Quantitative CT (QCT) analysis has wide acceptance as a tool to evaluate severity of chronic obstructive pulmonary disease (COPD) (12). Density mask technique is widely used to quantify emphysema. Threshold value of ≤ −950 Hounsfield unit (HU) on inspiratory CT is commonly used to define emphysema and correlates well with some spirometric measurements (345678910) and with pathologic severity of emphysema (111213). It may also correlate better with clinical status of COPD patients than spirometric evaluations (141516). However, a drawback of the density mask technique is that it is based solely on CT attenuation of voxels, and does not utilize morphological information provided by CT. It can vary depending on scanner model, scanner reconstruction parameters, degree of inspiration and smoking status (171819). Inclusion of morphological information in QCT analysis may more accurately define degrees of emphysema.

Low-attenuation (LA) cluster analysis calculates size and number of low-attenuation areas (LAA) seen in COPD on CT (20). This technique may provide additional morphological information regarding LAA, perhaps reproducing visual assessment of emphysema pattern (21). Although several investigators have used this measurement in studies (202122232425), its clinical relevance is not fully determined. Now both two-dimensional (2D) and three-dimensional (3D) methods (i.e., measuring cluster sizes two-dimensionally in each section as an area [mm²] and measuring cluster sizes three-dimensionally as a volume [mm³]) are available to calculate this measurement.

Thus the purpose of this study was to test if LA cluster analysis can provide information about severity of COPD independent of density mask technique by correlating CT measurements with physiological indices for assessment of COPD.

This study was approved by our Institutional Review Board (HS-2778). Written informed consent was obtained from all subjects, and the study was compliant with the Health Insurance Portability and Affordability Act. This study was retrospectively conducted in an institution using part of the prospectively acquired COPD Gene study data.

Subject characteristics and descriptive statistics of physiological measurements and QCT are summarized in Table 1. Results of correlation coefficients between 2D D value_{−950HU and −910HU} and 3D D value_{−950HU and −910HU}, and LAA%, clinical metrics and physiological measurements in subjects including subjects with mild COPD are shown in Tables 2 and 3, respectively. Data of 2D and 3D D values_−856HU are not shown because these values consistently revealed lower correlations than those of corresponding D values_{−950HU and −910HU}.

Two-dimensional and 3D D values had significant but mildly lower correlation coefficients with physiological measurements (r = −0.430, p < 0.001 between 2D D value_−910HU and %FEV1; r = −0.628, p < 0.001 between 3D D value_−910HU and D_LCO/V_A) than LAA% had (r = -0.549, p < 0.001 between LAA% and %FEV1; r = −0.646, p < 0.001 between LAA% and D_LCO/V_A) in subjects whereas in subjects with mild COPD some D values revealed slightly higher correlation coefficients with %FEV1 (r = −0.350, p < 0.001 between 2D D value_−910HU and %FEV1) and DLCO/VA (r = −0.652, p < 0.001 between 3D D value_−910HU vs. D_LCO/V_A) than LAA% did (r = −0.343, p = 0.001 between LAA% and %FEV1; r = -0.617, p < 0.001 between LAA% and D_LCO/V_A) (Fig. 2). Differences of correlation coefficients were not statistically significant (p = 0.168 for %FEV1 between 2D D value_−910HU vs. LAA%, p = 0.772 for D_LCO/V_A between 3D D value_−910HU vs. LAA% in subjects; p = 0.897 for %FEV1 between 2D D value_−910HU vs. LAA%, p = 0.631 for D_LCO/V_A between 3D D value_−910HU vs. LAA% in subjects with mild COPD) except for those between 3D D value_{−950HU and −910HU} and LAA% in subjects (p = 0.002 between 3D D value_−950HU vs. LAA%, p = 0.030 between 3D D value_−910HU vs. LAA%) (Table 4).

Among threshold CT values for calculating D value, 2D or 3D values at −910 HU always had the highest correlation coefficients with physiological measurements (r = −0.430, p < 0.001 between 2D D value_−910HU and %FEV1, and r = −0.628, p < 0.001 between 3D D value_−910HU and D_LCO/V_A in subjects; r = −0.350, p < 0.001 between 2D D value_−910HU and %FEV1, and r = −0.652, p < 0.001 between 3D D value_−910HU and D_LCO/V_A in subjects with mild COPD). Comparing between 2D and 3D techniques, 2D D values revealed higher correlations with physiological measurements except for D_LCO/V_A (r = −0.628, p < 0.001 for 3D D value_−910HU vs. r = −0.605, p < 0.001 for 2D D value_−910HU in subjects; r = −0.652, p < 0.001 for 3D D value_−910HU vs. r = −0.555, p < 0.001 for 2D D value_−910HU in subjects with mild COPD) than 3D D values did. Consequently, 2D D value_−910HU most frequently revealed highest correlations coefficients among D values. 2D D value_−910HU and 3D D value_−910HU had the highest correlation coefficients with %FEV1 and D_LCO/V_A, respectively, both in subjects (r = −0.430, p < 0.001 between 2D D value_−910HU and %FEV1; r = −0.628, p < 0.001 between 3D D value_−910HU and D_LCO/V_A) and in subjects with mild COPD (r = −0.350, p < 0.001 between 2D D value_−910HU and %FEV1; r = −0.652, p < 0.001 between 3D D value_−910HU and D_LCO/V_A), and values were used as an independent variable in the following multivariable analyses.

Multivariable analyses for %FEV1 revealed that LAA% (p < 0.001) in addition to BMI (p = 0.045) was a significant independent predictor of %FEV1 at an adjusted R² of 0.303 in subjects (Table 5) and that no independent variable remained significant with only 2D D value_−910HU revealing a marginal p value (0.050) at an adjusted R² of 0.119 in subjects with mild COPD (Table 6). Multivariable analyses for D_LCO/V_A revealed that LAA% (p < 0.001 both in subjects and subjects with mild COPD) and 3D D value_−910HU (p < 0.001 both in subjects and subjects with mild COPD) in addition to age (p < 0.001 in subjects and p = 0.022 in subjects with mild COPD) were significant independent predictors of D_LCO/V_A both in subjects (adjusted R² = 0.520) (Table 5) and subjects with mild COPD (adjusted R² = 0.555 ) (Table 6). Additionally, LAA% had the highest standardized β value among independent variables on the multivariable analysis for all subjects (−0.424) (Table 5) whereas 3D D value_−910HU had the highest β value (−0.410) (Table 6) for subjects with mild COPD.

Our study validated that cluster size analysis significantly correlated with physiological measurements for COPD although correlations were slightly weaker than those of the density mask technique (LAA%). When subjects were restricted to those with mild COPD, correlation coefficients between 2D and 3D D values and physiological measurements became higher whereas those between LAA% and physiological measurements became lower with 3D D values_{−950HU and −910HU} revealing higher correlation coefficients with D_LCO/V_A than LAA% although these differences were not statistically significant. Multivariable analyses also validated that only 2D D value_−910HU revealed a marginal p value (0.05) among independent variables in the regression model for %FEV1 and that 3D D value_−910HU had slightly higher standardized β value than LAA% in the regression model for D_LCO/V_A in subjects with mild COPD. Results may suggest that in mild COPD cluster analysis more accurately represents severity of diseases. Zach et al. (29) revealed that healthy non-smoking adults had low attenuation areas, with a mean LAA% of 2.0 percent perhaps related to image noise and/or air in small peripheral airways. LAA impair detection of early emphysema. Since such LAA are likely to occur in isolated voxels, cluster size analysis may be less affected by LAA unrelated to COPD. Conversely, in the case of far advanced COPD of which most LAA have reached maximum size, size distribution of LAA on cluster analysis may no longer considerably contribute to grading severity of COPD, and therefore LAA% may be more advantageous.

Cluster analysis of LAA has been used in several previous studies to evaluate severity of emphysema. Mishima et al. (20) observed that COPD patients with normal LAA% had significantly smaller D value (i.e., gentler slopes, larger LAA predominant) than healthy subjects and hypothesized that small LAA coalesce to form larger LAA in COPD patients, and D value can become smaller without increasing LAA%. Gietema et al. (21) validated that visually assessed regions of air trapping had smaller cluster sizes for a similar LAA% than those without air trapping. Tanabe et al. (25) documented that increase in LAA% and decrease in D were greater in COPD patients with acute exacerbation than those without and validated that not only enlargement of LAA but also coalescence of adjoining LAA caused progression of emphysema in patients with acute exacerbation. These reports, in addition to this study, support the concept that cluster analysis provides incremental information to density mask technique because it estimates sizes of LAA that cannot be assessed by density mask technique, namely LAA%.

The selection of LAA threshold level for defining extent of emphysema is of interest. Several previous studies revealed that threshold levels around −950 HU for LAA provide strongest correlation with pathologic extent of emphysema (12). In our study, too, LAA% at threshold of −950 HU had higher correlations with FEV1 and D_LCO/V_A than LAA% at thresholds of −910 HU and -856 HU (data not shown). However, as for LA cluster analysis in this study, −910 HU was a better threshold for correlations with physiological measurements. One possible explanation for this discrepancy is the effect of respiratory bronchiolitis that may increase attenuation of LAA. Several previous studies revealed that current smokers have lower LAA% than former smokers (171819). Shaker et al. (19) revealed that smoking cessation results in apparent increase in LAA%, due to improvement of respiratory bronchiolitis. Although this effect affects LAA% and cluster analysis, the impact may be significant on cluster analysis because effacement of thin connections between LAA that may occur by using a lower threshold results in substantial underestimation of sizes of LA clusters. The cluster size may be more accurately represented at higher threshold of HU.

Madani et al. (24) compared cluster analysis (D value at −960 HU and 1st percentile HU) with LAA% at the threshold of −960 HU and first percentile HU, and revealed that LAA% and first percentile HU significantly correlated with pulmonary functional tests and pathological extent of emphysema, while D value at −960 HU did not. Our results are contradictory to their observations. A couple of explanations may account for discrepancies. First, as revealed in our study, −910 HU was a better threshold than −950 HU for cluster analysis. The −960 HU they used for cluster analysis may have not been the optimal threshold for evaluating potential of cluster analysis. We used 2D and 3D techniques, each of which had advantages over each other. Second, as their results were based on investigation of resected lung specimens from lung cancer surgeries, whole lung correlations in our study may have revealed different results.

In comparison between 2D and 3D techniques, 2D D value_−910HU generally revealed highest correlations with physiological measurements. However, as for D_LCO/V_A, 3D D value_−910HU correlated better than 2D D value_−910HU. We cannot determine if 2D or 3D technique is better for evaluation of COPD based on our results. We currently recommend that 2D and 3D techniques are used in evaluation of COPD.

Our study had limitations. First, although we revealed threshold of −910 HU was better than those of −950 HU and −856 HU in cluster analysis, the best threshold should be determined based on a more detailed investigation focusing on threshold CT value with smaller CT value intervals (e.g., 10 HU). Second, R² were generally too low in regression models for subjects with mild COPD to make a definitive conclusion. The apparent advantage of cluster analysis in mild COPD subjects should be confirmed in a larger number of subjects. Third, methods of cluster analysis on CT are not standardized. Our data cannot simply be applied to other institutions.

In conclusion, cluster analysis is a clinically feasible method to evaluate severity of COPD, especially in mild COPD. The threshold of −910 HU in cluster analysis was better than those of −950 HU and −856 HU for evaluation of COPD. 2D and 3D techniques for cluster analysis were almost even and complementary to each other, and therefore we recommend using both techniques in evaluation of COPD.