Difference of the progression of pulmonary cysts assessed by computed tomography among COPD, lymphangioleiomyomatosis, and Birt-Hogg-Dubé syndrome

Many groups developed the methods to quantitatively analyze low attenuation area (LAA) on chest CT in patients with cystic lung diseases. Especially in COPD, it was reported that the cumulative size distribution of LAA clusters follows a power law characterized by the exponent D, which reflect the fractal dimension of terminal airspace geometry. We hypoyhesized that the quantitative charateristics of LAA clusters including fractal property might indicate the different features of the progression of cysts in cystic lung diseases. The aim of this study was to apply the CT image-based method of characterizing the size distribution of LAA clusters for lymphangioleiomyomatosis (LAM) and Birt-Hogg-Dubé syndrome (BHDS) to disclose their features of the progression of pulmonary cysts. 40 patients with COPD, 52 patients with LAM, and 18 patients with BHDS who had undergone CT scans at our institute between January 2002 and August 2009 were included. Differences among these diseases in the quantitative characteristics of LAA clusters {i.e., extent, number, size, fractal property, and the relationship between these quantitatives} were assessed. The Chi-sqsuare test, unpaired t-test, and one-way analyses of variance with Tukey post-hoc tests were used to compare groups, spline model with an interaction terms were used to assess the relationship between extent and number, and exponential regression model was used to assess the relationship between extent and size. Statistically significant differences separated the three diseases in extent and number (P < 0.001). Number was significantly correlated with extent in COPD (P < 0.001), but was not so in LAM and BHDS when extent exceeded 11.5% and 20.8%, respectively. Size was significantly correlated with extent in COPD and LAM (P < 0.001), but was not so in BHDS. The percentage of CT images with fractal property was higher in COPD than that in LAM and BHDS (95.8%, 92.9% and 63.0%, respectively). In conclusion, our study has demonstrated for the first time the different characteristics of the size distribution of LAA clusters among COPD, LAM and BHDS, and indicated that this method is useful for exploration of the pathophysiology in cystic lung diseases.

Introduction Both emphysema and pulmonary cysts are recognized as low attenuation areas (LAA) on chest computed tomography (CT). Recently, many groups developed the methods to quantitatively analyze LAA on chest CT in patients with cystic lung diseases {i.e., chronic obstructive pulmonary disease (COPD), lymphangioleiomyomatosis (LAM) and Birt-Hogg-Dubé syndrome (BHDS)}, and they reported the correlation between those quantitatives and pulmonary function or symptoms [1][2][3].
Especially in COPD, Mishima et al. analyzed the numbers and sizes of LAA clusters and found that the cumulative size distribution of LAA clusters follows a power law characterized by the exponent D [4]. The values of D reflect the fractal dimension of terminal airspace geometry and sensitively detect alveolar tissue destruction. Thereafter, the simulation analysis by Suki et. al. and the longitudinal CT study by Tanabe et. al. revealed that the fractal property of LAA clusters on chest CT in COPD was explained by applying a model of mechanical forcebased destruction in emphysema development [5][6][7]. Accordingly, the analysis of size distribution of LAA clusters on chest CT is thought to be a very useful method for consideration of pathogenesis and pathophysiologies in COPD.
To our knowledge, there is no study that examined the characteristics of size distribution of LAA clusters on chest CT in cystic lung diseases other than COPD, in which the pathogenesis of pulmonary cysts are distinctly different from that of COPD. We hypoyhesized that the quantitative charateristics of LAA clusters including fractal property would be quite different among cystic lung diseases, and might indicate their features of the progression of pulmonary cysts. Therefore, the purpose of the present study is to apply this CT image-based method for LAM and BHDS to disclose their features of the progression of pulmonary cysts.

Study population
This retrospective study was approved by the ethics committee of Juntendo Hospital (JIRB21-134), and the requirement to obtain informed consent was waived. This study included consecutive patients with COPD, LAM, and BHDS who had undergone thin-section CT scans at our institute between January 2002 and August 2009. COPD was diagnosed according to the criteria of the Global Initiative for Obstructive Lung Disease Workshop Report [8]. LAM was diagnosed from tissue biopsies or characteristic clinical pictures (recurrent pneumothorax and/or chylous pleural effusion) and CT findings (diffusely scattered thin-walled pulmonary cysts). Tuberous sclerosis complex-associated LAM (TSC-LAM) was diagnosed by means of established clinical criteria [9]. The diagnosis of BHDS was obtained with the FLCN gene mutation analysis [10]. Patients who had pneumothorax or pleural effusion were excluded.
In COPD, consecutive 40 patients were included {38 male with mean age of 66.7 years (age range, 43-81 years), and 2 female with mean age of 74.0 years (age range, 73-75)} without exclusions of patients. In LAM, after the exclusion of eight patients with pneumothorax and six patients with hydrothorax, 52 patients were finally enrolled {all female with mean age of

Thin-section CT techniques
All patients underwent thin-section CT with an 8-detector row CT scanner or 64-detector row CT scanner (Aquilion 8 or Aquilion 64; Toshiba Medical, Tokyo, Japan) with a 2-mm slice thickness and scanning parameters of 120 kVp, 150 mAs and a field of view of 320 mm. No contrast media were used. During the scan, the patients held their breath after a deep inspiration in the supine position. Each CT image was composed of a 512 × 512 matrix of numeric data (CT numbers) in Hounsfield units (HU) reconstructed using a lung algorithm (FC83). The area of 1 pixel is 0.39 mm 2 .

Analysis of LAA clusters
Three slices from each patient were analyzed: the upper slice taken 1cm above the upper margin of the aortic arch; the middle slice taken 1cm below the carina; and the lower slice taken 1cm above the top of the diaphragm. Each lung zone were thought to have different degrees of disease progression, therefore, all of the derived values from each lung slice were used for the evaluation as an independent value representing disease state (i.e., three data were obtained from one patient). The percentage of LAA (LAA%) was calculated automatically according to the method reported previously (1,4). We defined lung fields as areas with CT numbers less than −200 HU, whereas the cut-off level between LAA and normal lung density was −960 HU (Fig 1) [11]. The number and mean size of LAA clusters were quantitatively assesed automaltically (LAA_n and LAA_s, respectively). Fractal analysis. Fractals are self-similar structures characterized by power-law functions and noninteger dimensions (fractal dimension) (PNAS Sato). To assess the fractal properties of the distributions of LAA sizes, first, cumulative frequency distributions of the LAA_s in each CT image were plotted with a logarithmic scale (log-log plot). Fig 2 demonstrates that the cumulative frequency distributions of LAA_s decrease on log-log plots. The slopes of plots vary between subjects and as a function of LAA%. Next, power law functions were created to fit the slopes of the plots: Y = K Ã X -D (Y, the cumulative frequency distribution; X, LAA_s). Values of coefficient of determination (r 2 ) were taken to indicate the goodness-of-fit of the power law functions [4], and it was defined as having fractal property with an r 2 value of 0.9 or more.
The differences among the three diseases about quantitative characteristics of LAA clusters, characteristics of relationships between LAA% and LAA_n or LAA_s, and fractal property (r 2 ) were examined. The relationship between the LAA_n and LAA% was modeled by regression techniques using a spline model with an interaction terms [12]. First, LAA_n was set to objective variable, and LAA% and disease group (i.e., LAM, BHDS, or COPD) were set to explaining variables. Next, the number and location of knots were determined objectively, and Akaike's Information Criterion (AIC) was used to assess the goodness of model fitness. Therefore, four knots were made at 20, 40, 60, and 80 percentiles of LAA%, and five intervals of LAA% were made as follows; LAA% between 0% and 4.2% (interval I), 4.2 and 11.5% (interval II), 11.5% and 20.8% (interval III), 20.8% and 34.2% (interval IV), and LAA% of 34.2% or more (interval V).

Statistical analysis
The Chi-square test was used for qualitative variables and one-way analyses of variance with Tukey post-hoc tests were used for quantitative variables to compare groups. The unpaired ttest was used to compare two groups. All statistical analyses were performed using SPSS software (version 16.0, SPSS Inc., Chicago). Data were expressed as mean ± SD, and values of p < 0.05 were regarded as statistically significant.

Patients characteristics
Patients characteristics are presented in Table 1. There were statistical significant differences among the three groups in sex ratio, age, height and smoking history.

Analysis of LAA clusters
The results of the quantitative analyses of LAA clusters (LAA%, LAA_n, and LAA_s) are shown in Table 2. Significant differences separated the three groups in LAA% and LAA_n. Among the three diseases, BHDS had the lowest value in LAA% and LAA_s, while COPD had the highest value in LAA%, LAA_n, and LAA_s in the upper lung field. Fig 3 shows the relationship between LAA_n and LAA% modeled by regression techniques using a spline model in each disease. In COPD, LAA_n increased once and thereafter decreased with the increase of  LAA%. In all intervals, LAA_n-LAA% relationships were statistically significant. In LAM, LAA_n increased with the increase of LAA% from 0% to 11.5%, and thereafter LAA_n-LAA% relationship was not statistically significant. In BHDS, LAA_n increased with the increase of LAA% from 0% to 20.8%, and thereafter LAA_n-LAA% relationship was not statistically significant because only three images had LAA% higher than 20.8%. Table 3 shows the estimated slope of the correlation between LAA_n and LAA% in each interval of LAA% from linear regression with spline models. The comparison of LAA_n among five LAA%-intervals in each disease (Fig 4) can confirm the tendency of LAA_n-LAA% relationships demonstrated in Fig 3.  Fig 5 shows the relationship between LAA_s and LAA% in each disease. In COPD and LAM, LAA_s can be expressed as an exponential function of LAA%, and the coefficient of determination was higher in COPD than that in LAM (0.9435 and 0.7727, respectively). In BHDS, there was no significant relationship between LAA_s and LAA%. The comparison of LAA_s among five LAA%-intervals in each disease (Fig 6) can confirm the tendency of LAA_s-LAA% relationships demonstrated in Fig 5. Finally, based on the previous report [3], we categorized the size of LAA cluster into three groups as follows: small (< 10pixels = 3.9 mm 2 ), medium (10 -100pixels = 3.9-39 mm 2 ), and large (> 100 pixels = 39 mm 2 ), and calculated the percentage of each size-catergorized LAA_n in the entire LAA_n of each CT image. In LAM, the increase of medium-sized LAA_n percentage stands out more than other diseases (online supplementary S1 and S2 Figs for further details).  significantly described by a power law of the size of LAA clusters, compared with LAM and BHDS. The mean values of coefficient of determination (r 2 ) in COPD, LAM, and BHDS were 0.975, 0.944, and 0.912, respectively. The r 2 value in COPD was significantly higher than that in LAM and BHDS (p < 0.001), and the r 2 value in LAM was higher than that in BHDS (p < 0.001). Fig 7 shows the comparions of r 2 value among five LAA%-intervals in each disease. In COPD, interval I was lower than II, III and IV, and there was no statistically significant difference among interval II to V. The median r 2 value of each LAA%-interval was higher than 0.96. In LAM, interval IV was lower than interval I and II, and the r 2 value seemed to decrease step by step from interval I to IV. In BHDS, there was no statistically significant difference among all intervals, however, the r 2 value seemed to increase step by step from interval I to V.

Fig 3. The relationships between LAA_n and LAA% in all lung fields.
Solid and dashed line shows that the correlation is statistically significant and non-significant with the linear regression with spline models, respectively. A: In COPD, LAA_n increased once and thereafter decreased with increase of LAA%. In all intervals, LAA_n-LAA% relationships were statistically significant. B: In LAM, LAA_n increased with increase of LAA% from 0% to 11.5%, and thereafter LAA%-LAA_n relationship was not statistically significant. C: In BHDS, LAA_n increased with increase of LAA% from 0% to 20.8%, and thereafter LAA_n-LAA% relationship was not statistically significant. LAA%: percentage of lung field occupied by low attenuation areas; LAA_n: number of LAA cluster.

Discussion
Our study has demonstrated the different characteristics of the size distribution of LAA clusters among COPD, LAM and BHDS. In COPD, emphysema is thought to result from chronic inflammation followed by increased alveolar wall cell death and/or failure of alveolar wall maintenance [13,14]. Previous study assessed the cumulative size distributions of LAA clusters on chest CT and reported that emphysematous lung had fractal property regardless of LAA% value [4]. Also in our COPD patients, most of the CT images (95.8%) had fractal property regardless of LAA% value. Understandably, both of the LAA_n-LAA% and LAA_s-LAA% relationships were the same as previous reports (1). This phenomenon is thought to be due to the developmental process of emphysema which is explained by applying "elastic spring network model" or "model of mechanical force-based destruction" [4,[6][7]. These models mean that alveolar wall rupture is not uniform in the lung, and might be more severe in local regions of relatively advanced emphysema.
LAM is thought to be a neoplastic disorder, and metastatic pulmonary lesions induce protease-antiprotease imbalance in local milieus leading to connective tissue matrix degradation and cyst formation [15][16][17]. This putative mechanism is analogous to one of the mechanisms Fig 6. The comparison of LAA_s among five LAA%-intervals. A: In COPD, interval V was the highest and there was no statistically significant difference among intervals I to IV. B: In LAM, intervals I and II were lower and thereafter LAA_s increased step by step from intervals III to V. C: In BHDS, interval III was slightly higher than II, and there was no significant difference among the other intervals. The line across the box indicates the median. The bottom of the box is at 25th percentile, and the top is at 75th percentile value. The minimal and maximal values are shown with short horizontal lines ("whiskers"). BHDS: Birt-Hogg-Dubé syndrome; COPD: chronic obstructive pulmonary disease; LAA%: percentage of lung field occupied by low attenuation areas; LAA_s: mean size of LAA cluster; LAM: lymphangioleiomyomatosis for emphysema, but the substantial difference in both diseases is its extent in the parenchyma. The protease-antiprotease imbalance is diffuse in COPD with varying intensity, whereas the imbalance is likely to be limited to the sites of LAM cell infiltration. Moreover, LAM cell infiltration of distal airways which leads to airway narrowing and air trapping may also have a role in cyst formation [18]. In our LAM patients, the characteristics of LAA_n-LAA% relationships is quite different from those in COPD. Moreover, the percentage of CT images with fractal property is lower than that in COPD (92.9% in LAM and 95.8% in COPD), and tends to decrease when LAA% increases (Fig 7). These results can be explained by the fact that medium-sized LAA clusters increase as LAA% increases. From these analyses, it seems that "mechanical force-based destruction model" can not fully explain the way of progression of pulmonary cysts in LAM. Other than pulmonary cysts, LAM cells induce lymphatic vessel hyperplasia and lymphatic obstruction resulting in pulmonary lymphedema, and also involve pulmonary vessels resulting in pulmonary hemorrhage [17,[19][20][21]. These pathological changes may play a key role in preventing expansion and fusion of cysts.
BHDS is a rare autosomal dominant inherited genodermatosis, and the FLCN gene located in chromosome 17p11.2 has recently been identified to be defective [22][23][24]. The mechanism Difference of the progression of pulmonary cysts assessed by CT among cystic lung diseases for the progression of pulmonary cysts in BHDS remains unclear, but several hypotheses have been presented. Graham et al. postulated that folliculin may play a role in lung growth [25]. Painter et al. reported that pulmonary cysts may develop from inflammation caused by macrophages or fibroblasts in which folliculin mRNA is strongly expressed [26]. Recently, Furuya et al. reported that dysregulation of the mTOR pathway may induce cyst formation through proliferation of alveolar type II (ATII) cells [27], while Kumasaka et al. postulated that FLCN mutation may result in abnormalities at the alveolar-septal junction with few inflammatory changes [28]. And more recently, Goncharova et al. reported that pulmonary cysts in BHDS may result from an underlying defect in ATII survival, attributable to FLCN regulation of the E-cadherin-LKB1-AMPK axis [29]. In our BHDS patients, both of LAA_n-LAA% and LAA_s-LAA% relationships were quite different from those in COPD and LAM. The percentage of CT images with fractal property is very low (63.0%), and tends to increase as LAA% increases. These results can be explained by the fact that very large-sized LAA clusters already exist in CT images with low LAA%, and the number of small to medium-sized LAA clusters increase as LAA% increases. These analyses show that the mechanism of the progression of pulmonary cysts in BHDS has quite different characteristics from those in COPD and LAM. Therefore, the hypotheses of Graham, Kumasaka, and Goncharova may accord with the results of our image analyses.
Our study may have several limitations. First, this is a retrospective study, and the number of patients with BHDS included is relatively small. Second, this is a cross-sectional study. In patients with BHDS, it was reported that there was no significant correlation between pulmonary function and LAA% (3), and LAA% may not represent disease severity and/or activity. Therefore, longitudinal analysis is needed.
In conclusion, our study has demonstrated for the first time the different characteristics of the size distribution of LAA clusters among COPD, LAM and BHDS. These results seem to represent the difference in the mechanism of the progression of pulmonary cysts, and a longitudinal study is necessary to confirm it. Our study indicates that quantitative analysis of LAA clusters on CT images is useful for exploration of the pathophysiology in cystic lung diseases.