Association between the number and size of intrapulmonary lymph nodes and chronic obstructive pulmonary disease severity

CT scans and data

CT scans, clinical data, and spirometry outcomes were obtained with approval from the COPDGene study (which was approved by all participating institutional review boards) for the purpose of performing ancillary study number 271 (Regan et al., 2010). All participants signed the consent form. A Data Access Investigator Certification Form was signed before the data was provided; ethical board approval for our ancillary study was waived. The CT protocols can be found in Appendix 1 of the study design (Regan et al., 2010). The only subject inclusion criterium was that a baseline and a five-year follow-up CT images were available. We initially calculated a recommended sample size of 39 per group based on a power analysis to detect an effect size of 0.25—moderate according to Fritz, Morris & Richler (2012)—with one-way ANOVA of five groups (GOLD stages 0 to 4), where α = 0.05 and power = 0.80. A pilot study had been performed with a sample size of 200 where about 10% of the scans were missing (results not published); we determined that additional subjects should be included to compensate for this. For our study, 50 subjects were randomly selected from each of the following seven groups (based on classifications at baseline): smokers classified as GOLD stages 0 to 4, smokers classified as PRISm, and never-smoking controls (GOLD 0).

Since 2017, the GOLD strategy documents recommended an “ABCD assessment” where COPD patients are further divided into one of four groups (i.e., A, B, C, or D, from low to high risk) depending on symptoms and exacerbation frequency and severity (Singh et al., 2019). Each individuals’ category was determined as a parameter for COPD symptom severity. Additionally, two quantitative CT measures of emphysema and chronic bronchitis were obtained using CIRRUS Lung Quantification (Diagnostic Image Analysis Group; Fraunhofer MEVIS): emphysema score and bronchial wall thickness (Pi10) (Gallardo-Estrella et al., 2017; Kim et al., 2014). Emphysema score was calculated by calculating the percentage of lung voxels below −950 HU after resampling the CT images to 3 mm slice thickness, normalization, and bullae analysis (Gevenois et al., 1996; Gallardo-Estrella et al., 2016). Pi10 is the square root of the airway wall area for a theoretical 10 mm lumen perimeter airway derived using the linear regression of the square root of segmented wall area against the lumen perimeter (Charbonnier et al., 2019).

Pulmonary nodules

The baseline and follow-up CT scans were screened for nodules by a trained medical researcher with support from computer-aided detection software CIRRUS Lung Screening (Diagnostic Image Analysis Group; Fraunhofer MEVIS). Only nodules of longest diameter of at least 3 mm were considered on the premise that the morphological features of smaller nodules would not be perceptible. Size was extracted automatically from the nodule segmentation in the form of volume, longest diameter, and short-axis diameter. Size changes in nodules which were detected in both scans were recorded; if the difference was larger than the known measurement variance of ±22.3%, the nodule was labelled as growing or shrinking (De Hoop et al., 2009).

In the baseline scans, all opacities deemed to be non-calcified nodules were subsequently assessed independently by two radiologists: a general radiologist with over 30 years’ experience and a focus on chest radiology including 5 years of lung cancer screening experience, and a chest radiologist with more than 20 years of experience in thoracic imaging. The classification options were as follows: “PFN”, “non-PFN solid”, “part-solid”, “non-solid”, “calcified”, and “not a nodule”. Disagreements were arbitrated by a third radiologist with over 30 years’ experience.

All radiologists were familiar with the morphology of PFNs as described in the literature (Ahn et al., 2010; De Hoop et al., 2012; Mets et al., 2018; Schreuder et al., 2018). Non-calcified solid nodules with sharp margins and an oval or polygonal shape were eligible as PFNs. Despite the nomenclature, PFNs do not need to be directly adjacent to fissures  (De Hoop et al., 2012; Mets et al., 2018; Schreuder et al., 2018). Readers were instructed to only classify a nodule as a PFN if highly certain that it was a lymph node. All opacities agreed on as “calcified” or “not a nodule” and opacities which were labelled differently by all readers were excluded from the statistical analysis.

Statistical analysis

The observer study outcomes were imported into Microsoft Excel™ 365 and subsequently to R statistical analysis package version 3.4.3. The number of PFNs, non-PFN solid nodules, part-solid nodules, and non-solid nodules were counted per scan. Nodule characteristics were compared between PFNs and non-PFN solid nodules using the Mann–Whitney U test for continuous variables and the chi-squared test for categorical variables. Inter-reader agreement was measured using Cohen’s kappa with all six nodule classification options and with two classes (PFN or not) (Cohen, 1960).

Regression analysis was performed to test whether the number and size (volume, longest diameter, and short-axis diameter) of PFNs detected could predict COPD severity. Ordered logistic regression (R function “polr” from package “MASS”) was used to model ordinal dependent variables (GOLD staging and ABCD assessment), and linear regression for continuous variables (emphysema score and Pi10). This analysis was performed using four outcome measures of COPD: GOLD severity groups (“Healthy” [GOLD 0 never and ever smokers], “Mild” [GOLD 1 and 2 and PRISm], and “Moderate/severe” [GOLD 3 and 4]), ABCD assessment, emphysema score (ln (x + 1) transformed due to non-linearity of residuals), and Pi10.

PFN count was analysed as an independent variable at the subject level; PFN sizes were modelled at the nodule level. This analysis was performed with and without adjustment for the baseline variables age, sex (female VS male), race (non-Hispanic Black VS non-Hispanic White), and smoking status (current VS former VS never). Analysis of GOLD severity groups was not adjusted for smoking status because this variable was used to define the groups. Statistical significance was defined as p < 0.0253 after applying the Šidák correction for multiple comparisons (calculated for only two hypotheses because the measures of PFN size and measures of COPD severity were considered complementary)  (Šidák, 1967; Rothman, 1990).The Shapiro–Wilk test was used to test for normality; all numeric outcomes were not normally distributed (p < 0.001) which justified performing the analyses using non-parametric tests.

Post hoc analysis was performed using univariable linear regression to associate more variables of interest—age, sex, race (White and non-Hispanic Black), educational level, smoking status, smoking duration, exacerbation frequency, at least one severe exacerbation, Modified Medical Research Council (mMRC) Dyspnea Scale, exposure to a dusty job (for at least one year), and exposure to fumes at work (for at least one year)—to PFN count and PFN volume. mMRC score and educational level was transformed into a numeric trend; for the latter, 1 = “8th grade or less”, 2 = “high school, no diploma”, 3 = “high school graduate or General Educational Development certificate”, 4 = “some college or technical school, no degree”, 5 = “college or technical school graduate (Bachlor’s or Associate degree)”, and 6 = Master’s or Doctoral degree”. For these 22 comparisons, the level of significance was lowered to 0.0023 (Šidák, 1967).

Observer study

The inclusion criteria and study design of COPDGene are described elsewhere (Regan et al., 2010). In short, it is a multicenter observational study collecting clinical, functional, imaging, and genetic data related to COPD. Ten thousand non-Hispanic Whites and Blacks between 45 and 80 years with and without COPD were enrolled at baseline (between January 2008 and June 2011). All participants required at least 10 pack-years of smoking, except for the control group of 133 never-smokers. A second screening round was performed five years after baseline.

The subject inclusion process is shown in Fig. 1. Of the 350 subjects included, 26 (7.4%) were excluded because the baseline and/or follow-up scans were missing from the database. The baseline characteristics across GOLD stages are shown in Table 1. Out of 324 baseline scans, 592 nodules were included in the observer study; 211 scans (65.1%) contained at least one nodule. After the observer study had been performed independently by Readers A and B, there were 163 disagreements (27.5%) which were arbitrated by Reader C. After arbitration, six opacities (6/592, 1.0%) were not considered nodules, five (5/592, 0.8%) were calcified, and six (6/592, 1.0%) were given different classifications by all three readers; these were excluded from the statistical analysis. Fourteen subjects (14/324, 4%) had missing values for mMRC score (therefore also lacking an ABCD assessment), 10 (3%) had missing values for exposure to a dusty job (for at least one year), and 20 (6%) for exposure to fumes at work (for at least one year); these were excluded from the relevant analyses.

Inter-rater agreement

Agreement in the form of Cohen’s kappa between readers A and B for 592 nodules was 0.50 (95% confidence interval: 0.43 to 0.56) with six classification options and 0.54 (0.47 to 0.61) when only distinguishing between a PFN and all other classification options, interpretable as moderate agreement. Among the 163 disagreements, the six-class kappa between Readers A and C and Readers B and C were 0.06 (−0.06 to 0.17) and 0.17 (0.05 to 0.29), respectively; the two-class kappa were 0.11 (−0.04 to 0.26) and 0.10 (−0.05 to 0.25), respectively; these values indicate that there was no significant agreement (Cohen, 1960).

Nodule characteristics

The characteristics of the nodules analysed are summarized in Table 2. Most nodules were classified as PFNs (366/575, 63.7%); Fig. 2 is an example of what is considered a typical PFN. Compared to non-PFN solid nodules (180/575, 31.3%), PFNs were more often smaller, flatter in shape, found in the middle or lower lobes, and remained the same size at follow-up; note that no nodule in the study was found to have a volume doubling time of less than 600 days. A much smaller number of subsolid nodules (29/575, 5.0%) was detected.

Perifissural nodule count

The prevalence and distribution of PFNs across COPD GOLD groups are shown in Table 3 and Fig. 3, and the data is provided as a supplementary file which also includes basic characteristics. Roughly one-third of the scans contained no nodules (118/324, 36.4%) and half contained no PFNs (154/324, 47.5%). The most PFNs found in one scan was 10. The average number of PFNs ranged from 0.92 (1.50) in the GOLD 0 ever-smokers group to 1.40 (1.95) in the GOLD 4 group. PFN count was not a statistically significant predictor for GOLD severity groups (p = 0.50), ABCD assessment (p = 0.94), emphysema score (p = 0.13), or Pi10 (p = 0.70). The respective p values were 0.33, 0.69, 0.52, and 0.94 after adjusting for other factors. There were no significant associations of PFN count with other variables (Table 4), though there may be a trend where non-Hispanic Blacks have fewer PFNs than Whites (p = 0.005).

Perifissural nodule size

The distribution of PFN sizes across GOLD stages are listed in Table 3; Fig. 4 shows a bar plot of the PFN density (number of PFNs in the size category divided by the number of PFNs in the GOLD stage) versus volume per GOLD stage. The median volume, longest diameter, and short-axis diameter of the PFNs analysed was 22.3 mm³ (interquartile range = 18 mm³), 4.8 mm (1.7 mm), and 3.2 mm (1.0 mm), respectively.

Volume, longest diameter, and short-axis diameter were larger on average in subjects more severe GOLD stages (p = 0.050, p = 0.023, and p = 0.021, respectively). The differences were largest between the Mild and Moderate/severe groups; the difference between the Healthy and Mild groups were not significant for longest diameter (p = 0.51) and short-axis diameter (p = 0.83). Univariable analysis for ABCD assessment showed a stronger trend, where larger PFNs were usually found in symptomatic participants at risk of exacerbation (volume p = 0.002; longest diameter p = 0.011; short-axis p = 0.012). Only PFN volume was a significant predictor of emphysema score (p = 0.043; longest diameter p = 0.053; volume p = 0.07, respectively). None of the PFN sizes could be used as predictors of Pi10 (p = 0.14, p = 0.19, and p = 0.23, respectively).

After adjusting for other factors (excluding smoking status), only short-axis diameter remained a statistically significant predictor of GOLD severity group (p = 0.012; volume p = 0.050; longest diameter p = 0.07). The respective p values for adjusted analysis of ABCD assessment are 0.076, 0.017, and 0.053. A stronger correlation was found (similar to the univariable analyses) when excluding smoking status from the list of adjustment factors (volume p = 0.002; longest diameter p = 0.011; short-axis p = 0.011). The most statistically significant post hoc finding was the positive trend between PFN volume and mMRC score (p = 0.004), though this only explained 2% of the variation (adjusted r-squared).