A CT-based nomogram for predicting the risk of adenocarcinomas in patients with subsolid nodule according to the 2021 WHO classification

A total of 1054 patients with SSNs who underwent surgery at our institution between April 2019 and December 2020 were retrospectively analyzed. All cases were enrolled based on a strict inclusion and exclusion criteria. The inclusion criteria were as follows: (1) patients with SSNs on CT scan; (2) maximum diameter ≤ 30 mm; (3) complete surgical resection. Exclusion criteria were: (1) no detailed pathology; (2) history of other malignancies; (3) neither precursor glandular lesions or adenocarcinomas; (4) anti-tumor therapy or biopsy; (5) CT scan and surgery were separated by more than one month; (6) missed complete thin-slice images (1 mm); (7) poor CT image quality. Finally, 656 cases were enrolled in this study. A flowchart showing participant recruitment is shown in Fig. 1. All clinical characteristics CT images of the participants were extracted from the Picture and Communication Systems (PACS) and the hospital’s electronic medical records (EMR) systems. The Ethics Committee of our institution approved the study (approval No. k2020-009) and waived the requirement of informed consent.

Chest CT images were acquired using a 64-slice spiral CT system (SOMATOM Definition AS+; Siemens Healthcare, Germany) using a pulmonary window center of -500 Hounsfield units (HU), a window width of 1500 HU, mediastinal window center of 50 HU, and a window width of 350 HU. Prior to examination, participants were trained on how to take a deep breath and hold their breath. Each participant was scanned while in the headfirst supine position, with arms raised. CT scans were acquired at the end of inspiration, with the CT scanning ranging from the tip of the lung to the diaphragm. Scanning parameters were as follows: 120 kVp tube voltage, automatic tube current modulation, a pitch of 1.2, a matrix of 512×512, 7 mm slice thickness, and reconstructed slice thickness of 1 mm with high-resolution reconstruction algorithm.

CT images were examined by radiologists A, B and C (with 4, 9 and 20 years of chest CT imaging experience, respectively), using a commercially available advanced workstation (VE40A, Siemens Healthineers) containing a tumor quantitative analysis software package (Siemens Healthineers). All radiologists were blinded to all clinical data and pathological results, except for age and sex of the patients. The software package has a semi-automated lesion segmentation tool. Raw 1 mm CT imaging data (DICOM format) were loaded into the software package. SSNs were segmented manually by tracing the segmentation boundaries of the nodules slice by slice, on axial images, excluding areas where large vessels and bronchi predominated. Nodule segmentation was conducted by radiologist A using the semi-automated tool. Segmentation boundaries were checked by radiologist B, and any discrepancies of segmentation boundaries of SSNs between the two radiologists were resolved by radiologist C. The segmentation tool produced long axis, short axis, mean CT value, and volume of the lesion. Lesion size was calculated as the average of the long- and short-axis diameters. Nodule mass was calculated as using the formula: mass(g) = (mean CT value + 1000) × volume(cm³)/1000 [12]. Morphological features of the SSNs were examined by radiologists A and B through multiple planar reconstruction (MPR), maximal intensity projection (MIP), volume rending technique (VRT), and minimum intensity projection (MinIP). The morphological features of SSNs including the following: (1) vascular change was abnormal vascular broadening or distortion [13], and was classified as present or absent; (2) bronchiole change was bronchus with dilated or tortuous lumen [14], and was classified as present or absent; (3) lobulation was defined as the outline of the lesion is not purely circular or oval, was classified as present or absent; (4) bubble was defined as a gaseous density with maximum diameter < 5mm [15], was classified as present or absent; (5) pleural attachment was defined as the pleura that was pulled to the lesion by a linear structure, and was classified as present or absent; (6) spiculation was defined as linear strands extending beyond the lesion [16], and was classified as present or absent; (7) lesion-lung interface, the border between the lesion and the normal lung tissues, and was classified as clear and blurry. Any discrepancies in the morphological features of SSNs between radiologists A and B were resolved by radiologist C.

Agreements between radiologists A, B, and C were evaluated with the Intraclass correlation coefficient (ICC) for quantitative parameters and Kappa coefficient for categorical variables. 30 SSNs (5%) were randomly selected for segmentation and morphological feature evaluation. At first, radiologists A, B, and C conducted segmentation and morphological feature evaluation independently. After 2 weeks, the segmentation and morphological feature evaluation of the SSNs were performed by three radiologists a second time.

All enrolled pathological specimens were independently reviewed by two senior pathologists, and the final pathological results were obtained through consensus. Pathological diagnosis was classified as AAH, AIS, MIA, and IAC based on the 2015 WHO classification of pulmonary adenocarcinomas [4]. Notably, AAH and AIS were classified as precursor glandular lesions and MIA and IAC as adenocarcinomas based on the 2021 WHO classification [6].

SPSS 26.0 (IBM) and R 3.5.1 (http://​www.​r-project.​org) were used for statistical analyses. Normality of quantitative parameters was tested using the Shapiro-Wilk test. Quantitative parameters satisfying the normal distribution were expressed as ‾x ± s. Otherwise, quantitative parameters were expressed as median (P25, P75). Student’s t test or Mann-Whitney U test were used to compare differences between continuous data. Categorical variables were compared using the chi-square test. Univariate and multivariate analyses using logistic regression were performed for the derivation cohort to identify independent risk factors for adenocarcinoma. The independent risk factors were then used to construct a nomogram. The nomogram’s discriminative capacity was first internally validated using 1,000 bootstrap samples to acquire a Harrell concordance index (C-index) in the derivation cohort. The nomogram was then tested on the validation cohort for external validation. Two-sided P <0.05 indicated statistical significance.

A total of 656 participants (681 SSNs) were enrolled, including 241 males and 415 females, with a mean age of 52.03 ± 12.26. There were 546 never smokers and 110 current or former smokers. Moreover, 407 patients (423 SSNs) operated between April 2019 and April 2020 were into the derivation cohort and 249 patients (258 SSNs) operated between May 2020 and December 2020 were assigned into the validation cohort. The baseline clinical characteristics and CT features of SSNs are shown in Tables 1 and 2, respectively. In derivation cohort, 110 SSNs were diagnosed as precursor glandular lesions (AAH =7, AIS =103), and 313 SSNs were diagnosed as adenocarcinomas (MIA =144, IAC =169). There were significant differences between precursor glandular lesions and adenocarcinomas subgroups in the lesion size (P <0.001), mean CT value (P <0.001), volume (P <0.001), mass (P <0.001), vascular change (P <0.001), bronchiole change (P <0.001), lobulation (P <0.001), pleural attachment (P <0.001), spiculation (P =0.032), and lesion-lung interface (P <0.001), except for bubble (P =0.081). In validation cohort, 72 SSNs were diagnosed as precursor glandular lesions (AAH =4, AIS =68), and 186 SSNs were diagnosed as adenocarcinomas (MIA =89, IAC =97). There were significant differences between precursor glandular lesions and adenocarcinomas subgroups in the lesion size (P <0.001), mean CT value (P <0.001), volume (P <0.001), mass (P <0.001), vascular change (P <0.001), bronchiole change (P =0.003), lobulation (P <0.001), pleural attachment (P =0.006), lesion-lung interface (P =0.003), except for bubble (P =0.359), and spiculation (P =0.092). The CT features and pathological results are shown in Table 3. The CT and pathological images from the 2 examples are shown in Fig. 2.

Intraobserver and intraobserver agreements between three radiologists were near perfect, the ICC values of quantitative parameters and kappa coefficients of categorical variables were all greater than 0.75 (supplemental Tables 1 and 2).

Univariate analysis of the derivation cohort indicated lesion size (OR =1.373; 95% CI, 1.061-1.777, P =0.016), mean CT value (OR =1.005; 95% CI, 1.001-1.010, P =0.024), vascular change (OR =5.125; 95%CI, 1.437-18.281, P =0.012), lobulation (OR =6.196; 95%CI, 2.007-19.127, P =0.002), and spiculation (OR =2.436; 95%CI, 1.055-5.625, P =0.037) correlated with adenocarcinomas. However, volume, mass, bronchiole change, bubble, pleural attachment, and lesion-lung interface did not correlate with adenocarcinomas (P >0.05). Stepwise multivariate analysis showed that lesion size (OR =1.335; 95% CI, 1.178-1.512, P <0.001), mean CT value (OR =1.005; 95% CI, 1.002-1.008, P =0.002), vascular change (OR =5.771; 95% CI, 1.659-20.074, P =0.006), and lobulation (OR =6.528; 95% CI, 2.173-19.608, P =0.001) were independent risk factors for adenocarcinomas (Table 4).

Based on univariate and multivariate logistic regression analysis results, an individualized nomogram was generated by incorporating the 4 independent risk factors, namely lesion size, mean CT value, vascular change and lobulation (Fig. 3a). The nomogram showed that lesion size was the most important contributor to discrimination, followed by mean CT value, lobulation, and vascular change. Each independent risk factor in the nomogram was assigned a point based on regression coefficient and a straight line drawn based on total points. Finally, the probabilities of individual values were determined using the function conversion relationship of total points (Fig. 3b).

In the derivation cohort, the C-index of the nomogram in predicting adenocarcinomas was 0.867 (95% CI, 0.833-0.901) which exceeded that of the lesion size (C-index =0.779; 95% CI, 0.733-0.825), mean CT value (C-index =0.740; 95% CI, 0.688-0.793), vascular change (C-index =0.723; 95% CI, 0.691-0.754), and lobulation (C-index =0.734, 95% CI, 0.701-0.767) (Fig. 4a, Table 5). Lesion size 8.5mm and mean CT value -579.5 HU were the optimal threshold values for adenocarcinomas.

Furthermore, the C-index of 0.877 (95% CI, 0.836-0.917) indicated that the nomogram had good discrimination in the validation cohort (Fig. 4b, Table 5). Evaluation of the nomograms’ performance using calibration curves, with the 45-degree line indicating best performance, revealed that the predicted results were strongly consistent with the actual results in both derivation and validation cohorts (Fig. 5a, d). Decision curve analysis of the nomogram’s value and clinical impact curve analysis revealed that the nomogram had good standardized net benefit and prediction performance (Fig. 5b, c, e, f).