Introduction

The density of areas of emphysema depicted on computed tomography (CT) typically ranges from −900 to −1,024 Hounsfield Units (HU) [1]. Hence, a straightforward way of quantitatively assessing the extent of emphysema is to compute the low attenuation areas (LAA) under a specific threshold. Originally, Müller et al. [2] suggested a threshold of −910 HU for this density mask analysis. Later, Gevenois et al. [3] explained that a threshold of −950 HU correlated more closely with the macroscopic pathological features of emphysema. Since then, a threshold of −950 HU has been commonly used for density-based emphysema quantification [4–7]. However, recently, Madani et al. [8] suggested that a cut-off of −960 HU or −970 HU might yield more accurate assessment of the extent of emphysema. Therefore, there is no consensus on the optimal threshold for density-based emphysema quantification.

Often, a global index is computed to quantify the severity of emphysema by measuring the ratio of the low attenuation area below a specific threshold as a fraction of the entire lung volume [9]. As a global measure, this index ignores the fact that the density of emphysema may vary regionally throughout the lung. Considering that pulmonary lobes are relatively independent functional units with their own vascular and airway branches, there has been increasing interest in investigating the lobar distribution of emphysema [10–14]. Given the heterogeneity of emphysema, it remains an open question whether different thresholds are needed to accurately quantify emphysema in individual lobes. Different thresholds have been employed by investigators to quantify the extent of emphysema at the individual lobe level. For example, when the threshold was set at −910 HU, Saitoh et al. [10] found that the distribution of emphysema was different in the upper and lower lobes; when the threshold was set at −950 HU, Mohamed Hoesein et al. [14] found that subjects with upper lobe predominant emphysema tended to have a more rapid decline in measures of pulmonary function than those with lower lobe predominant emphysema; when investigating the relationship between pulmonary function and normal lobar volume, Matsuo et al. [13] also used −950 HU as the cut-off threshold. Given the heterogeneity of emphysema, it may be desirable to have a standardised cut-off threshold for density mask analysis in clinical practice.

In this study, we systematically investigated how LAA% varies with different thresholds on quantitative CT analysis of pulmonary emphysema at levels of both the individual lobe and the entire lung. A large and diverse dataset consisting of 573 chest CT examinations was collected from a chronic obstructive pulmonary disease (COPD) study. Here, our objective was to determine whether an optimal threshold can be used for more accurate and consistent quantification of emphysema for each pulmonary lobe. At the same time, we also assessed how CT densitometric measurements among these lobes correlate with the COPD severity and the commonly used pulmonary function test (PFT) measurements (e.g. forced expiratory volume in 1 s [FEV1]).

Materials and methods

Study population

The study population consisted of 573 consenting subjects participating in an NIH-sponsored Specialized Centre for Clinically Oriented Research (SCCOR) in COPD at the University of Pittsburgh. SCCOR subjects were primarily recruited from the Pittsburgh Lung Screening Study cohort. It is a tobacco-exposed cohort and includes a spectrum of obstructive lung disease severity. Inclusion criteria for enrolment required an age >40 years and at least a 10-pack-year history of tobacco use. The subjects underwent pre- and post-bronchodilator spirometry and plethysmography tests, measurements of lung diffusion capacity, impulse oscillometry, demographic and medical history questionnaires and a chest CT examination. All subjects were smokers with an average smoking history of 58.3 pack-years. These individuals were classified using a five-category classification, namely: (1) no airflow obstruction (None), (2) mild obstruction (GOLD I), (3) moderate obstruction (GOLD II), (4) severe obstruction (GOLD III) and (5) very severe obstruction (GOLD IV). Subjects that have FEV1/FVC <70 % were classified into GOLD stages I-IV according to the Global Initiative for Obstructive Lung Disease (GOLD) [15]. The subjects that do not meet the GOLD classification were classified as “None”. All procedures were performed under a University of Pittsburgh Institutional Review Board-approved protocol (no. 0612016) and written informed consent was obtained for each subject. The involved SCCOR subject demographics are summarised in Table 1.

Table 1 Subject demographics (n = 573)
Full size table

Acquisition of thin-section CT examinations

The CT examinations were performed using 64-detector CT (LightSpeed VCT; GE Healthcare, Waukesha, WI, USA) with subjects holding their breath at end-inspiration. The CT datasets were acquired using a helical technique without contrast medium at the following parameters: 32 × 0.625 mm detector configuration, 0.969 pitch, 120-kVp tube energy, 250-mA tube current, and 0.4-s gantry rotation (or 100 mAs). Images were reconstructed to encompass the entire lung field in a 512 × 512-pixel matrix using the GE “bone” kernel at 0.625-mm section thickness and 0.625-mm interval. The “bone” kernel was used because of its ability to analyse both the parenchyma and airways [16].

Quantification of the extent of emphysema

Computerised schemes were used to process the CT examinations objectively and efficiently. The extent of emphysema at the levels of both the entire lung and individual lobes was quantified. First, a three-dimensional (3D) adaptive border marching algorithm described by Pu et al. [17] was used to identify the entire lung volume. Second, pulmonary fissures were detected using a computational geometry approach [18]. Third, individual lobes were identified by representing the pulmonary fissures as implicit surface functions [19]. Detailed descriptions of these computerised schemes have been reported elsewhere [17–19]. The final segmentation results were carefully inspected and verified by three independent image analysts. If the computerised schemes failed to accurately identify individual lobes, we manually delineated or refined the pulmonary fissure before applying the automated lobe segmentation scheme. Several examples in Fig. 1 are used to demonstrate the performance of the identified lobe segmentation scheme in the presence of severe emphysema.

Fig. 1
figure 1

Examples demonstrating the performance of the computerised schemes [16–18] in lobe identification in the presence of severe emphysema (GOLD = 4 for all three examples). (a)–(c) Example 1; (d)–(f) Example 2; and (g)–(i) Example 3

Full size image

In this study, the percentage of the LAA divided by lung or lobe volume(s) was used as an index of the extent of emphysema [20–23]. A wide range of thresholds, ranging from −850 HU to −1,000 HU in intervals of 5 HU, were tested. The upper bound of −850 HU was selected because of the weak correlation between emphysema severity and LAA% at the threshold of −856 HU [24]. The lower boundary of −1,000 HU was chosen because the radiodensity of air is typically defined as −1,000 HU [25]. LAA% was computed at the levels of the entire lung, individual lobes (RUL, right upper lobe; RML, right middle lobe; RLL, right lower lobe; LUL, left upper lobe; LLL, left lower lobe), the upper lobe (RUL + RML + LUL), the lower lobe (RLL + LLL), the left lung and the right lung.

Statistical data analysis

All statistical analyses were performed using Excel (Microsoft, Redmond, WA, USA) and SAS (SAS Institute, Cary, NC, USA). We computed the mean volumes and the LAA% for each individual lobe and the entire lung. These results are expressed as mean ± standard deviation (SD). Pearson correlation coefficients were used to assess the correlation between LAA% under different thresholds in individual lobes or the whole lung and PFT measures. In particular, we investigated whether the correlation would be different when the non-COPD subjects were included or excluded. When assessing the correlation between the five-category classification and LAA% under different cut-off thresholds, Spearman analysis was used. In addition, the correlations between lobe volume and emphysema severity based on our five-category classification were assessed. A P value less than 0.001 was considered statistically significant. The post-bronchodilator PFT measurements used in the correlation analyses included: (1) the residual volume (RV)/total lung capacity (TLC) ratio, (2) RV % predicted, (3) forced expiratory volume in 1 s (FEV1), (4) FEV1 % predicted, (5) FEV1/functional vital capacity (FVC) ratio and (6) diffusing lung capacity of carbon monoxide (DLco %) predicted.

Results

After the application of the lobe segmentation scheme to the collected 573 CT examinations (Table 1), the volumes of each individual lobe, the upper lobe, the lower lobe, the left lung, the right lung and the entire lung along their means (± SD) is summarised in Table 2. In Table 2, the LAA and LAA% obtained at the threshold of −950 HU were also listed. The LAA% ranged from 12.3 % to 18.4 %. In LAA%, there was no obvious difference between the left lung and the right lung, but there were differences between the upper lobe and the lower lobe. The average LAA% for each individual lobe under different thresholds is shown in Fig. 2, and the average LAA% for the upper lobe, the lower lobe, the right lung, the left lung and the entire lung under different thresholds is shown in Fig. 3.

Table 2 Volume, low attenuation area (LAA), and the percentage of LAA (LAA%) for each individual lobe, the upper lobe, the lower lobe, the left lung, the right lung and the entire lung
Full size table
Fig. 2
figure 2

Average LAA% for each individual lobe under different thresholds

Full size image
Fig. 3
figure 3

Average LAA% for the upper lobe, the lower lobe, the left lung, the right lung and the entire lung under different thresholds

Full size image

The Spearman correlations between LAA% under different thresholds for individual lobes and the five-category classification are shown in Fig. 4, demonstrating a strong association between LAA% and the five-category classification regardless of the threshold. It can be seen that the thresholds corresponding to the highest correlation vary for each lobe. For example, for RUL, the optimal threshold is −935 HU (r = 0.649, P < 0.001); for RML, the optimal threshold is −945 HU (r = 0.521, P < 0.001). The computed results at different thresholds of the above correlations at the level of the upper lobe, the lower lobe, the left lung, the right lung and the entire lung are shown in Fig. 5, demonstrating a strong association for all thresholds (r > 0.631, all P values < 0.001). It can be seen that the LAA% of the upper lobe and the right lung have a slightly higher association with the five-category classification than the LAA% of the lower lobe and the left lung at the same thresholds. The differences between the highest correlation coefficients and those at −950 HU were summarised in Table 3.

Fig. 4
figure 4

Correlations between the LAA% and the five-category classification under different thresholds for each individual lobe. Highest correlation coefficient r and the corresponding threshold: (1) RUL, r = 0.649 (−935 HU); (2) RML, r = 0.521 (−945 HU); (3) RLL, r = 0.617 (−940 HU); (4) LUL, r = 0.623 (−940 HU); (5) LLL, r = 0.596 (−945 HU). P < 0.001 for all r values

Full size image
Fig. 5
figure 5

Correlations between the LAA% and the five-category classification under different thresholds for the upper lobe, the lower lobe, the left lung, the right lung and the entire lung. The highest correlation coefficient r and the corresponding threshold: (1) the upper lobe, r = 0.649 (−940 HU); (2) the lower lobe, r = 0.631 (−940 HU); (3) the left lung, r = 0.657 (−945 HU); (4) the right lung, r = 0.677 (−940 HU) and the entire lung, r = 0.676 (−940 HU). P < 0.001 for all r values

Full size image
Table 3 Comparison of the highest correlation coefficients and those at −950 HU
Full size table

The computed Pearson correlation coefficients between the LAA% and the PFT measurements are listed in Table 4. The LAA% showed strong correlations with airflow obstruction measurements, such as FEV1% predicted (r = 0.578–0.726, P < 0.001), FEV1/FVC ratio (r = 0.634–0.795, P < 0.001) and DLco% predicted (r = 0.528–0.707, P < 0.001). For lung capacity-related measures, such as RT/TLC ratio and RV% predicted, the LAA% only showed moderate associations. Conversely, the LAA% of RML showed comparatively weak correlations with the PFT measures compared with the other lobes. It can be seen that the highest correlation coefficients and their corresponding cut-off thresholds were similar, no matter whether the non-COPD subjects were excluded or not.

Table 4 Correlations between the percentage of low attenuation area (LAA%) and PFT measurements. The results when the non-COPD subjects were excluded were presented in parentheses
Full size table

Similar to the first correlation analysis between the LAA% and the five-category classification, the thresholds corresponding to the highest correlation coefficients ranged from −925 HU to −965 HU, and the differences between the highest correlation coefficients and those computed at the threshold of −950 HU are summarised in Table 5. In addition, the significances (P values) of the differences between the highest correlation coefficients and those computed at −950 HU were also provided in Tables 3 and 5. Lastly, the correlations between lobe volumes and emphysema severity indexed in terms of five-category classification are shown in Fig. 6.

Table 5 Differences between the highest correlation coefficients and those computed at −950 HU
Full size table
Fig. 6
figure 6

Correlations between the five-category classification and lobe volumes. (1) RUL: r = 0.386, P < 0.001; (2) RML: -0.001, P = 0.976; (3) RLL: 0.135, P < 0.001; (4) LUL: 0.269, P < 0.001; (5) LLL: r = 0.200, P < 0.001

Full size image

Discussion

The density mask method has been commonly used to assess the severity of COPD. Here, we systematically investigated the impact of threshold selection on the accuracy of assessing the extent of emphysema. This study has a number of unique merits. First, a relatively large, diverse dataset was used. To the best of our knowledge, few investigations have used a dataset as large as ours. A large dataset enables a robust investigation of the underlying relationship between density mask and emphysema extent. Second, the correlations between emphysema extent at different thresholds and pulmonary functions are studies not only at the level of the entire lung but also at the level of individual lobes. This regional analysis may verify whether a uniform threshold is sufficient to quantitatively assess the extent of emphysema. As a result, the ultimately determined optimal threshold may facilitate a more accurate assessment of emphysema extent. Third, our study verified that using a single cut-off threshold for emphysema assessment could achieve a reasonable accuracy. Finally, compared with previous investigations, this study considers a larger number of PFT measures.

In this study, not only the LAA% of each individual lobe but also the LAA% of the entire lung were computed. We found that the five-category classification was strongly correlated with LAA% under thresholds ranging from −850 to −1,000 HU. There were slight variations in the optimal thresholds corresponding to the highest correlation coefficients for individual lobes. However, the highest correlation coefficients and those obtained at −950 HU do not vary significantly, and the percentage difference ranges from 0.2 % to 0.3 % (Table 3). This may suggest that the threshold of −950 HU is appropriate for emphysema quantification at the level of individual lobes. We arrive at the same threshold of −950 HU when similar analysis is repeated between the LAA% and pulmonary function (the percentage difference ranges from 0 % to 2.0 %). In fact, our results verified the conclusion reported in Pescarolo et al. [26], where the authors explained that there was significant correlation (r = −0.7, P < 0.001) between the five-category classification and the extent of emphysema. However, in the study of Pescarolo et al. [26], only 43 CT examinations were used.

When studying the correlation between emphysema extent and pulmonary function, Saitoh et al. [10] showed a strong correlation in the upper lobes, and Matsuo et al. [13] found that the LAA% of RLL and LLL had stronger correlations with VC, FEV1 and DLco. However, in our study with a relatively large dataset, we failed to observe obvious difference in correlation coefficients between the upper lobes and the lower lobes, although the correlation coefficients for the upper lobes were also higher than those for the lower lobes. The small sample sizes (i.e. n = 50 of Saitoh et al. [10] and n = 111 of Matsuo et al. [13]) may heavily skew the results towards predominantly emphysematous lobes. Nevertheless, we observed that the average LAA% values of the RUL and the LUL were slightly higher than those of the RLL and the LLL. This is in agreement with the results obtained by Mohamed Hoesein et al. [14]. In particular, we observed that the five-category classification and the PFT measurements showed weak correlations with the LAA% of the RML. This weak correlation had been reported previously by other investigators [10, 13] and may be largely attributed to the relatively small size of the RML. When studying the relationship between the five-category classification and the lobe volumes, as shown in Fig. 6, we found that the volume of the RML did not statistically correlate with the five-category classification (r = −0.001, P = 0.976). This may indicate that the RML inflation is only minimally affected when the severity of emphysema increases. A possible explanation is that the RML is less prone to emphysema. This may be explained by the fact that the average LAA% of RML was smaller than of the other lobes within the range of −920 to −1,000 HU (Fig. 3). In particular, we observed that CT emphysema measurements correlated significantly with the diffusing lung capacity of carbon monoxide (DLco) at levels of both individual lobes and the entire lung. Our results are in consistence with those in [10, 27, 28], suggesting that DLco might be a useful measurement for emphysema diagnosis.

We are aware of the limitations of this study. First, despite the relatively large size of the dataset, the CT examinations in this study are not equally distributed in each category. We note that it is very difficult to have a dataset that can meet such a criterion in practice. Fortunately, our conclusions may not be affected in this study because of the large number of CT examinations in each category. Second, it has been shown that the CT acquisition protocols, such as dose, image reconstruction kernels, and equipment, may have an impact on CT-based quantification [8, 25, 29]. In our study, we did not consider the details in the CT acquisition protocols. For example, the “bone” kernel used in this study may alter the displayed Hounsfield unit number. However, we had shown that the differences in correlation coefficients were very small under the threshold range −925 HU to −965 HU. In addition, Boedeker et al. [30] demonstrated that the “bone” kernel might shift the obtained density mask volume by 2.4 % compared with the “standard” kernel. The underlying implication is that even under different image acquisition protocols we should draw similar conclusions. Nevertheless, this study could be improved by using datasets acquired under different protocols. Third, the CT examinations used in this study were acquired at the end of full inspiration. It is likely that the lung density will vary when chest CT examinations are acquired at the end of full expiration, thereby leading to a completely different optimal threshold for the quantification of emphysema extent. In clinical practice, the CT for emphysema assessment is typically acquired at the end of full inspiration. It may be interesting, however, to investigate the optimal thresholds for emphysema assessment when CT examinations are acquired at other respiratory stages (e.g. full expiration).

In summary, our study investigated the optimal threshold for computerised tomographic assessment of the extent of emphysema using a large dataset consisting of 573 CT examinations. The density mask method was not only applied to the entire lung but also to each individual lobe. Our results show that the single threshold of −950 HU is optimal for CT densitometry analysis of emphysema when the CT examinations are acquired at full inspiration.