CT Density Distribution Analysis in Patients with Cystic Fibrosis: Correlation with Pulmonary Function and Radiologic Scores grugilliers, MSc, Isaure de Lavernhe, MD, Alain Le Blanche, MD, PhD, Lo€ıc De Marie-France Carette, MD, PhD, Sam Bayat, MD, PhD Rationale and Objectives: The progressive changes in lung morphology observed in cystic fibrosis (CF) can potentially affect the statistical distribution of computed tomography (CT) density values. This study aimed to characterize the lung CT density distributions by quantifying indices of the kurtosis and skewness of the lung density distribution and to compare these indices to radiologic scores and lung function parameters in children and young adults with CF. Materials and Methods: CT scans and lung function of 26 patients with CF were retrospectively examined. The Bhalla radiologic scoring was performed separately, in random order, by two expert radiologists, blinded to the patient’s identity, age, clinical status, results of lung function tests, and the other paired observer’s score. Results: Positive relations were evidenced between the log indices of lung density distribution kurtosis (iKurtosis) and the overall radiologic scores (RS) of both observers (R = 0.58; P < .001 vs RS1 and R = 0.71; P < .001 vs RS2). A similar relationship was evidenced with the log index of the degree of distribution asymmetry (iSkewness; R = 0.62; P < .001 vs RS1 and R = 0.62; P < .001 vs RS2). Log-iKurtosis and log-iSkewness were related to FEV1 (R = 0.56; P < 105 and R = 0.55; P < 105) and to residual volume (R = 0.40; P < .001 and R = 0.45; P < .001, respectively). Both radiologic scores showed significant relation with lung function. The correlation between RS1 and RS2 was excellent (R = 0.93), with a Cohen weighted kappa of 0.43. Conclusions: Characteristic indices of lung CT density distribution are correlated to lung function and radiologic scores in patients with CF and merit further evaluation as part of more comprehensive automated methods for quantifying CF lung CT data. Key Words: Cystic fibrosis; computed tomography; lung function; image analysis. ªAUR, 2015
ADVANCES IN KNOWLEDGE
A
utomatically calculated indices used to quantify the sharpness of the lung computed tomography density distribution peak (kurtosis) and the degree of distribution asymmetry (skewness) show significant relations with both lung function and the radiologic score of Bhalla et al. (eg, log kurtosis index: R = 0.56; P < 105 vs 1-second forced expiratory volume; R = 0.58; P < .001 vs total score of
Acad Radiol 2015; 22:179–185 From the Department of Pediatric Pulmonary Medicine, Amiens University Hospital, 1 Place Victor Pauchet, Amiens Cedex 1 80054, France (I.L., L.D., de Formation S.B.); University of Versailles Saint-Quentin-en-Yvelines, Unite decine de Paris-Ile-de-France Ouest Simone Veil et de Recherche de Me Versailles, France (A.L.B.); Department of Diagnostic and Interventional Dubos Hospital, Paris, France (A.L.B.); Gramfc Laboratory Radiology, Rene Inserm U1105, University of Picardie Jules Verne (L.D., S.B.); Department of Radiology, Faculty of Medicine, Pierre and Marie Curie University Paris VI, Tenon University Hospital, Paris, France (M.F.C.). Received May 18, 2014; accepted September 2, 2014. Parts of the data contained in this article were presented at the 2009 American Thoracic Society Annual Meeting. Funding Sources: This study was funded by the European Regional Development Fund and by the Picardie Regional Council (#REG08009); the ERDF and Regional council are a single grant. to S.B. Address correspondence to: S.B. e-mail: [email protected] ªAUR, 2015 http://dx.doi.org/10.1016/j.acra.2014.09.003
observer 1; R = 0.71; P < .001 vs observer 2) in patients with cystic fibrosis. Introduction
Cystic fibrosis (CF), the most frequent autosomal recessive disease in Caucasians, leads to inflammatory changes in the airways and lung parenchyma due to chronic bacterial infection. Life expectancy in this disease remains severely compromised due to respiratory complications. Lung function tests (LFTs) remain the gold standard for monitoring the progression of respiratory disease in children with CF. High-resolution computed tomography (CT) is also used on a periodic basis in some centers to follow-up the progression of structural changes due to CF lung disease. Previous investigations have demonstrated that radiologic scores allowing semiquantitative assessment of various structural lung changes are significantly correlated to lung functional parameters (1–4). Some studies have suggested that changes in CT radiologic scores may be detected earlier than those in LFTs as they provide a more sensitive outcome measurement in patients with CF (5). However, the routine use of CT scoring has some drawbacks as it is time consuming and observer 179
DE LAVERNHE ET AL
dependent (6). Moreover, interobserver agreement is often poor and, at best, moderate (6–8). In the lung parenchyma, CT attenuation is determined by the relative contents of air, tissue, and blood within the image voxel. CF is characterized by structural changes due to chronic inflammation that tend to increase the lung parenchymal attenuation. On the other hand, chronic airway obstruction, closure, and increased gas trapping in this disease tend to decrease attenuation. In the normal lung, CT density histograms show a single mode or peak value at approximately 800 Hounsfield units (HU), with an asymmetrical distribution around this peak value that appears skewed. The progressive structural changes in the lung parenchyma observed in CF can potentially affect the statistical distribution of CT densities. We hypothesize that the changes in the shape of the CT density distribution contain functional relevance in this disease. If this hypothesis is verified, it would offer promising perspectives for rapid, objective, and automated analysis of CT images to provide clinically relevant data. The primary goals of the present study were 1) to characterize the lung CT density value distribution, by quantifying the sharpness of the distribution peak (kurtosis) and the degree of distribution asymmetry (skewness) and 2) to compare these indices with both semiquantitative radiologic scores and lung function parameters in children and young adults with CF. As a secondary goal, the changes in density distribution indices, CT scores, and LFTs over time and with disease progression were retrospectively assessed over a 3-year period. SUBJECTS AND METHODS Study Population
CT examinations and LFTs of 26 patients with CF, which were performed annually as part of the routine medical follow-up from 2006 to 2010 at a university hospital CF center, were retrospectively examined. The diagnosis of CF was based on a pilocarpine iontophoresis sweat chloride > 60 mEq/L and confirmed by CF gene mutation analysis in all patients. Caucasian subjects aged 4–19 years, able to perform reproducible LFTs were included in the study (Table 1). Subjects with a respiratory tract infection within the month preceding CT and LFTs were not eligible for the study. Internal review board approval and the requirement for informed consent were waived for this retrospective study. CT Imaging
The patients were examined in supine position and instructed to inhale maximally and hold their breath until the completion of volumic acquisition of the thorax. Thoracic imaging was performed using a GE multidetector LightSpeed scanner (General Electric Medical Systems, Milwaukee, WI). Contiguous 1.2mm thick sections were obtained at 112 12 kV and 228 91 mA (mean standard deviation [SD]) with a 0.6 seconds rotation time and a field of view of 350 mm. Images were reconstructed with a Detail algorithm and a 512 512 matrix. 180
Academic Radiology, Vol 22, No 2, February 2015
Radiologic Evaluation
Radiologic scoring was performed separately and in random order by two expert radiologists, blinded to the patient’s identity, age, clinical status, LFTresults, and the other paired radiologist’s score. Anonymized images were analyzed using a standard lung +1600 WR 600 WL window protocol. The radiologic score was based on the score of Bhalla et al. (1) that semiquantitatively evaluates the severity and extent of bronchiectasis, bronchial wall thickening, mucous plugging, atelectasis, bullae, consolidation, sacculations or abscesses, septal thickening, and air trapping. This latter criterion was termed ‘‘emphysema’’ in the original description of the scoring system (1). CT Density Distribution Analysis
The lung parenchyma was automatically segmented by thresholding regions of interest with a CT density of 250 to 1024 HU. The frequency distributions of lung density were fit with a lognormal function in each image, using Nelder–Mead minimization. The goodness-of-fit parameter was assessed using the coefficient of determination (R2). Fits with an R2 < 0.96 were rejected. An index of sharpness of the histogram peak (iKurtosis) was calculated as follows (9): . 1
2
2
2
e4s þ 2e3s þ 3e2s 6
where s is the SD extracted from the fitted function. An index of the degree of distribution asymmetry (iSkewness) was computed as (9) follows: .h 1
es þ 2 2
pffiffiffiffiffiffiffiffiffiffiffiffiffiffii es2 1
The indices represent the reciprocal values of kurtosis and skewness, calculated as such for the values to change in the same direction as the radiologic scores. Both indices were calculated for each image of a given study and averaged over all image slices. Lung Function
The LFTs were performed on the same day as the CTexaminations in 49 of 94 occasions. The mean time interval between CT and LFT was 2.3 8.4 days, and both tests were performed within 1 month in all patients. Lung function testing was performed using a Jaeger MasterLab system (CareFusion, San Diego, CA), according to the American Thoracic Society/European Respiratory Society guidelines for acceptability and reproducibility. The forced expiratory volume in 1 second (FEV1), forced vital capacity (FVC), FEV1/FVC ratio, peak expiratory flow (PEF), forced expiratory flow between 25% and 75% of expiratory vital capacity (FEF25–75%), airway resistance (Raw), residual volume (RV), and total lung capacity (TLC) were recorded and expressed as percentage of predicted values (10). Body plethysmography was performed in 78 of 94
Academic Radiology, Vol 22, No 2, February 2015
CT DENSITY DISTRIBUTION ANALYSIS IN CF
TABLE 1. Patient Characteristics, Mean Values of Lung Function, and Radiologic Parameters and Scores
age (mean), years range (minimum–maximum) N Gender (M/F) Time from CT1, years FEV1 (% pred) FVC (% pred) FEV1/FVC (% pred) PEF (% pred) FEF25–75 (% pred) Raw (% pred) RV (% pred) TLC (% pred) TcSaO2 (%) RS1 RS2 MLA (HU) iSkewness iKurtosis
CT1
CT2
CT3
CT4
P Value
9.9 4.0–16.5 19 (7/14) 0 88.3 28.7 89.6 24.0 98.1 10.3 89.8 22.6 77.7 38.1 178.4 63.7 123.0 60.4 99.3 23.1 97.0 1.6 6.8 5.4 7.8 6.2 787.7 95.7 0.79 0.19 1.21 0.67
10.6 5.0–17.5 26 (11/15) 1.0 0.2 88.9 28.1 91.3 23.5 95.9 10.8 86.4 23.3 74.7 40.8 225.3 89.1 154.3 62.1 101.2 18.9 97.0 1.3 7.2 6.7 8.0 6.6 808.4 70.8 0.83 0.27 1.42 1.37
12.3 6.6–18.5 23 (11/13) 1.9 0.4 83.9 27.1 90.6 21.9 91.7 12.1 82.4 30.1 64.3 42.3 232.0 123.9 164.7 61.4 99.9 20.3 97.6 1.5 7.8 6.63 8.8 7.03 818.6 57.8 0.83 0.26 1.41 1.27
12.9 6.9–19.5 23 (10/13) 2.9 0.4 81.8 26.8 89.4 22.5 89.7 13.8 82.4 27.6 63.2 38.1 230.2 119.8 137.8 64.0 98.4 15.3 97.6 1.3 8.2 6.5 9.2 6.67 809.2 71.4 0.79 0.18 1.20 0.66
.040 .869 .011 .253 .003 .005 .001 .714 .062 .036 .030 .326 .403 .135
CT, computed tomography; FEF25–75, mean forced expiratory flow between 25% and 75% of forced vital capacity; FEV1, forced expiratory volume in 1 second (expressed as percent predicted value [% pred]); FVC, forced vital capacity; HU, Hounsfield units; iKurtosis, index of the lung density distribution kurtosis; iSkewness, index of the lung density distribution skewness; MLA, mean lung attenuation in Hounsfield units; PEF, peak expiratory flow; Raw, airway resistance; RS1, RS2, total radiologic scores provided by each observer; RV, residual volume; TcSaO2, transcutaneous oxygen saturation; TLC, total lung capacity. Data are mean standard deviation.
tests. Transcutaneous oxygen saturation (TcSaO2) was measured using a pulse oximeter (Nellcor N65; Mansfield, MA). Statistical Analysis
Data are expressed as mean SD. Mean values were compared using the Friedman one-way repeated-measures analysis of variance on ranks, with the Student–Neuman–Keuls procedure for multiple pairwise comparisons. The Pearson correlation coefficient was used to assess the relations between total radiologic score, log-transformed indices of kurtosis and skewness of the lung density distribution, and LFT parameters. Univariate linear regression analysis was used to assess the relation between total radiologic scores of each observer (RS1, RS2), iSkewness, and iKurtosis and between individual radiologic scores criteria and FEV1. Interobserver agreement of the radiologic scores was evaluated by Cohen weighted kappa, linear regression and the Bland–Altman plot (11). Statistical analysis was performed using the Sigmaplot software (Version 11; Systat Software Inc, Chicago, IL). A P < .05 was considered significant for all statistical tests. RESULTS Study Population
Patient characteristics mean values of lung function parameters, total radiologic scores, and indices describing the shape
of the density distribution (iSkewness and iKurtosis) are summarized in Table 1. Of the 26 patients enrolled in the study, both CT and LFT results were obtained on all four time points in 19 patients. Lung Function, Radiologic Score, and Indices of Lung Density Distribution Skewness and Kurtosis
On average, lung functional compromise was mild, with significant variability among subjects (Table 1). The FEV1, FEV1/FVC, FEF25–75, Raw, and RV significantly changed over the study period. Conversely, FVC, TLC, and TcSaO2 did not show significant changes. Total radiologic scores also showed a small but significant increase over time. However, mean lung attenuation and indices of skewness and kurtosis of the lung density distribution did not change significantly over the studied time interval. Figure 1 shows examples of density distribution analysis in two patients with radiologic scores of 0 and 23 and 21 for observers 1 and 2, respectively. Correlations between total radiologic scores from both observers, indices of skewness and kurtosis, and functional parameters are shown in Table 2. Total radiologic scores were correlated to lung functional parameters. The strongest negative correlations were observed with FEV1, FVC, and FEV1/FVC ratio. The total radiologic scores were also positively correlated to RV, a parameter that reflects air trapping, although these correlations were not as strong. Log-transformed indices of
181
DE LAVERNHE ET AL
Academic Radiology, Vol 22, No 2, February 2015
Figure 1. Sample computed tomography (CT) images without contrast injection in two patients with minimal (upper row) and pronounced (lower row) morphologic changes based on radiologic scores. (a) Original CT images, (b) the segmented lung based on grey level threshold, and (c) the raw density distributions (grey curve) and lognormal fits (black curve) with the corresponding radiologic scores of both observers (RS1 and RS2) and the calculated indices of skewness (iSkewness) and kurtosis (iKurtosis). HU, Hounsfield units; RS, radiologic scores.
skewness and kurtosis of the lung density distribution also showed negative correlations with lung functional parameters altered by airway obstruction such as FEV1, FVC, and FEV1/FVC ratio. Similarly, significant positive correlations were found between these indices and RV. The absolute values of the 95% confidence intervals of the correlation coefficients were not significantly different between density distribution–derived indices versus lung function parameters, and radiologic scores versus lung function parameters: P = .635 and .288 for log-skewness versus RS1 and RS2, and P = .913 and .432 for log-kurtosis versus RS1 and RS2, respectively by the Student paired t test. The relation between total radiologic scores and logtransformed indices of skewness and kurtosis are shown in Figure 2. Log-iKurtosis showed a significant relation with the total radiologic score of both observers: R = 0.58; P < .001 and R = 0.71; P < .001versus RS1 and RS2, respectively. Similarly, iSkewness was significantly correlated to both radiologic scores: R = 0.62; P < .001 and R = 0.62; P < .001versus RS1 and RS2, respectively. Interobserver Agreement
A strong correlation was found between both radiologic scores, with R = 0.93 and P < .001 (Fig 3). The Bland–Altman plot of the radiologic scores is presented in Figure 3 and shows the limits of agreement between the observers. The average bias was small (0.94 2.50). However, Cohen weighted kappa showed only moderate agreement between the two observers for the total radiologic score (Table 3). This parameter showed significant variability between individual score criteria. For example, agreement was almost
182
perfect for the generations of bronchial divisions involved, whereas it was fair for bronchial wall thickening. DISCUSSION This study aimed at determining whether lung CT density distributions contain clinically relevant information in patients with CF. The main finding of this study is that automatically calculated indices quantifying the sharpness of the lung CT density distribution peak (kurtosis) and the degree of distribution asymmetry (skewness) showed significant relations with both lung function and the radiologic score of Bhalla et al. in patients with CF. Several studies have previously shown the significant relationship between lung morphologic changes quantified using radiologic scores, with lung function (1–4). These data are in agreement with the findings of the present study, which also find significant correlations between overall radiologic scores and lung function. Longitudinal analysis of CT data has shown progression of bronchiectasis even in the presence of stable lung function (5,12). This may be due to the fact that LFT assesses global parameters at the airway opening which are dependent on patient cooperation and less sensitive to focal structural changes. Several radiologic scores have been proposed to quantitatively assess the changes in lung morphology in CF (1,2,13,14). Although these have shown that CT abnormalities can be assessed in a reproducible fashion, radiologic scores are observer dependent, and perfect interobserver agreement is rarely found (15). Furthermore, proper radiologic scoring of lung CT is time consuming, which sets a limitation for the study of large data sets.
FEV1
FEV1/FVC
PEF
FEF25–75
RV
TLC
TcSaO2
0.56 (0.69 to 0.39)
ADVANCES IN KNOWLEDGE
A
utomatically calculated indices used to quantify the sharpness of the lung computed tomography density distribution peak (kurtosis) and the degree of distribution asymmetry (skewness) show significant relations with both lung function and the radiologic score of Bhalla et al. (eg, log kurtosis index: R = 0.56; P < 105 vs 1-second forced expiratory volume; R = 0.58; P < .001 vs total score of
Acad Radiol 2015; 22:179–185 From the Department of Pediatric Pulmonary Medicine, Amiens University Hospital, 1 Place Victor Pauchet, Amiens Cedex 1 80054, France (I.L., L.D., de Formation S.B.); University of Versailles Saint-Quentin-en-Yvelines, Unite decine de Paris-Ile-de-France Ouest Simone Veil et de Recherche de Me Versailles, France (A.L.B.); Department of Diagnostic and Interventional Dubos Hospital, Paris, France (A.L.B.); Gramfc Laboratory Radiology, Rene Inserm U1105, University of Picardie Jules Verne (L.D., S.B.); Department of Radiology, Faculty of Medicine, Pierre and Marie Curie University Paris VI, Tenon University Hospital, Paris, France (M.F.C.). Received May 18, 2014; accepted September 2, 2014. Parts of the data contained in this article were presented at the 2009 American Thoracic Society Annual Meeting. Funding Sources: This study was funded by the European Regional Development Fund and by the Picardie Regional Council (#REG08009); the ERDF and Regional council are a single grant. to S.B. Address correspondence to: S.B. e-mail: [email protected] ªAUR, 2015 http://dx.doi.org/10.1016/j.acra.2014.09.003
observer 1; R = 0.71; P < .001 vs observer 2) in patients with cystic fibrosis. Introduction
Cystic fibrosis (CF), the most frequent autosomal recessive disease in Caucasians, leads to inflammatory changes in the airways and lung parenchyma due to chronic bacterial infection. Life expectancy in this disease remains severely compromised due to respiratory complications. Lung function tests (LFTs) remain the gold standard for monitoring the progression of respiratory disease in children with CF. High-resolution computed tomography (CT) is also used on a periodic basis in some centers to follow-up the progression of structural changes due to CF lung disease. Previous investigations have demonstrated that radiologic scores allowing semiquantitative assessment of various structural lung changes are significantly correlated to lung functional parameters (1–4). Some studies have suggested that changes in CT radiologic scores may be detected earlier than those in LFTs as they provide a more sensitive outcome measurement in patients with CF (5). However, the routine use of CT scoring has some drawbacks as it is time consuming and observer 179
DE LAVERNHE ET AL
dependent (6). Moreover, interobserver agreement is often poor and, at best, moderate (6–8). In the lung parenchyma, CT attenuation is determined by the relative contents of air, tissue, and blood within the image voxel. CF is characterized by structural changes due to chronic inflammation that tend to increase the lung parenchymal attenuation. On the other hand, chronic airway obstruction, closure, and increased gas trapping in this disease tend to decrease attenuation. In the normal lung, CT density histograms show a single mode or peak value at approximately 800 Hounsfield units (HU), with an asymmetrical distribution around this peak value that appears skewed. The progressive structural changes in the lung parenchyma observed in CF can potentially affect the statistical distribution of CT densities. We hypothesize that the changes in the shape of the CT density distribution contain functional relevance in this disease. If this hypothesis is verified, it would offer promising perspectives for rapid, objective, and automated analysis of CT images to provide clinically relevant data. The primary goals of the present study were 1) to characterize the lung CT density value distribution, by quantifying the sharpness of the distribution peak (kurtosis) and the degree of distribution asymmetry (skewness) and 2) to compare these indices with both semiquantitative radiologic scores and lung function parameters in children and young adults with CF. As a secondary goal, the changes in density distribution indices, CT scores, and LFTs over time and with disease progression were retrospectively assessed over a 3-year period. SUBJECTS AND METHODS Study Population
CT examinations and LFTs of 26 patients with CF, which were performed annually as part of the routine medical follow-up from 2006 to 2010 at a university hospital CF center, were retrospectively examined. The diagnosis of CF was based on a pilocarpine iontophoresis sweat chloride > 60 mEq/L and confirmed by CF gene mutation analysis in all patients. Caucasian subjects aged 4–19 years, able to perform reproducible LFTs were included in the study (Table 1). Subjects with a respiratory tract infection within the month preceding CT and LFTs were not eligible for the study. Internal review board approval and the requirement for informed consent were waived for this retrospective study. CT Imaging
The patients were examined in supine position and instructed to inhale maximally and hold their breath until the completion of volumic acquisition of the thorax. Thoracic imaging was performed using a GE multidetector LightSpeed scanner (General Electric Medical Systems, Milwaukee, WI). Contiguous 1.2mm thick sections were obtained at 112 12 kV and 228 91 mA (mean standard deviation [SD]) with a 0.6 seconds rotation time and a field of view of 350 mm. Images were reconstructed with a Detail algorithm and a 512 512 matrix. 180
Academic Radiology, Vol 22, No 2, February 2015
Radiologic Evaluation
Radiologic scoring was performed separately and in random order by two expert radiologists, blinded to the patient’s identity, age, clinical status, LFTresults, and the other paired radiologist’s score. Anonymized images were analyzed using a standard lung +1600 WR 600 WL window protocol. The radiologic score was based on the score of Bhalla et al. (1) that semiquantitatively evaluates the severity and extent of bronchiectasis, bronchial wall thickening, mucous plugging, atelectasis, bullae, consolidation, sacculations or abscesses, septal thickening, and air trapping. This latter criterion was termed ‘‘emphysema’’ in the original description of the scoring system (1). CT Density Distribution Analysis
The lung parenchyma was automatically segmented by thresholding regions of interest with a CT density of 250 to 1024 HU. The frequency distributions of lung density were fit with a lognormal function in each image, using Nelder–Mead minimization. The goodness-of-fit parameter was assessed using the coefficient of determination (R2). Fits with an R2 < 0.96 were rejected. An index of sharpness of the histogram peak (iKurtosis) was calculated as follows (9): . 1
2
2
2
e4s þ 2e3s þ 3e2s 6
where s is the SD extracted from the fitted function. An index of the degree of distribution asymmetry (iSkewness) was computed as (9) follows: .h 1
es þ 2 2
pffiffiffiffiffiffiffiffiffiffiffiffiffiffii es2 1
The indices represent the reciprocal values of kurtosis and skewness, calculated as such for the values to change in the same direction as the radiologic scores. Both indices were calculated for each image of a given study and averaged over all image slices. Lung Function
The LFTs were performed on the same day as the CTexaminations in 49 of 94 occasions. The mean time interval between CT and LFT was 2.3 8.4 days, and both tests were performed within 1 month in all patients. Lung function testing was performed using a Jaeger MasterLab system (CareFusion, San Diego, CA), according to the American Thoracic Society/European Respiratory Society guidelines for acceptability and reproducibility. The forced expiratory volume in 1 second (FEV1), forced vital capacity (FVC), FEV1/FVC ratio, peak expiratory flow (PEF), forced expiratory flow between 25% and 75% of expiratory vital capacity (FEF25–75%), airway resistance (Raw), residual volume (RV), and total lung capacity (TLC) were recorded and expressed as percentage of predicted values (10). Body plethysmography was performed in 78 of 94
Academic Radiology, Vol 22, No 2, February 2015
CT DENSITY DISTRIBUTION ANALYSIS IN CF
TABLE 1. Patient Characteristics, Mean Values of Lung Function, and Radiologic Parameters and Scores
age (mean), years range (minimum–maximum) N Gender (M/F) Time from CT1, years FEV1 (% pred) FVC (% pred) FEV1/FVC (% pred) PEF (% pred) FEF25–75 (% pred) Raw (% pred) RV (% pred) TLC (% pred) TcSaO2 (%) RS1 RS2 MLA (HU) iSkewness iKurtosis
CT1
CT2
CT3
CT4
P Value
9.9 4.0–16.5 19 (7/14) 0 88.3 28.7 89.6 24.0 98.1 10.3 89.8 22.6 77.7 38.1 178.4 63.7 123.0 60.4 99.3 23.1 97.0 1.6 6.8 5.4 7.8 6.2 787.7 95.7 0.79 0.19 1.21 0.67
10.6 5.0–17.5 26 (11/15) 1.0 0.2 88.9 28.1 91.3 23.5 95.9 10.8 86.4 23.3 74.7 40.8 225.3 89.1 154.3 62.1 101.2 18.9 97.0 1.3 7.2 6.7 8.0 6.6 808.4 70.8 0.83 0.27 1.42 1.37
12.3 6.6–18.5 23 (11/13) 1.9 0.4 83.9 27.1 90.6 21.9 91.7 12.1 82.4 30.1 64.3 42.3 232.0 123.9 164.7 61.4 99.9 20.3 97.6 1.5 7.8 6.63 8.8 7.03 818.6 57.8 0.83 0.26 1.41 1.27
12.9 6.9–19.5 23 (10/13) 2.9 0.4 81.8 26.8 89.4 22.5 89.7 13.8 82.4 27.6 63.2 38.1 230.2 119.8 137.8 64.0 98.4 15.3 97.6 1.3 8.2 6.5 9.2 6.67 809.2 71.4 0.79 0.18 1.20 0.66
.040 .869 .011 .253 .003 .005 .001 .714 .062 .036 .030 .326 .403 .135
CT, computed tomography; FEF25–75, mean forced expiratory flow between 25% and 75% of forced vital capacity; FEV1, forced expiratory volume in 1 second (expressed as percent predicted value [% pred]); FVC, forced vital capacity; HU, Hounsfield units; iKurtosis, index of the lung density distribution kurtosis; iSkewness, index of the lung density distribution skewness; MLA, mean lung attenuation in Hounsfield units; PEF, peak expiratory flow; Raw, airway resistance; RS1, RS2, total radiologic scores provided by each observer; RV, residual volume; TcSaO2, transcutaneous oxygen saturation; TLC, total lung capacity. Data are mean standard deviation.
tests. Transcutaneous oxygen saturation (TcSaO2) was measured using a pulse oximeter (Nellcor N65; Mansfield, MA). Statistical Analysis
Data are expressed as mean SD. Mean values were compared using the Friedman one-way repeated-measures analysis of variance on ranks, with the Student–Neuman–Keuls procedure for multiple pairwise comparisons. The Pearson correlation coefficient was used to assess the relations between total radiologic score, log-transformed indices of kurtosis and skewness of the lung density distribution, and LFT parameters. Univariate linear regression analysis was used to assess the relation between total radiologic scores of each observer (RS1, RS2), iSkewness, and iKurtosis and between individual radiologic scores criteria and FEV1. Interobserver agreement of the radiologic scores was evaluated by Cohen weighted kappa, linear regression and the Bland–Altman plot (11). Statistical analysis was performed using the Sigmaplot software (Version 11; Systat Software Inc, Chicago, IL). A P < .05 was considered significant for all statistical tests. RESULTS Study Population
Patient characteristics mean values of lung function parameters, total radiologic scores, and indices describing the shape
of the density distribution (iSkewness and iKurtosis) are summarized in Table 1. Of the 26 patients enrolled in the study, both CT and LFT results were obtained on all four time points in 19 patients. Lung Function, Radiologic Score, and Indices of Lung Density Distribution Skewness and Kurtosis
On average, lung functional compromise was mild, with significant variability among subjects (Table 1). The FEV1, FEV1/FVC, FEF25–75, Raw, and RV significantly changed over the study period. Conversely, FVC, TLC, and TcSaO2 did not show significant changes. Total radiologic scores also showed a small but significant increase over time. However, mean lung attenuation and indices of skewness and kurtosis of the lung density distribution did not change significantly over the studied time interval. Figure 1 shows examples of density distribution analysis in two patients with radiologic scores of 0 and 23 and 21 for observers 1 and 2, respectively. Correlations between total radiologic scores from both observers, indices of skewness and kurtosis, and functional parameters are shown in Table 2. Total radiologic scores were correlated to lung functional parameters. The strongest negative correlations were observed with FEV1, FVC, and FEV1/FVC ratio. The total radiologic scores were also positively correlated to RV, a parameter that reflects air trapping, although these correlations were not as strong. Log-transformed indices of
181
DE LAVERNHE ET AL
Academic Radiology, Vol 22, No 2, February 2015
Figure 1. Sample computed tomography (CT) images without contrast injection in two patients with minimal (upper row) and pronounced (lower row) morphologic changes based on radiologic scores. (a) Original CT images, (b) the segmented lung based on grey level threshold, and (c) the raw density distributions (grey curve) and lognormal fits (black curve) with the corresponding radiologic scores of both observers (RS1 and RS2) and the calculated indices of skewness (iSkewness) and kurtosis (iKurtosis). HU, Hounsfield units; RS, radiologic scores.
skewness and kurtosis of the lung density distribution also showed negative correlations with lung functional parameters altered by airway obstruction such as FEV1, FVC, and FEV1/FVC ratio. Similarly, significant positive correlations were found between these indices and RV. The absolute values of the 95% confidence intervals of the correlation coefficients were not significantly different between density distribution–derived indices versus lung function parameters, and radiologic scores versus lung function parameters: P = .635 and .288 for log-skewness versus RS1 and RS2, and P = .913 and .432 for log-kurtosis versus RS1 and RS2, respectively by the Student paired t test. The relation between total radiologic scores and logtransformed indices of skewness and kurtosis are shown in Figure 2. Log-iKurtosis showed a significant relation with the total radiologic score of both observers: R = 0.58; P < .001 and R = 0.71; P < .001versus RS1 and RS2, respectively. Similarly, iSkewness was significantly correlated to both radiologic scores: R = 0.62; P < .001 and R = 0.62; P < .001versus RS1 and RS2, respectively. Interobserver Agreement
A strong correlation was found between both radiologic scores, with R = 0.93 and P < .001 (Fig 3). The Bland–Altman plot of the radiologic scores is presented in Figure 3 and shows the limits of agreement between the observers. The average bias was small (0.94 2.50). However, Cohen weighted kappa showed only moderate agreement between the two observers for the total radiologic score (Table 3). This parameter showed significant variability between individual score criteria. For example, agreement was almost
182
perfect for the generations of bronchial divisions involved, whereas it was fair for bronchial wall thickening. DISCUSSION This study aimed at determining whether lung CT density distributions contain clinically relevant information in patients with CF. The main finding of this study is that automatically calculated indices quantifying the sharpness of the lung CT density distribution peak (kurtosis) and the degree of distribution asymmetry (skewness) showed significant relations with both lung function and the radiologic score of Bhalla et al. in patients with CF. Several studies have previously shown the significant relationship between lung morphologic changes quantified using radiologic scores, with lung function (1–4). These data are in agreement with the findings of the present study, which also find significant correlations between overall radiologic scores and lung function. Longitudinal analysis of CT data has shown progression of bronchiectasis even in the presence of stable lung function (5,12). This may be due to the fact that LFT assesses global parameters at the airway opening which are dependent on patient cooperation and less sensitive to focal structural changes. Several radiologic scores have been proposed to quantitatively assess the changes in lung morphology in CF (1,2,13,14). Although these have shown that CT abnormalities can be assessed in a reproducible fashion, radiologic scores are observer dependent, and perfect interobserver agreement is rarely found (15). Furthermore, proper radiologic scoring of lung CT is time consuming, which sets a limitation for the study of large data sets.
FEV1
FEV1/FVC
PEF
FEF25–75
RV
TLC
TcSaO2
0.56 (0.69 to 0.39)
