol 2012;22(10):2085– 2093. 14. Schroeder JD, McKenzie AS, Zach JA, et al. Relationships between airflow obstruction and quantitative CT measurements of emphysema, air trapping, and airways in subjects with and without chronic obstructive pulmonary disease. AJR Am J Roentgenol 2013;201(3):W460–W470. 15. Hogg JC, Wright JL, Wiggs BR, Coxson HO, Opazo Saez A, Paré PD. Lung structure and function in cigarette smokers. Thorax 1994;49(5):473–478. 16. Gelb AF, Hogg JC, Müller NL, et al. Contribution of emphysema and small airways in COPD. Chest 1996;109(2):353–359. 17. Bhatt SP, Sieren JC, Dransfield MT, et al. Comparison of spirometric thresholds in diagnosing smoking-related airflow obstruction. Thorax 2014;69(5):410–415. 18. Cerveri I, Dore R, Corsico A, et al. As sessment of emphysema in COPD: a functional and radiologic study. Chest 2004;125(5):1714–1718. 19. Desai SR, Hansell DM, Walker A, MacDonald SL, Chabat F, Wells AU. Quantification of emphysema: a composite physiologic index derived from CT estimation of disease extent. Eur Radiol 2007;17(4):911–918.
8. Gevenois PA, De Vuyst P, de Maertelaer V, et al. Comparison of computed density and microscopic morphometry in pulmonary emphysema. Am J Respir Crit Care Med 1996;154(1):187–192.
20. Hesselbacher SE, Ross R, Schabath MB, et al. Cross-sectional analysis of the utility of pulmonary function tests in predicting emphysema in ever-smokers. Int J Environ Res Public Health 2011;8(5):1324–1340.
9. Gevenois PA, De Vuyst P, Sy M, et al. Pulmonary emphysema: quantitative CT during expiration. Radiology 1996;199(3):825–829.
21. Karimi R, Tornling G, Forsslund H, et al. Lung density on high resolution computer tomography (HRCT) reflects degree of inflammation in smokers. Respir Res 2014;15:23.
10. 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(2):315–321. 11. Nakano Y, Muro S, Sakai H, et al. Computed tomographic measurements of airway dimensions and emphysema in smokers: correlation with lung function. Am J Respir Crit Care Med 2000;162(3 Pt 1): 1102–1108. 12. Camiciottoli G, Bartolucci M, Maluccio NM, et al. Spirometrically gated high-resolution CT findings in COPD: lung attenuation vs lung function and dyspnea severity. Chest 2006;129(3):558–564.
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
22. From the Global Strategy for the Diagno sis, Management and Prevention of COPD, Global Initiative for Chronic Obstructive Lung Disease (GOLD). http://www.goldcopd.org/. Published 2014. Accessed June 26, 2014. 23. Miller MR, Hankinson J, Brusasco V, et al. Standardisation of spirometry. Eur Respir J 2005;26(2):319–338. 24. Wanger J, Clausen JL, Coates A, et al. Standardisation of the measurement of lung volumes. Eur Respir J 2005;26(3):511–522. 25. Macintyre N, Crapo RO, Viegi G, et al. Standardisation of the single-breath deter-
7
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
mination of carbon monoxide uptake in the lung. Eur Respir J 2005;26(4):720–735. 26. Cherkassky V, Ma Y. Multiple model regression estimation. IEEE Trans Neural Netw 2005;16(4):785–798. 27. Anscombe FJ. Graphs in statistical analysis. Am Stat 1973;27(1):17–21. 28. Agusti A, Calverley PM, Celli B, et al. Characterisation of COPD heterogeneity in the ECLIPSE cohort. Respir Res 2010; 11:122. 29. Wedzicha JA. The heterogeneity of chronic obstructive pulmonary disease. Thorax 2000; 55(8):631–632.
8
30. Pistolesi M. Beyond airflow limitation: another look at COPD. Thorax 2009;64(1): 2–4. 31. Han MK, Agusti A, Calverley PM, et al. Chronic obstructive pulmonary disease phenotypes: the future of COPD. Am J Respir Crit Care Med 2010;182(5):598–604. 32. Casanova C, de Torres JP, Aguirre-Jaíme A, et al. The progression of chronic obstructive pulmonary disease is heterogeneous: the experience of the BODE cohort. Am J Respir Crit Care Med 2011;184(9):1015–1021.
Paoletti et al
typing of COPD. Chest 2011;140(3):634– 642. 34. Vestbo J, Anderson W, Coxson HO, et al. Evaluation of COPD longitudinally to identify predictive surrogate end-points (ECLIPSE). Eur Respir J 2008;31(4):869–873. 35. Regan EA, Hokanson JE, Murphy JR, et al. Genetic epidemiology of COPD (COPDGene) study design. COPD 2010;7(1):32–43.
33. Bafadhel M, Umar I, Gupta S, et al. The role of CT scanning in multidimensional pheno-
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
Purpose:
To determine whether the relationship between pulmonary function and computed tomographic (CT) lung attenuation in chronic obstructive pulmonary disease (COPD), which is traditionally described with single univariate and multivariate statistical models, could be more accurately described with a multiple model estimation approach.
Materials and Methods:
The study was approved by the local ethics committee. All participants provided written informed consent. The prediction of the percentage area with CT attenuation values less than 2950 HU at inspiration (%LAA2950insp) and less than 2910 HU at expiration (%LAA2910exp) obtained with single univariate and multivariate models was compared with that obtained with a multiple model estimation approach in 132 patients with COPD.
Results:
At univariate analysis, %LAA2950insp and %LAA2910exp values higher than the mean value of this cohort (19.1% and 22.0%) showed better correlation with percentage of predicted diffusing capacity of lung for carbon monoxide (Dlco%) than with airflow obstruction (forced expiratory volume in 1 second [FEV1]/vital capacity [VC]). Conversely, %LAA2950insp and %LAA2910exp values lower than the mean value were correlated with FEV1/VC but not with Dlco%. Multiple model estimation performed with two multivariate regressions, each selecting the most appropriate functional variables (FEV1/VC for mild parenchymal destruction, Dlco% and functional residual capacity for severe parenchymal destruction), predicted better than single multivariate regression both %LAA2950insp (R2 = 0.75 vs 0.46) and %LAA2910exp (R2 = 0.83 vs 0.63).
Conclusion:
The relationship between pulmonary function data and CT densitometric changes in COPD varies with the level of lung attenuation impairment. The nonlinear profile of this relationship is accurately predicted with a multiple model estimation approach. RSNA, 2015
q
Online supplemental material is available for this article. 1
From the Department of Experimental and Clinical Medicine, University of Florence, Largo Brambilla 3, 50134 Florence, Italy. Received July 28, 2014; revision requested September 18; revision received December 14; accepted January 6, 2015; final version accepted January 15. Address correspondence to M. Pistolesi (e-mail: massimo. [email protected]). RSNA, 2015
q
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
1
Imaging
Matteo Paoletti, PhD Lucia Cestelli, MD Francesca Bigazzi, MD Gianna Camiciottoli, MD Massimo Pistolesi, MD
Original Research n Thoracic
Chronic Obstructive Pulmonary Disease: Pulmonary Function and CT Lung Attenuation Do Not Show Linear Correlation1
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
C
omputed tomography (CT) is currently considered the method of choice for providing accurate in vivo information about the pathologic changes in the lung that occur in chronic obstructive pulmonary disease (COPD) (1,2). Many CT studies have been dedicated to in vivo quantification of the extent and severity of
Advances in Knowledge nn Considering the entire range of CT attenuation variability in chronic obstructive pulmonary disease (COPD), airflow obstruction (forced expiratory volume in 1 second [FEV1]/vital capacity [VC]) shows better correlation with CT attenuation values compatible with milder parenchymal destruction (FEV1/VC vs percentage area with CT attenuation values less than 2950 HU at inspiration [%LAA2950insp] , 19.1%, r = 20.40; FEV1/VC vs percentage area with CT attenuation values less than 2910 HU at expiration [%LAA2910exp] , 22.0%, r = 20.64; FEV1/VC 19.1% vs %LAA2950insp, r = 20.22, and vs %LAA2910exp 22.0%, r = 20.24), whereas the percentage of predicted diffusing capacity of lung for carbon monoxide (Dlco%) shows better correlation with CT attenuation values compatible with more severe parenchymal destruction (Dlco% vs %LAA2950insp , 19.1%, r = 20.17; Dlco% vs %LAA2910exp , 22.0%, r = 20.18; Dlco% vs %LAA2950insp 19.1%, r = 20.51; Dlco% vs %LAA2910exp 22.0%, r = 20.62). nn A statistical approach that considers the nonlinearity of the relationship between pulmonary function and CT attenuation values may predict with considerable accuracy CT-determined emphysema by using pulmonary function measurements (R2 = 0.75 for %LAA2950insp and 0.83 for %LAA2910exp). 2
emphysema (3–9). The percentage of lung area with CT attenuation values below predetermined thresholds is the most commonly adopted parameter with which to quantitatively assess lung parenchymal emphysematous destruction as reflected by reduced CT lung attenuation. Microscopic and macroscopic morphometric correlation studies (7–9) showed that the percentage of lung area with x-ray attenuation values of less than 2950 HU and less than 2910 HU can be used to approximate anatomic emphysema on CT scans obtained in inspiration and expiration, respectively. However, because CT measurements are not always available, it would be helpful to be able to estimate the CT-determined severity of emphysema by using functional measurements. The percentage of lung area with CT attenuation values compatible with emphysema has been shown to be related to functional measurements of airflow obstruction (10–14). However, several morphologic and CT studies (15–17) have shown the intrinsic limitations of spirometry for reflecting the different pathophysiologic mechanisms underlying airflow obstruction in COPD. Detection and quantification of emphysema extent with multivariate analysis was improved by adding diffusing capacity and hyperinflation indexes to spirometric data (18–20). Accurate prediction of the percentage of lung area with reduced x-ray attenuation values at CT by using pulmonary function could be of help to further characterize
Implications for Patient Care nn Prediction of CT-determined emphysema with pulmonary function measurements might be of use in clinical practice and in clinical or pharmacologic studies in which CT is not feasible or cost-effective. nn Single linear relationships between pulmonary function measurements and CT lung attenuation in COPD may underestimate the complexity and heterogeneity of the disease.
Paoletti et al
the lung impairment in either patients with COPD who are undergoing routine clinical evaluation or subjects enrolled in clinical and pharmacologic trials in which CT data acquisition is not routinely performed. To the best of our knowledge, the relationship between lung function and CT attenuation changes in COPD has always been studied by using standard linear methods such as Pearson correlation or multiple linear regression analysis. However, the pathophysiology of emphysema is complex and the net CT lung attenuation in COPD could result from intravoxel summation of reduced x-ray attenuation caused by overinflation and/or parenchymal destruction together with increased x-ray attenuation secondary to inflammatory changes (21). It is unlikely that such pathophysiologic processes will sum to an output that is well described with a single linear function. We performed this study to determine whether the relationship between Published online before print 10.1148/radiol.2015141769 Content code: Radiology 2015; 000:1–8 Abbreviations: COPD = chronic obstructive pulmonary disease Dlco% = percentage of predicted diffusing capacity of lung for carbon monoxide FEV1 = forced expiratory volume in 1 second FRC% = percentage of predicted functional residual capacity FVC = forced VC %LAA = percentage of lung attenuation area %LAA2910exp = percentage area with CT attenuation values less than 2910 HU at expiration %LAA2950insp = percentage area with CT attenuation values less than 2950 HU at inspiration TLC% = percentage of predicted total lung capacity VC = vital capacity Author contributions: Guarantors of integrity of entire study, F.B., M. Pistolesi; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; manuscript final version approval, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, L.C., F.B., G.C., M. Pistolesi; clinical studies, L.C., F.B., G.C., M. Pistolesi; statistical analysis, M. Paoletti; and manuscript editing, L.C., F.B., G.C., M. Pistolesi Conflicts of interest are listed at the end of this article.
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
pulmonary function and CT lung attenuation in COPD, which is traditionally described with single univariate and multivariate statistical models, could be more accurately described with a multiple model estimation approach.
Materials and Methods Patients The study was approved by the local ethics committee. All participants provided written informed consent. From January 2012 to December 2013, we recruited through the outpatient clinic 132 nonconsecutive patients with COPD (mean age 6 standard deviation, 66 years 6 8). There were 103 men (mean age, 66.1 years) and 29 women (mean age, 63.3 years). The age difference between men and women was not significant with the t and Welch tests (P = .11 with both tests). All patients were smokers, with a mean smoking exposure of 47.8 pack-years 6 34.5. Patients recruited are a part of those participating in the Clinical Identification of Phenotypes of COPD, or CLIP-COPD, study. COPD was diagnosed according to Global Initiative for Chronic Obstructive Lung Disease criteria (22). We included patients aged 40–85 years with stage I–IV COPD and a smoking history of more than 10 pack-years who showed postbronchodilator nonreversible airflow obstruction and who gave written informed consent to undergo chest CT within 48 hours of pulmonary function evaluation. We excluded 193 subjects within 1 month of an exacerbation or with clinical conditions that could interfere with the assessment of pulmonary function or chest CT quantitative parameters, including asthma, diffuse bronchiectasis, cystic fibrosis, interstitial lung disease, acute heart failure, previous chemotherapy and/or radiation therapy, lung cancer, previous lung surgery, known or suspected pregnancy, and metal objects in the chest. Pulmonary Function Testing All patients underwent complete pulmonary function evaluation with use of a mass-flow sensor and multigas
analyzer (V6200 Autobox Body Plethysmograph; Sensor Medics, Yorba Linda, Calif). Pre- and postbronchodilator spirometric data, static lung volumes, and single-breath diffusing capacity for carbon monoxide were measured according to American Thoracic Society–European Respiratory Society guidelines (23–25).
CT Examination All patients underwent volumetric chest CT at full inspiration and full expiration. Patients had been previously instructed on how to perform the respiratory maneuvers while lying in the CT scanner acquisition bed. No contrast medium was used. All scans were obtained by the same team of diagnostic radiology personnel and with the same CT scanner (Sensation 64; Siemens, Erlangen, Germany). CT parameters were as follows: 120 kVp, 200 mAs (inspiratory), 50 mAs (expiratory), B31f reconstruction kernel, section thickness of 0.75 mm, and section interval of 0.5 mm. Image analysis was performed with a workstation (Pulmonary Workstation Apollo 1.0; VIDA Diagnostics, Coralville, Iowa). The extent of lung impairment was assessed by using the percentage of lung area with x-ray attenuation values of less than 2950 HU at full inspiration (%LAA2950insp) and the percentage of lung area with x-ray attenuation values of less than 2910 HU at full expiration (%LAA2910exp). Interactive quality control of the software performance was done by a bioengineer (M. Paoletti) with 11 years of experience in image analysis and medical statistics. Data Analysis Pearson r correlation coefficients were calculated to obtain a preliminary measurement of the association between pulmonary function and percentage of lung attenuation area (%LAA). A forward multiple linear step-wise regression analysis was then performed by using pulmonary function data, age, body mass index, and pack-years as independent variables to define models estimating %LAA2950insp and %LAA2910exp. Analysis of residuals was
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
Paoletti et al
performed and shown in conditioned box plots. The mean value of %LAA was arbitrarily adopted as a cut-off to compare patients with milder lung impairment (%LAA , mean value) and more severe lung impairment (%LAA . mean value). Separate r coefficients were calculated between pulmonary function indexes and the two %LAA ranges. An iterative learning routine (multiple model estimation) (26) was implemented to identify the optimum models reflecting the %LAA observed data distributions. Specific multivariate regression models were identified to estimate %LAA at different levels of lung attenuation. The %LAA thresholds for defining the two submodels were automatically determined during the learning process to optimize the R2 and the standard estimation error. For each submodel, the optimal set of functional predictors was identified by means of a step-wise selection. The identified models undergo internal validation by means of 10-fold crossvalidation (also known as rotation estimation). Additional details about the validation process and practical examples of the prospective application of the models are provided in Appendix E1 (online). P , .05 was considered indicative of a statistically significant difference. Data analysis and statistical analysis were performed by using software (SPSS/PC WIN 11.5.1 [SPSS, Chicago, Ill] and R, Mathcad, version 2001 [Mathsoft, Cambridge, Mass]), and the iterative learning routine was written by using ANSI C++ programming language.
Results The mean anthropometric data from the 132 patients enrolled in the study, together with their smoking habits, lung function data, %LAA2950insp, and %LAA2910exp, are shown in Table 1. Table 2 shows the correlation between CT quantitative values of %LAA and pulmonary function indexes. With the exception of TLC%, Pearson r values were found to be higher for %LAAthan for %LAA2950insp for all 2910exp pulmonary function variables tested. 3
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
Table 2
Table 1 Patient Characteristics Parameter No. of patients No. of men No. of women Age (y) Body mass index (kg/m2) Pack-years FVC% FEV1% FEV1/ VC FEV1/FVC TLC% RV% RV/TLC FRC% Dlco% %LAA2950insp %LAA2910exp
Value 132 103 29 66 6 8 25.4 6 4.2 47.8 6 34.5 85.8 6 21.7 54.1 6 21.4 44.3 6 12.8 50.6 6 12.7 113.1 6 15.1 155.7 6 47.8 51.5 6 10.9 136.9 6 29.8 69.6 6 22.5 19.16 12.4 22.0 6 15.9
Note.—Except where indicated, data are means 6 standard deviations. Dlco% = percentage of predicted diffusing capacity of lung for carbon monoxide, FEV1 = forced expiratory volume in 1 second, FEV1% = percentage predicted forced expiratory volume in 1 second, FRC% = percentage of predicted functional residual capacity, FVC = forced vital capacity, FVC% = percentage of predicted forced vital capacity, RV = residual volume, RV% = percentage of predicted residual volume, TLC = total lung capacity, TLC% = percentage of predicted total lung capacity, VC = vital capacity.
The highest value (r = 0.74, P , .001) was found for the correlation between FEV1/VC and %LAA2910exp.
Multivariate Regression Analysis Table 3 shows the single multivariate linear regression models in which %LAA2950insp and %LAA2910exp were used as dependent variables and pulmonary function variables, age, body mass index, and pack-years were used as independent variables. The step-wise algorithm selected FEV1/VC, Dlco%, and TLC% as predictors of %LAA2950insp (R2 = 0.46, P , .001) and FEV1/VC, Dlco%, and FRC% as predictors of %LAA2910exp (R2 = 0.63, P , .001). Figure 1 illustrates the relationships between %LAA and the functional predictors of the models selected with the step-wise algorithm. Mean values of %LAA2950insp and %LAA2910exp (19.1% and 22.0%, respectively) are 4
Table 3
Correlation between Pulmonary Function Indexes and CT Lung Attenuation Parameter* FVC% FEV1% FEV1/ VC FEV1/FVC TLC% RV% RV/TLC FRC% Dlco%
Paoletti et al
%LAA2950insp
%LAA2910exp
20.11 (.2) 20.44 (,.001) 20.54 (,.001) 20.53 (,.001) 0.44 (,.001) 0.44 (,.001) 0.22 (.01) 0.52 (,.001) 20.59 (,.001)
20.33 (,.001) 20.67 (,.001) 20.74 (,.001) 20.72 (,.001) 0.41 (,.001) 0.54 (,.001) 0.49 (,.001) 0.56 (,.001) 20.64 (,.001)
Single Multivariate Linear Regression Models to Predict CT Lung Attenuation with Pulmonary Function Indexes Variable and Predictor
Note.—Data are Pearson r correlation values. Numbers in parentheses are P values. * FEV1% = percentage predicted forced expiratory volume in 1 second, FVC% = percentage of predicted forced vital capacity, RV = residual volume, RV% = percentage of predicted residual volume, TLC = total lung capacity.
represented in the diagrams. The relationship of FEV1/VC and Dlco% with inspiratory and expiratory %LAA, as shown in Figure 1, has a nonlinear profile depending on the magnitude of %LAA changes. FEV1/VC is better related with %LAA values below the mean value line, whereas Dlco% is better related with %LAA values above the mean value line. Numerically, the Pearson correlation coefficients between the functional predictors and %LAA in the two separate ranges (higher and lower than the mean value) were calculated. FEV1/VC showed consistent correlation with a %LAA2950insp of less than 19.1% and a %LAA2910exp of less than 22.0% (r = 20.40, P , .001 and r = 20.64, P , .001, respectively). For a %LAA2950insp of at least 19.1% and a %LAA2910exp of at least 22.0%, however, the correlations decreased to lower levels (r = 20.22, P = .02 and r = 20.24, P = .004, respectively). Conversely, Dlco% shows a significant correlation only with a %LAA2950insp of at least 19.1% and a %LAA2910exp of at least 22.0% (r = 20.51, P , .001 and r = 20.62, P , .001, respectively), whereas the correlations become not significant for a %LAA2950inof less than 19.1% and a %LAA2910exp sp of less than 22.0% (r = 20.17, P = .21
%LAA2950insp FEV1/VC Dlco% TLC% Intercept %LAA2910exp FEV1/VC Dlco% FRC% Intercept
Coefficient
20.19 20.23 0.25 14.94 20.60 20.21 0.08 52.07
Note.—The step-wise algorithm selected FEV1/VC, Dlco%, and TLC% as predictors of %LAA2950insp (R2 = 0.46, P , .001) and FEV1/VC, Dlco%, and FRC% as predictors of %LAA2910exp (R2 = 0.63, P , .001).
and r = 20.18, P = .12, respectively). Although with great dispersion, FRC% showed a more stable trend over the entire attenuation range (r = 0.38, P , .001 with %LAA2910exp , 22.0% and r = 0.39, P , .001 with %LAA2910exp 22.0%), whereas TLC% had a weak correlation only with a %LAA2950insp of at least 19.1% (r = 0.32, P , .01) and was not significant with a %LAA2950insp of less than 19.1% (r = 0.17, P = .14). The box-and-whiskers plot in Figure 2 shows the distribution of the standardized residuals of %LAA2950insp and %LAA2910exp with multivariate linear regressions. The standardized residuals represent the differences between the observed and estimated values. The residuals are not randomly dispersed around the horizontal axis, which indicates that the associations are not linear. In particular, %LAA was systematically overestimated (negative residuals) in patients with %LAA lower than its mean value and underestimated (positive residuals) in patients with %LAA higher than its mean value.
Multiple Model Estimation Table 4 shows the characteristics of the multivariate regressions developed by applying the multiple model estimation approach. Two specific
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
Paoletti et al
Figure 1
Figure 1: Relationships between %LAA and the functional predictors of single multivariate linear regression models. Correlations between pulmonary function variables and %LAA show lower dispersion when considering %LAA2910exp with respect to %LAA2950insp. Horizontal lines represent mean values of %LAA2950insp and %LAA2910exp (19.1% and 22.0%, respectively). A nonlinear relationship profile, depending on level of %LAA increase, can be observed for both FEV1/VC and Dlco%.
Figure 2
Figure 2: Box-and-whiskers plots show distribution of standardized residuals obtained after applying multivariate linear regression to model %LAA2950insp and %LAA2910exp by pulmonary function variables in patients with different severity of x-ray lung attenuation at CT. To condition box-and-whiskers plots, cohort was divided into two groups and the cut-off values were selected at %LAA2950insp and %LAA2910exp mean values (19.1% and 22.0%, respectively). Both models systematically overestimated %LAA (negative residuals) in patients with milder reduction of x-ray lung attenuation at CT and underestimated %LAA (positive residuals) in patients with greater reduction of x-ray lung attenuation at CT.
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
multivariate regression models were identified: submodel A, which predicts lower levels of reduced x-ray attenuation (%LAA2950insp 21%, %LAA2910exp 26%), and submodel B, which predicts higher levels of reduced x-ray attenuation (%LAA2950insp . 21%, %LAA2910exp . 26%). The %LAA ranges of each submodel were automatically determined to optimize the R2 and the standard error of the two regressions (the two best-fitting models were selected). Among all functional regressors identified with the step-wise process, only FEV1/ VC was selected to predict both %LAA2950insp of 21% or less and %LAAof 26% or less. Conversely, 2910exp Dlco% was the predictor selected for both %LAA2950insp greater than 21% and %LAA2910exp greater than 26%. FRC% was selected together with Dlco% to predict %LAA2950insp greater than 21%. The predicting performances of the multiple models 5
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
Table 4 Multiple Model Estimation to Predict CT Lung Attenuation with Pulmonary Function Indexes Variable, Submodel, and Predictor %LAA2950insp Submodel A FEV1/VC Intercept Submodel B Dlco% FRC% Intercept %LAA2910exp Submodel A FEV1/VC Intercept Submodel B Dlco% Intercept
Coefficient
20.25 23.81 20.12 0.12 20.88
20.51 38.06 20.42 62.87
Note.—Submodel A: %LAA2950insp 21% and %LAA2910exp 26%. Submodel B: %LAA2950insp . 21% and %LAA2910exp . 26%. The threshold values were automatically selected during the learning process to optimize the R2 and the standard error of the two regressions. The predicting performances of the multiple models were much higher than those of the single models (R2 = 0.75 [P , .001] vs 0.46 [P , .001], respectively, for %LAA2950insp and R2 = 0.83 [P , .001] vs 0.63 [P , .001] for %LAA2910exp).
were much higher than those of the single models (R2 = 0.75 vs 0.46, respectively, for %LAA2950insp and R2 = 0.83 vs 0.63 for %LAA2910exp) (Tables 3, 4), which shows that the twofold models are able to cover almost the entire variability range of the original CT data.
Discussion The main results of this study are as follows. First, COPD pulmonary function measurements are not linearly related to CT lung attenuation over the full range of observed attenuations. Second, a multiple model approach that combines measurements of airflow obstruction (FEV1/VC), overinflation (FRC%), and parenchymal destruction (Dlco%) can accurately predict the inspiratory and expiratory %LAA over a wide range of values. These results are in keeping with the 6
notion that COPD is a heterogeneous condition in which inflammatory and destructive changes that affect airways and parenchyma may combine to cause the lung attenuation changes measured with CT. Similar to previous investigators (18–20), we found that linear multivariate analysis only partially improves the predictive ability of each of the single functional variables of obstruction, overinflation, and parenchymal destruction. We also noticed that the relationships between some functional predictors and %LAA were not linear but varied depending on the degree of the CT densitometric alteration. In particular, despite the overall fair Pearson correlation coefficients, for %LAA values compatible with greater parenchymal destruction a very weak association with FEV1/ VC was evidenced, whereas for %LAA values compatible with lower or absent parenchymal destruction no significant association with Dlco% was demonstrated. Accordingly, if we adopted a standard correlation-based single-model approach, we could reach the misleading conclusion that CT parenchymal destruction varies proportionally to the severity of airflow limitation or to the reduction in diffusing capacity, and vice-versa. In our dataset, this is true only in limited ranges of %LAA. This twofold profile heavily affects the performances of traditional multivariate regression models. The analysis of the residuals in our cohort confirms the notion that forcing a single linear model approach in the presence of nonlinear relationships may produce a systematic under- or overestimation of the predicted parameter (27). The complexity of COPD cannot be expressed with a simple measurement of expiratory airflow obstruction (28–32). Even though the percentage of emphysema has been reported to increase with increasing Global Initiative for Chronic Obstructive Lung Disease stages (14,28), there is a considerable amount of discordance unexplained by the current classification approach (14–17). Recent data obtained in a large cohort of patients with COPD
Paoletti et al
demonstrated that, when correlated with %LAA2950insp, spirometric indexes of obstruction (eg, FEV1 and FEV1/ FVC) fail to determine with accuracy the presence and the severity of emphysema as reflected by quantitative CT evaluation (14,17). In this line of evidence, the results of the present study may be of clinical significance. Our results showed that, for high values of %LAA2950insp, the contribution of FEV1/VC was not significant to predict parenchymal destruction, whereas Dlco% accounted for a large amount of that variation with a small contribution by FRC%. Conversely, Dlco% did not significantly predict lower values of %LAA2950insp. Another interesting finding was that, considering %LAA as a parameter of reference, FEV1/VC instead of FEV1/ FVC was the airflow obstruction index invariably selected by the stepwise algorithms. This may point to the known greater accuracy of FEV1/VC with respect to FEV1/FVC in the assessment of airway obstruction in COPD. Recently, Bafadhel et al (33) showed that diffusing capacity has great accuracy in the identification of CT-detected emphysema. In general, the results of our study may indicate that this relationship could be particularly accurate in those patients who are more severely affected by emphysema and confirm that diffusing capacity has a relevant role in predicting the level of parenchymal destruction as assessed with quantitative CT. The most recent large-scale clinical trials considered the measurement of %LAA2950insp at CT as an accurate assessment of the amount of emphysema (34,35). The relationship between pulmonary function data and CT quantitative parameters observed in this study, if confirmed and prospectively validated in larger samples, may be of use to obtain a fairly accurate evaluation of lung attenuation impairment for clinical and pharmacologic studies in which the use of CT might not be feasible or cost effective. Our study has some limitations. The number of patients should be increased to achieve a more reliable
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
learning process in the identification of the specific predicting models. Another intrinsic limitation is the absence of a reference standard that could help verify the real extent of emphysema, inflammation, and other pathophysiologic processes. However, the values of %LAA2950insp in our cohort could be considered to substantially cover the variability in parenchymal destruction reported in similar larger-scale studies. We also included expiratory %LAA2910exp in our analysis. The main limit of this measurement is the lower accuracy of the parameter itself in representing pulmonary emphysema. In fact, both pathologic alterations due to loss of elastic recoil and remodeling of small airways could be potentially reflected in the CT detection of decreased expiratory lung attenuation (9). A further limitation is that the CT data were acquired, as in the majority of similar studies, without control of the level of lung inflation. This could have increased the variation of the relationships found. In summary, this study demonstrated that the relationship between lung function and CT parenchymal destruction in COPD is nonlinear. In particular, only FEV1/VC was found to show significant correlation with mild parenchymal destruction. Dlco% was, conversely, the variable accounting for most of the variation in the presence of more severe parenchymal destruction. To overcome the observed nonlinearity of the relationship, we identified two specific linear models that better fit the actual data distribution over its full range of variability and that may be used in each patient to predict with an acceptable level of accuracy “CT-determined emphysema” in clinical and pharmacologic studies. Disclosures of Conflicts of Interest: M.P. disclosed no relevant relationships. L.C. disclosed no relevant relationships. F.B. disclosed no relevant relationships. G.C. disclosed no relevant relationships. M.P. disclosed no relevant relationships.
References 1. Lynch DA, Al-Qaisi MA. Quantitative computed tomography in chronic obstruc-
tive pulmonary disease. J Thorac Imaging 2013;28(5):284–290. 2. Litmanovich DE, Hartwick K, Silva M, Bankier AA. Multidetector computed tomographic imaging in chronic obstructive pulmonary disease: emphysema and airways assessment. Radiol Clin North Am 2014;52(1):137–154. 3. Hayhurst MD, MacNee W, Flenley DC, et al. Diagnosis of pulmonary emphysema by computerised tomography. Lancet 1984; 2(8398):320–322. 4. Müller NL, Staples CA, Miller RR, Abboud RT. “Density mask:” an objective method to quantitate emphysema using computed tomography. Chest 1988;94(4):782–787. 5. Knudson RJ, Standen JR, Kaltenborn WT, et al. Expiratory computed tomography for assessment of suspected pulmonary emphysema. Chest 1991;99(6):1357– 1366. 6. Miniati M, Filippi E, Falaschi F, et al. Radiologic evaluation of emphysema in patients with chronic obstructive pulmonary disease: chest radiography versus high resolution computed tomography. Am J Respir Crit Care Med 1995;151(5):1359– 1367. 7. Gevenois PA, de Maertelaer V, De Vuyst P, Zanen J, Yernault JC. Comparison of computed density and macroscopic morphometry in pulmonary emphysema. Am J Respir Crit Care Med 1995;152(2):653– 657.
Paoletti et al
13. Xie X, de Jong PA, Oudkerk M, et al. Morphological measurements in computed tomography correlate with airflow obstruction in chronic obstructive pulmonary disease: systematic review and metaanalysis. Eur Radiol 2012;22(10):2085– 2093. 14. Schroeder JD, McKenzie AS, Zach JA, et al. Relationships between airflow obstruction and quantitative CT measurements of emphysema, air trapping, and airways in subjects with and without chronic obstructive pulmonary disease. AJR Am J Roentgenol 2013;201(3):W460–W470. 15. Hogg JC, Wright JL, Wiggs BR, Coxson HO, Opazo Saez A, Paré PD. Lung structure and function in cigarette smokers. Thorax 1994;49(5):473–478. 16. Gelb AF, Hogg JC, Müller NL, et al. Contribution of emphysema and small airways in COPD. Chest 1996;109(2):353–359. 17. Bhatt SP, Sieren JC, Dransfield MT, et al. Comparison of spirometric thresholds in diagnosing smoking-related airflow obstruction. Thorax 2014;69(5):410–415. 18. Cerveri I, Dore R, Corsico A, et al. As sessment of emphysema in COPD: a functional and radiologic study. Chest 2004;125(5):1714–1718. 19. Desai SR, Hansell DM, Walker A, MacDonald SL, Chabat F, Wells AU. Quantification of emphysema: a composite physiologic index derived from CT estimation of disease extent. Eur Radiol 2007;17(4):911–918.
8. Gevenois PA, De Vuyst P, de Maertelaer V, et al. Comparison of computed density and microscopic morphometry in pulmonary emphysema. Am J Respir Crit Care Med 1996;154(1):187–192.
20. Hesselbacher SE, Ross R, Schabath MB, et al. Cross-sectional analysis of the utility of pulmonary function tests in predicting emphysema in ever-smokers. Int J Environ Res Public Health 2011;8(5):1324–1340.
9. Gevenois PA, De Vuyst P, Sy M, et al. Pulmonary emphysema: quantitative CT during expiration. Radiology 1996;199(3):825–829.
21. Karimi R, Tornling G, Forsslund H, et al. Lung density on high resolution computer tomography (HRCT) reflects degree of inflammation in smokers. Respir Res 2014;15:23.
10. 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(2):315–321. 11. Nakano Y, Muro S, Sakai H, et al. Computed tomographic measurements of airway dimensions and emphysema in smokers: correlation with lung function. Am J Respir Crit Care Med 2000;162(3 Pt 1): 1102–1108. 12. Camiciottoli G, Bartolucci M, Maluccio NM, et al. Spirometrically gated high-resolution CT findings in COPD: lung attenuation vs lung function and dyspnea severity. Chest 2006;129(3):558–564.
Radiology: Volume 000: Number 0— 2015 n radiology.rsna.org
22. From the Global Strategy for the Diagno sis, Management and Prevention of COPD, Global Initiative for Chronic Obstructive Lung Disease (GOLD). http://www.goldcopd.org/. Published 2014. Accessed June 26, 2014. 23. Miller MR, Hankinson J, Brusasco V, et al. Standardisation of spirometry. Eur Respir J 2005;26(2):319–338. 24. Wanger J, Clausen JL, Coates A, et al. Standardisation of the measurement of lung volumes. Eur Respir J 2005;26(3):511–522. 25. Macintyre N, Crapo RO, Viegi G, et al. Standardisation of the single-breath deter-
7
THORACIC IMAGING: Pulmonary Function and CT Attenuation in Chronic Obstructive Pulmonary Disease
mination of carbon monoxide uptake in the lung. Eur Respir J 2005;26(4):720–735. 26. Cherkassky V, Ma Y. Multiple model regression estimation. IEEE Trans Neural Netw 2005;16(4):785–798. 27. Anscombe FJ. Graphs in statistical analysis. Am Stat 1973;27(1):17–21. 28. Agusti A, Calverley PM, Celli B, et al. Characterisation of COPD heterogeneity in the ECLIPSE cohort. Respir Res 2010; 11:122. 29. Wedzicha JA. The heterogeneity of chronic obstructive pulmonary disease. Thorax 2000; 55(8):631–632.
8
30. Pistolesi M. Beyond airflow limitation: another look at COPD. Thorax 2009;64(1): 2–4. 31. Han MK, Agusti A, Calverley PM, et al. Chronic obstructive pulmonary disease phenotypes: the future of COPD. Am J Respir Crit Care Med 2010;182(5):598–604. 32. Casanova C, de Torres JP, Aguirre-Jaíme A, et al. The progression of chronic obstructive pulmonary disease is heterogeneous: the experience of the BODE cohort. Am J Respir Crit Care Med 2011;184(9):1015–1021.
Paoletti et al
typing of COPD. Chest 2011;140(3):634– 642. 34. Vestbo J, Anderson W, Coxson HO, et al. Evaluation of COPD longitudinally to identify predictive surrogate end-points (ECLIPSE). Eur Respir J 2008;31(4):869–873. 35. Regan EA, Hokanson JE, Murphy JR, et al. Genetic epidemiology of COPD (COPDGene) study design. COPD 2010;7(1):32–43.
33. Bafadhel M, Umar I, Gupta S, et al. The role of CT scanning in multidimensional pheno-
radiology.rsna.org n Radiology: Volume 000: Number 0— 2015
