Introduction

International Association for the Study of Lung Cancer (IASLC) proposed the eighth edition of the clinical and pathological T descriptor classification of lung adenocarcinomas [1]. IASLC recommended the classification of cT according to the solid portion size on CT scans and the determination of pT according to the histopathological invasive part size [1]. For the part-solid nodules (PSNs), several studies have shown that the solid portions in PSNs are well correlated to the invasive component of the pathological specimens [2,3,4]. However, for the pure ground-glass nodules (pGGNs), which do not have any measurable solid portion on CT scans, ‘cTis’ should be assigned to all pGGNs of total size ranging from 0.6 to 3.0 cm [1].

Nevertheless, it is well known that a subset of pGGNs exhibits a pathological invasive component [5]. Moon et al. [6] reported that 20.5% of the pGGNs in their study were invasive pulmonary adenocarcinomas (IPAs), and Lim et al. [7] stated that 39% of the pGGNs ≥10 mm in diameter were IPAs. In the absence of a measurable solid portion, researchers have searched for other potential surrogates including nodule size, volume, mass, histogram features or texture features which could be used to predict IPAs appearing as pGGNs [5,6,7,8,9,10,11].

Among those tentative surrogates, nodule mass was reported as a less variable parameter than nodule volume. It was also reported that the mass could reflect the change in volume and attenuation simultaneously, given that it is calculated as the product of volume and attenuation [12]. Nonetheless, the usefulness of nodule mass among pGGNs for the differentiation of IPAs and its potential as a surrogate for the pathological invasive component have not been investigated thoroughly [7, 10]. The strength of mass over other size metrics in the diagnosis of pGGNs has not yet been established.

Therefore, we aimed to investigate the diagnostic advantage of mass in differentiating IPAs among pGGNs over other volumetric measurements on both non-enhanced and contrast-enhanced CT scans. We also analysed the correlation between the volumetric measurements including mass and the pathological invasive component size on surgical specimens.

Materials and methods

This retrospective analysis was approved by the Institutional Review Board of Seoul National University Hospital, and the requirement of written informed consent was waived.

Study population

Patients who had pGGNs at the preoperative chest CT scans and subsequently underwent surgical resections at our hospital between November 2011 and July 2017 were retrospectively identified from the electronic medical records by a board-certified thoracic radiologist (H.K. with 7 years of experience in CT imaging). Among 124 patients, 7 patients (5.6%) were excluded because of semi-automatic nodule segmentation failure. Consequently, a total of 117 patients, 53 men and 64 women, were included in this study. The mean age (± standard deviation) of the overall population was 57.3 ± 11.1 years (men, 58.4 ± 11.7 years; women, 56.3 ± 10.6 years; p = 0.311).

A single nodule was selected from each patient. A dominant lesion was chosen for patients with multiple resected pGGNs. Among 117 patients, 35 patients had both non-enhanced and contrast-enhanced CT scans; the rest of the population (n = 82) had either non-enhanced (n = 51) or contrast-enhanced CT scans (n = 31) (Fig. 1). The median interval between CT and surgical resection was 2 days [interquartile range (IQR), 1-23 days]. Sixty-six nodules (56.4%) were located in the upper lobes. For the surgical methods, lobectomy was performed in 23 patients, segmentectomy in 33 patients and wedge resection in 61 patients. Pathology revealed that there were 14 atypical adenomatous hyperplasias (AAHs), 35 adenocarcinomas in situ (AISs), 34 minimally invasive adenocarcinomas (MIAs) and 34 IPAs. The median invasive component size was 2 mm (IQR, 2-3 mm) in MIAs (n = 31; measurements were not available in 3 patients) and was 6 mm (IQR, 6-8 mm) in IPAs (n = 11; measurements were not available in 23 patients).

Fig. 1
figure 1

Flowchart of patient selection. CECT contrast-enhanced CT scan, NECT non-enhanced CT scan, pGGN pure ground-glass nodule

Full size image

Part of the study population was reported in past studies [13,14,15,16,17]. However, none of those studies focused on the differentiation of pGGNs, and the topics were completely different from that of the present study.

CT acquisition

CT scans were performed with seven different scanners from three manufacturers (Brilliance 64, Ingenuity, iCT and iQon, Philips Healthcare, Best, The Netherlands; Discovery CT750HD, GE Healthcare, Waukesha, WI, USA; Sensation 16 and Definition, Siemens Healthcare, Forchheim, Germany). Our hospital is a tertiary medical centre, and it operates multiple CT scanners from various vendors. Thus, heterogeneity in imaging acquisition was inevitable during the retrospective data collection. All patients underwent CT scans from the lung apex to base at suspended maximum inspiration. Scans were performed with 120 kVp and mAs ranging approximately from 20 to 200 mAs with or without automatic exposure control of each vendor. It is noteworthy that all CT scans were reconstructed with slice thickness/interval of 0.625, 1 or 1.25 mm to assure adequate volumetry analysis. Lung kernel or sharp kernel was used for the reconstruction. For contrast enhancement, a total of 70-90 mL of 370 mgI/mL of the contrast material, most of which was iopamidol (Pamiray 370; Dongkook Pharmaceutical, Seoul, Korea) or iopromide (Ultravist 370; Schering, Berlin, Germany), was injected at a rate of 2.0-2.3 mL/s using a power injector. CT scanning was performed with a 60-s delay.

CT image analysis

All image analysis was performed by a single radiologist (H.K.) with commercial software (AVIEW, Coreline Soft, Seoul, Korea). This software enables semi-automatic segmentation of nodules in two steps. The reader determined the target by dragging a line traversing the nodule on an axial plane and then selected the nodule type (pGGN). Then, the software automatically defined the nodule margin, calculated the volumetric profiles, and provided quantitative measurements. The reader was allowed to adjust the nodule margin in cases of suboptimal segmentation by tuning two segmentation parameters: nodule shape and attenuation threshold. In our study, nodule segmentation was evaluated subjectively, and segmentation success was defined as correct segmentation proportion of 80% or greater, according to past literature [13, 14, 18]. Volumetric features of nodule volume, attenuation, mass, two-dimensional (2D) average diameter and three-dimensional (3D) longest diameter were obtained. Mass was calculated as follows.

$$ \mathrm{Mass}=\mathrm{Volume}\times \left[\left(\mathrm{Attenuation}+\mathrm{1,000}\right)\times 0.001\right] $$

Pathological diagnosis

In our institution, all surgical specimens containing subsolid nodules (SSNs) were fixed by infusing 10% buffered formalin through the transpleural and transbronchial approach to produce an inflated state and precisely measure the invasive adenocarcinoma component [2]. Haematoxylin and eosin were used for staining the specimens. All pathological diagnoses were established based on the surgical specimens by attending pulmonary pathologists of Seoul National University Hospital according to the 2011 IASLC/ American Thoracic Society (ATS)/ European Respiratory Society (ERS) pulmonary adenocarcinoma classification [19]. Pathological diagnosis and invasive component size of each nodule were recorded [19]. The invasive component size was recorded only for pT1mi and pT1a according to our institutional protocol.

Statistical analysis

We calculated descriptive statistics of the volumetric features according to the pathologic diagnosis. After testing the normality of variables using Shapiro-Wilk test, values were described with either median (IQR) or mean ± standard deviation. Then, the features between preinvasive lesions/MIAs and IPAs were compared. Independent t-test or Mann-Whitney U test was used for the continuous variables (age, 3D longest diameter, 2D average diameter, volume, attenuation, and mass), and chi-squared test was used for the categorical variable (sex). Subsequently, receiver operating characteristic (ROC) curve analysis was performed to evaluate the diagnostic performance of each volumetric feature in discriminating IPAs. Areas under the ROC curve (AUCs) were obtained and compared between the features using DeLong et al.’s method [20]. For multiple comparisons, Bonferroni correction was applied to minimise α (type I error). Therefore, the significance level was adjusted to 0.01 for comparisons of volumetric features between preinvasive lesions/MIAs and IPAs and 0.0125 for comparisons of AUCs between mass and the other four volumetric features.

To analyse the correlation between the invasive component size of the surgical specimens and the volumetric features obtained from the semi-automated volumetry, Spearman’s rank correlation coefficient (rho) was calculated with 95% confidence interval (CI). Correlation coefficient between 0.0-0.2 was regarded as negligible, 0.2-0.4 as weak, 0.4-0.7 moderate, 0.7-0.9 strong and 0.9-1.0 very strong correlation [21].

All analyses were conducted for patients who underwent non-enhanced CT scans and those who had contrast-enhanced CT scans, respectively, to exclude bias caused by the contrast enhancement. A subset of patients (n = 35) who had both non-enhanced and enhanced CT scans belonged to both analysis groups. All statistical analyses were performed using two commercial software programs (MedCalc version 12.3.0, MedCalc Software, Mariakerke, Belgium and SPSS 19.0, IBM SPSS Statistics, Armonk, NY, USA). A p value <0.05 was considered as statistically significant.

Results

Volumetric measurements on non-enhanced CT scans and pathological diagnosis

Descriptive statistics of the volumetric features for each pathological diagnosis are shown in Table 1. All volumetric features exhibited an increasing trend according to the pathologic disease spectrum from AAH to IPA.

Table 1 Volumetric measurements of pure ground-glass nodules on non-enhanced CT scans according to the pathological diagnosis
Full size table

There were significant differences between the preinvasive lesions/MIAs and IPAs for all volumetric features (all p < 0.001) (Electronic Supplementary Table 1). ROC analysis for differentiating IPAs among pGGNs demonstrated that AUC was highest for mass (0.792; 95% CI, 0.691-0.872) and lowest for 3D longest diameter (0.739; 95% CI, 0.634-0.828) (Table 2). For the pairwise comparisons of AUCs between mass and the other four volumetric features, mass was superior to 3D longest diameter (p = 0.009) and volume (0.747; p < 0.001). However, AUCs were not significantly different between mass and 2D average diameter (0.780; p = 0.501) and between mass and attenuation (0.752; p = 0.598).

Table 2 ROC analysis of the volumetric measurements to discriminate IPAs among pure ground-glass nodules on non-enhanced CT scans
Full size table

Volumetric measurements of contrast-enhanced CT scans and pathological diagnosis

Table 3 shows descriptive statistics of the volumetric features for each pathological diagnosis. All volumetric features showed increasing tendency from AAH to IPA according to the pathological disease spectrum.

Table 3 Volumetric measurements of pure ground-glass nodules on contrast-enhanced CT scans according to the pathological diagnosis
Full size table

There were significant differences between the preinvasive lesions/MIAs and IPAs for attenuation (p = 0.002) and mass (p = 0.005) (Electronic Supplementary Table 2). Two-dimensional average diameter showed marginal statistical significance (p = 0.010). Three-dimensional longest diameter (p = 0.029) and volume (p = 0.025) were not significantly different for the preinvasive lesions/MIAs and IPAs after adjustment of the statistical significance level for the multiple comparisons. ROC analysis for differentiating IPAs among pGGNs demonstrated that AUC was highest for attenuation (0.761; 95% CI, 0.640-0.857) and lowest for 3D longest diameter (0.679; 95% CI, 0.553-0.789) (Table 4). AUC for mass was 0.730 (95% CI, 0.607-0.832). For the pairwise comparisons of AUCs, there were no significant differences between mass and the other four volumetric parameters [p = 0.067 between mass and 3D longest diameter; p = 0.319 between mass and 2D average diameter (AUC, 0.700); p = 0.015 between mass and volume (AUC, 0.683); p = 0.749 between mass and attenuation] (Fig. 2).

Table 4 ROC analysis of the volumetric measurements to discriminate IPAs among pure ground-glass nodules on contrast-enhanced CT scans
Full size table
Fig. 2
figure 2

Representative cases of pure ground-glass nodules. a Adenocarcinoma-in-situ in a 42-year-old man. Two-dimensional average diameter was 7.6 mm, and nodule mass was 0.11 g (contrast-enhanced CT). Both measurements corresponded well with the diagnosis. b Adenocarcinoma-in-situ in a 47-year-old man. The two-dimensional average diameter was 13.3 mm, and nodule mass was 0.49 g (non-enhanced CT). Based on the volumetric measurements, invasive adenocarcinoma was more likely than a preinvasive lesion. c Adenocarcinoma in a 45-year-old male. Two-dimensional average diameter was 9.2 mm, and nodule mass was 0.17 g (non-enhanced CT). Volumetric measurements, which are suggestive of a preinvasive lesion, were discordant with the diagnosis. Invasive component size was 6 mm in this lesion

Full size image

Correlation between volumetric features of non-enhanced CT and invasive component size

Invasive component size was recorded in a subset of patients (n = 66); median invasive component size was 1 mm (IQR, 0-3 mm). The correlations between the volumetric features and the pathological invasive component size were significant for 3D longest diameter (p < 0.001; rho = 0.422; moderate), 2D average diameter (p = 0.001; rho = 0.401; moderate), volume (p = 0.002; rho = 0.375; weak) and mass (p = 0.001; rho = 0.410; moderate) (Electronic Supplementary Table 3). Attenuation had no significant correlation with invasive component size (p = 0.176; rho = 0.168; negligible).

Correlation between volumetric features of contrast-enhanced CT and invasive component size

Invasive component size was obtained in a subset of the study population (n = 54). Median invasive component size was 1 mm (IQR, 0-3 mm). The correlations between volumetric features and pathological invasive component size were significant for 3D longest diameter (p = 0.022; rho = 0.310; weak), 2D average diameter (p = 0.044; rho = 0.276; weak), volume (p = 0.039; rho = 0.281; weak) and mass (p = 0.035; rho = 0.287; weak) (Electronic Supplementary Table 4). Attenuation did not exhibit a significant correlation with invasive component size (p = 0.394; rho = 0.118; negligible).

Discussion

In this study, we demonstrated that nodule mass is not superior to other volumetric measurements, such as 2D average diameter, for the discrimination of IPAs manifesting as pGGNs. The correlation between the volumetric measurements (mass, 3D longest diameter and 2D average diameter) and the invasive component size was moderate on non-enhanced CT scans and weak on contrast-enhanced CT scans.

The advantages of volumetric measurements for lung nodules have repeatedly been reported in the literature regarding measurement accuracy [22,23,24] and reproducibility [25,26,27]. The feasibility of volumetric analysis for the SSNs has also been assessed in the past decade, and a few studies reported the practicability and suggested the measurement variability ranges for the volume and mass [28,29,30]. Furthermore, Kim et al. [16] recently reported that the effective diameter obtained from semi-automatic volumetry exhibited higher diagnostic accuracy and reproducibility between readers for the differentiation of IPAs among SSNs than manual average diameter measurements. In this circumstance, it is assumed that the volumetric approach is a convincing diagnostic option for the SSNs. Among multiple volumetric measurements obtainable from the semi-automatic segmentation, mass was suggested as a less variable [12] and useful parameter for the evaluation of SSNs [7, 10]. This is particularly interesting for the pGGNs as they do not have a measurable solid portion which is a direct indicator of pathological invasiveness. Nevertheless, the diagnostic performance of the nodule mass was not superior to 2D average diameter as analysed in the present study.

The current standard for the measurement of SSNs uses the average diameter, which is measured manually with an electronic calliper on thin-section CT scans [31, 32]. Although manual measurement was not analysed in our study, we revealed that there are no significant differences in the diagnostic capability between 2D average diameter on an axial image and nodule mass at both non-enhanced and contrast-enhanced CT scans. Considering that nodule mass is affected by contrast material injection and the threshold for the diagnosis should thus be changed according to the enhancement status as demonstrated in our study (i.e. 0.23 g in non-enhanced CT and 0.37 g in contrast-enhanced CT; 61% increase in enhanced CT), it would be more practical and easy to use a single threshold of 2D average diameter (i.e. 10 mm; sensitivity 85.2% and specificity 64.4% at non-enhanced CT; sensitivity 88.2% and specificity 49.0% at enhanced CT), which can be applied regardless of the enhancement status. In fact, the contrast-enhancement profile of a nodule is influenced by the patient factor, contrast medium factor and CT scanning factor [33]. Therefore, mass, which reflects nodule attenuation, is inherently a vulnerable parameter for the acquisition setting. Thus, we cautiously suggest that the evaluation of pGGNs based on the 2D average diameter is sufficient and practical for clinical practice.

Recently, Heidinger et al. [8] reported a similar finding that measuring nodule volume and attenuation of pGGNs had no advantage over average diameter measurement. They performed correlation analysis between the CT measurements and the invasive foci (size and number) [8]. The results were similar to those in our study, and the correlation coefficients were 0.417 between the average diameter and the invasive focus, 0.401 between the nodule volume and the invasive focus, and 0.237 between the mean density and the invasive focus [8]. However, the diagnostic performance of the quantitative features for identifying IPAs was not evaluated, and the nodule mass was not calculated in their study. Also, the correlation analysis cannot be directly translated into a classification (differentiation) task. Thus, in this respect, we believe that our study provided additional supporting evidence for the usage of the conventional measurement metric, average diameter, rather than mass or volume for pGGNs.

Interestingly, the diagnostic performance of volumetric measurements and the correlation between the CT measurements and the invasive component size were lower at contrast-enhanced CT scans. This is plausible given that any penetrating or abutting vessels to the nodule may increase the nodule attenuation and size as well as nodule volume substantially [14]. The software we used in the present study applied a vessel removal algorithm to solve this issue but the measurements led to the suboptimal results. Cohen et al. [14] previously reported that the volumetric parameters of PSNs, including diameter and volume, significantly increased in enhanced CT. Although pGGNs were not analysed in that study, similar results may be inferred for the pGGNs.

Lastly, 10 out of 117 pGGNs (8.5%) had an invasive component size larger than 10 mm (data not shown). In our institution, an invasive component size larger than 10 mm is described with a greater than symbol (>10 mm) in the pathology report and the precise size is not measured. Thus, these nodules were excluded from the correlation analysis in our study. Weak-to-moderate correlation between the invasive component size and the volumetric measurements might be partly attributed to the exclusion of these nodules. A narrow range of variables may result in small correlation coefficients [34]. Nevertheless, it is noteworthy that a subset of pGGNs may have a substantially large invasive component. The median average diameter of these nodules was 13.9 mm (IQR, 13.2-17.5 mm).

There were several limitations to this study. First, the CT scans included in our study had heterogeneous acquisition settings including various manufacturers. Our hospital is a tertiary medical centre, which operates a few different CT scanners. Thus, this was an inevitable limitation of this retrospective study's design. Second, one commercial software was used, and our study results might be software specific. In addition, this software requires further validation for the solid nodules as well as SSNs. Third, manual measurements were not performed for the nodules. We focused on semi-automated volumetric measurements as the nodule mass cannot be accurately assessed with manual measurements. Comparison of the volumetric parameters, including mass with manual average diameter in terms of the diagnostic performance and reproducibility, is warranted. Fourth, measurement variability was not analysed. The inter- and intra-reader agreements of the software measurements for SSNs were excellent (intra-class correlation coefficient, 0.90-0.98) according to our prior study [35]. Fifth, our study population comprised patients who underwent surgical resection, and this might have induced selection bias. Small indolent pGGNs, which are usually followed-up with serial CT scans, were not analysed. Thus, the invasive component size in this study may not accurately represent that in screening-detected or incidentally detected pGGNs. Lastly, findings in this retrospective observational study should be validated in a future prospective cohort.

In conclusion, nodule mass measurement had no strength over other volumetric parameters for the prediction of invasiveness of lung adenocarcinoma in pGGNs. There was a moderate correlation between the volumetric parameters and the pathological invasive component.