Development and validation of growth prediction models for multiple pulmonary ground-glass nodules based on CT features, radiomics, and deep learning

Introduction

With the widespread application of low-dose and multi-slice spiral computed tomography (CT), the detection rate of ground-glass nodules (GGNs) has significantly increased (1,2). GGN, identifiable by hazy opacity on high-resolution CT that does not obscure underlying bronchial structures or vessels and measures ≤3 cm in diameter, can be classified as pure or mixed based on internal solid components. Recently, diagnoses of multiple GGNs have increased, with 25.6–48.5% of patients having multiple lesions (2,3). These lesions are predominantly multiple primary lung adenocarcinomas or the glandular precursor lesions rather than metastatic tumours (4-7). GGNs span a pathological spectrum from atypical adenomatous hyperplasia (AAH) to adenocarcinoma in situ (AIS), minimally invasive adenocarcinomas (MIA), and invasive adenocarcinoma (IAC). The presence of GGNs can cause patients to fall into severe anxiety and depression. Managing multiple GGNs remains challenging, as current guidelines addressing GGNs (8-11) come with no consensus on management strategies. Excessive follow-up and treatment are common, especially in Asia. Key questions include optimal follow-up schedules, timing for surgical intervention, and prioritising which GGNs to treat while preserving lung function. Answering these questions hinges on understanding the natural growth history and growth prediction for multiple pulmonary GGNs.

The growth rate and pattern of GGNs significantly influence their follow-up and intervention strategies (8-12). For nodules exhibiting slow growth, periodic follow-up may be considered appropriate. However, in cases of rapid-growing GGNs, closer observation is recommended to monitor growth and facilitate timely intervention. The GGNs exhibit an indolent growth pattern. On the whole, the median volume doubling time (VDT) and the median mass doubling time (MDT) of GGNs were approximately 800 days (13-15). Age, bubble lucency, lobulated sign, diameter, volume, and mass, and the presence and development of solid components were significant risk factors for GGN growth (14,16-19). With the rapid advancement of radiomics and deep learning (DL), several scholars have developed predictive models for GGN growth based on radiomics and DL, achieving promising predictive performance, and often outperforming traditional radiologic models (20-22). However, most of these studies have focused on solitary GGN, and there remains a significant gap in understanding the natural growth patterns and developing growth prediction models for multiple GGNs.

Multiple GGNs may grow simultaneously or at different time points, with variations between patients. A retrospective study found that 32% of multiple GGNs grew within 36 months, and if one GGN grew, the remaining GGNs also grew in 41% of patients (23). Among patients with multiple GGNs, 7.6–26.5% of the remaining lesions progressed after resection of the main lesion (24-26). Despite these findings, studies on the growth of multiple GGNs and the development of their predictive models remain limited. Therefore, this study aimed to investigate the natural history of multiple GGNs and develop and validate models based on CT features, radiomics, and DL to predict the growth of GGNs as well as compare the predictive performance of these models. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-1039/rc).

Methods

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by ethics committee of National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College (No. 22/434-3636) and individual consent for this study was waived due to the retrospective nature of the study.

Patient selection

Patients with multiple GGNs who underwent thin-section (≤1.25 mm) chest CT scans between October 2010 and November 2023 were retrospectively reviewed. Inclusion criteria were (I) patients with two or more persistent GGNs and CT follow-up of at least 3 years; (II) no prior radiotherapy, chemotherapy, or surgery; (III) for patients with five or fewer GGNs, all GGNs were included; for patients with more than five GGNs, only the five largest GGNs were analysed. Exclusion criteria included: (I) patients with only one GGN; (II) multiple GGNs with follow-up less than 3 years; (III) diffuse GGNs, as well as fibrotic alterations, suspicion of interstitial lung disease or bronchiolitis; (IV) lesions ≥3 cm; (V) failure to perform thin-section CT scans; (VI) inability to accurately detect and segment GGNs using the Dr. Wise system (developed by Deepwise AI Laboratory; https://www.deepwise.com/product-drwise) (Figure 1).

Figure 1 Cohort selection flowchart. Flow diagram shows the inclusion and exclusion criteria for patients with multiple GGNs. CT, computed tomography; GGNs, ground-glass nodules.

Finally, 732 GGNs from 231 patients were included. The 732 enrolled GGNs were randomly divided into a training set (n=518) and an internal validation set (n=214) at a 7:3 ratio.

CT examination

All CT scans were performed using 64-detector row scanners, including the LightSpeed VCT, Discovery CT 750 HD, and Optima CT 660 from General Electric Medical Systems, and the TOSHIBA Aquilion from TOSHIBA Medical Systems Scans were performed with patients fully inhaling, using standardised parameters: a tube voltage of 120 kV, automatic mA adjustments (200–350 mA), a noise index of 13, pitches of 0.992 or 0.984, rotation time of 0.5 seconds, and a slice thickness of ≤1.25 mm. Reconstructions were performed with slice thicknesses of 1.25 or 1.0 mm, accompanied by an interval of 0.8 mm, utilizing a high-spatial-frequency reconstruction algorithm. The imaging analysis involved assessment at lung (window width, 1,600 HU; window level, −600 HU) and mediastinal (window width, 350 HU; window level, 40 HU) settings.

Clinical and radiological analyses

One radiologist searched for all enrolled patients’ clinical and demographic features. The radiological features were independently reviewed by two radiologists with 8 and 3 years of experience in diagnosing thoracic neoplasms, respectively. Any disagreements were resolved by a consensus consultation with a senior radiologist with 30 years of experience in chest CT.

GGNs were categorized into pure GGNs and mixed GGNs and their morphological characteristics including GGN subtype and location, lobulation, spiculation, vacuole, air bronchogram, cystic sign, and pleural adhesion/retraction were observed.

The Dr. Wise System facilitated the automatic and accurate detection, segmentation, quantitative analysis, and follow-up comparison of GGNs (National Device Registration approval No. 20203210920), calculating mean diameter, mean CT value, volume, mass, VDT, and MDT at the baseline and follow-up CT scans.

Definition of GGN growth

The enrolled GGNs were divided into growth and non-growth groups according to interval changes during the follow-up period. Growth of GGNs was defined by a mean diameter increase of at least 2 mm, a volume increase of at least 30%, or the emergence or enlargement of solid components by at least 2 mm (15,16,27).

Model development

A total of six predictive models were developed in this study: two clinical models (Clinical model 1 and Clinical model 2), one Radiomics model, one DL model, one clinical-radiomics combined model (Clinical-Radiomics model) and one clinical-DL combined model (Clinical-DL model). The establishment of the models is shown in Figure 2.

Figure 2 Overview of the framework of the predictive models. Flowchart of the clinical model (solid blue box). Flowchart of the radiomics model (solid red box). Flowchart of the deep learning model (solid orange box). 4D, four-dimensional; AUC, area under the receiver operating characteristics curve; DFL, discriminative filter learning; LASSO, least absolute shrinkage and selection operator; ROI, region of interest.

Univariate and multivariate logistic regression analyses were performed to select clinical features. Variables with P<0.05 in univariate analysis were included in multivariate analysis. Lobulation, vacuole sign, initial volume, and mass were identified as independent predictors and used to construct Clinical model 1. Considering the potential linear correlation between diameter and volume, as well as mass and CT values, Clinical model 2 were developed with lobulation, vacuole sign, and initial mean diameter.

For the Radiomics model, a total of 1,454 radiomic features were extracted from each region of interest (ROI) after image preprocessing and normalization. Feature selection was performed using univariate logistic regression followed by minimum redundancy maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) regression to reduce dimensionality and eliminate redundant features. The selected features were used to develop the Radiomics model with a support vector machine (SVM) classifier. To construct the Clinical-Radiomics model, the significant clinical variables and the selected radiomics features were combined and input into an SVM classifier. Detailed information on the Radiomics model and the Clinical-Radiomics model development is presented in the Supplementary file (Appendix 1).

For the DL Model, a multi-branch architecture was implemented, consisting of global, medium, and local feature extraction branches. The global branch extracted contextual information using a 3D DenseNet backbone, while the medium and local branches focused on medium-resolution and fine-grained local features of the nodules, respectively. All convolutional and pooling layers were modified from 2D to 3D to capture volumetric information. The Clinical-DL model was developed by integrating the DL-extracted features with the significant clinical factors using a fully connected layer for final classification. Detailed information on the DL model and the Clinical-DL model development is presented in the Supplementary file (Appendix 1).

All models were trained using the training set with five-fold cross-validation, and tested on the remaining independent validation set. The optimal cut-off values for classification were determined according to the maximum Youden index derived from the validation set within the cross-validation process, and these thresholds were then applied to the independent validation set for performance evaluation.

Statistical analysis

Continuous data with a normal distribution were expressed as mean ± standard deviation, while non-normally distributed data were expressed as median [interquartile range (IQR)]. Categorical data were presented as counts (percentages). Statistical analyses of clinical and radiological characteristics between the groups with GGN growth and without growth were performed. For continuous variables, t-tests or Mann-Whitney U tests were applied, while categorical variables were analysed using Chi-squared tests or Fisher’s exact tests. Univariate analyses were conducted to identify factors predictive of GGN growth, including age, sex, nodule location, morphological characteristics, and initial measurements (diameter, CT attenuation value, volume, and mass). Kaplan-Meier estimates and log-rank tests were utilized to assess these variables. Independent predictors of growth were identified using multivariate Cox regression analysis with backward selection, focusing on variables with univariate P values <0.20. Statistically significant indicators were used as risk factors for the growth of pulmonary multiple GGNs. Univariate and multivariate logistic regression analyses were performed to select clinical features. Variables with P<0.05 in univariate analysis were included in multivariate analysis. Independent predictors were used to construct clinical models. Variables, including initial mean diameter, volume, and mass, were further categorised. Enrolled GGNs were stratified into subgroups based on initial mean diameter (<8 and ≥8 mm), volume (<326 and ≥326 mm³), and mass (<107 and ≥107 mg). The Kaplan-Meier method and log-rank test were performed to compare the time-to-growth curves among these subgroups. The area under the receiver operating characteristic curve (AUC) was used to evaluate the performance of the Clinical, Radiomics, DL, Clinical-Radiomics, and Clinical-DL models. In addition, sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were computed. Comparisons of the AUC among models were conducted using the DeLong test. P values were calculated using a two-sided test, and P values <0.05 were considered to indicate statistical significance. Data were analysed using SPSS (version 26.0; IBM Corp.) and the Python 3 programming language (version 3.6; IBM Corp.).

Results

Clinical and radiological characteristics of the enrolled patients and their multiple GGNs

A total of 732 GGNs from 231 patients were evaluated in this study. The median age of the patients was 55.0 (IQR, 47.0–60.0) years, with 26.4% (61/231) male and 73.6% (170/231) female. The median follow-up period was 2,164 (IQR, 1,951–2,628) days, with 197 patients followed up for at least 5 years and 34 for less than 5 years. During follow-up, 37.2% (272/732) of GGNs exhibited growth, 61.6% (451/732) remained stable or decreased, and 1.2% (9/732) resolved. The growth group included 272 GGNs, whereas the non-growth group included 460 GGNs. Patient age (P=0.02) and lobulation (P=0.03) significantly differed between the growth and non-growth groups; while sex (P=0.46), GGNs location (P=0.60), GGN type (P=0.21), initial mean diameter (P=0.37), initial mean CT value (P=0.35), initial volume (P=0.75), initial mass (P=0.88), follow-up period (P=0.06), and some morphological characteristics including spiculation (P=0.15), vacuole (P=0.052), air bronchogram (P=0.61), cystic sign (P=0.56) and pleural adhesion/retraction (P=0.07) did not significantly differ between the two groups.

The training set comprised 163 patients with 518 GGNs, and the validation set included 68 patients with 214 GGNs. The clinical and radiological characteristics of the enrolled patients are summarised in Table 1. The details for the training and validation sets are presented in Tables 2,3, respectively. These two sets were used for subsequent model construction and validation.

Multiple GGNs growth

On a per-patient basis, 67.5% (156/231) of the enrolled patients had growth in at least one GGN, with a median follow-up of 2,164 (IQR, 1,951–2,628) days. Among these patients, 48.7% (76/156), 34.6% (54/156), 12.2% (19/156), 2.6% (4/156) and 1.9% (3/156) of the patients showed one, two, three, four, and five GGNs growth, respectively. All GGNs in the 12 patients (12/156, 7.7%) showed growth. The largest GGN in mean diameter showed growth in 94 patients (94/156, 60.3%); among the 94 patients, 32 (32/94, 34.0%) showed only the largest GGN growth, and the remaining 62 (62/94, 66.0%) had extra 2–5 GGNs growth. Among the 156 patients with GGN growth, the fastest growing GGN showed a VDT and MDT of 2,285 (IQR, 1,369–3,545) and 2,438 (IQR, 1,361–4,140) days, respectively; and the median time-to-GGN growth was 1,109 (IQR, 593–1,644) days (Figure 3).

Figure 3 Example of follow-up CT for a patient with multiple GGNs. GGNs were first detected on June 7, 2017, and the final follow-up time was February 20, 2023. Follow-up time was 68 months. Pure GGN (one) with an initial diameter of 7.0 mm × 6.0 mm and initial volume of 136.0 mm³ showed growth during the follow-up period, with a VDT of 5,151 days. Pure GGN (two) with an initial diameter of 6.0 mm × 4.0 mm and an initial volume of 114.6 mm³ showed growth during the follow-up period, with a VDT of 896 days. Pure GGN (three) with an initial diameter of 6.0 mm × 5.0 mm and initial volume of 125.0 mm³ remains stable during the follow-up period. Pure GGN (four) with an initial diameter of 9.0 mm × 6.0 mm and initial volume of 291.9 mm³ remains stable during the follow-up period. Pure GGN (five) with an initial diameter of 6.0 mm × 5.0 mm and initial volume of 130.0 mm³ showed growth during the follow-up period, with a VDT of 1,835 days. The red arrow indicates GGN lesions. CT, computed tomography; GGNs, ground-glass nodules; VDT, volume doubling time.

On a per-nodule basis, 272 (37.2%) exhibited growth during follow-up. Notably, 34.2% (93/272) of the growing GGNs showed a mean diameter increase of at least 2 mm, 92.3% (251/272) exhibited a volume increase of at least 30%, and 22.4% (61/272) had a new or increased solid component. The 272 growing GGNs demonstrated a median VDT of 2,934 (IQR, 1,648–4,491) days and MDT of 2,875 (IQR, 1,619–5,148) days. The median time-to-GGN growth was 1,212 (IQR, 688–1,717) days. Additionally, among the 93 GGNs with a mean diameter increase of at least 2 mm, the median VDT and MDT were 1,647 (IQR, 1,168–2,434) and 1,630 (IQR, 1,138–2,596) days, respectively, and the median time-to-GGN growth was 944 (IQR, 541–1,362) days.

The time-to-GGNs growth is shown in Figure 4A-4D. The 1-, 2-, 5-, and 8-year cumulative percentages of GGN growth were 4.5%, 10.4%, 29.6%, and 45.5%, respectively (Figure 4A). The enrolled GGNs were divided into two subgroups based on the initial mean diameter: <8 mm group (n=637) and ≥8 mm group (n=95). A significant difference in the cumulative percentages of GGN growth was observed between the two subgroups (P<0.001), with a mean diameter ≥8 mm exhibiting a significantly higher cumulative percentage of growth compared to those with a mean diameter <8 mm (Figure 4B). GGNs were subcategorised based on a cutoff volume of 326 mm³ and a cutoff mass of 107 mg between the growth and non-growth groups. GGNs with initial volume ≥326 mm³ (n=77) demonstrated significantly higher cumulative percentages of growth than those with initial volume <326 mm³ (n=655) (P<0.001) (Figure 4C). Consistently, GGNs with initial mass ≥107 mg (n=653) showed significantly higher cumulative percentage of growth than those with an initial mass <107 mg (n=79) (P<0.001) (Figure 4D).

Figure 4 Kaplan-Meier plot for time-to-GGN growth according to (A) all enrolled nodules, (B) initial mean diameter, (C) initial volume, and (D) initial mass. (A) 1-, 2-, 5-, and 8-year cumulative percentages of GGN growth were 4.5%, 10.4%, 29.6%, and 45.5%, respectively. (B) GGNs with initial mean diameter ≥8 mm showed significantly higher cumulative percentages of growth than those with mean diameter <8 mm. (C) GGNs with initial volume ≥326 mm³ showed significantly higher cumulative percentages of growth than those with initial volume <326 mm³ (P<0.001). (D) GGNs with initial mass ≥107 mg showed significantly higher cumulative percentages of growth than those with initial mass <107 mg (P<0.001). GGN, ground glass nodule.

Risk factors for the growth of multiple GGNs

The outcomes of both univariate and multivariate analyses for the clinical and radiological characteristics predicting GGN growth in the training set are shown in Table 4. Multivariate Cox proportional hazards regression analysis was further used to analyze the clinical variables in Clinical model 1 and Clinical model 2. Results showed that in multivariate Cox proportional hazards regression analysis, the categorical features lobulation (P=0.049), vacuoles (P=0.009), initial volume (P=0.01), and initial mass (P=0.01) in Clinical model 1 were significant risk factors for predicting the growth of multiple GGNs. For Clinical model 2, the features lobulation (P=0.02), vacuoles (P=0.01), and initial mean diameter (P=0.03) were significant predictors.

Performance of the predictive models

Receiver operating characteristic curves (ROCs) showed the performance of six models in Figure 5. Among the six predictive models, the Clinical-Radiomics model demonstrated superior predictive accuracy, achieving the highest AUC value of 0.908 [95% confidence interval (CI): 0.861–0.954], outperforming the other five models. The AUC values for the other models were as follows: Clinical model 1, 0.876 [95% CI: 0.823–0.929]; Clinical model 2, 0.869 [95% CI: 0.815–0.924]; Radiomics model, 0.845 [95% CI: 0.786–0.904]; DL model, 0.735 [95% CI: 0.663–0.807]; Clinical-DL model, 0.887 [95% CI: 0.836–0.938]. Their overall performance is presented in Table 5. The Clinical-Radiomics model performed better in terms of accuracy (85.0%), sensitivity (82.3%), specificity (86.7%), and NPV (89.3%) compared to the other five models. The specificity for the six models is similar, around 80.0%. The PPV of the Clinical-Radiomics model was slightly lower than that of Clinical model 2, with values of 78.3% versus 86.8%, respectively.

Figure 5 Receiver operating characteristic curves show the performance of the Clinical model 1, Clinical model 2, Radiomics model, DL model, Clinical-Radiomics model, and Clinical-DL model in (A) training and (B) validation sets. Clinical-Radiomics model possessed the best predictive accuracy. AUC, area under the receiver operating characteristic curve; CI, confidence interval; DL, deep learning.

Discussion

Multiple pulmonary GGNs show indolent biological behaviour. We explored the risk factors of GGN growth and found that lobulation, vacuole, initial volume, and mass were significant predictors of GGN growth. We established six models based on CT features, radiomics, and DL to predict GGN growth. The Clinical-Radiomics model performed best in predicting the growth of multiple GGNs, with an AUC of 0.908, a sensitivity of 82.3%, and a specificity of 86.7%.

This study investigated the natural history of multiple GGNs and found that multiple GGNs showed a longer growth time, with a median VDT and MDT of 2,934 and 2,875 days, respectively. The time-to-GGN growth was longer than those of previous studies (16,27-29). This discrepancy might be because many enrolled pure GGNs were ≤5 mm in mean diameter, and growth was defined by volume changes. However, the VDT and MDT of GGNs with a mean diameter increase of at least 2 mm aligned with previous studies. The VDT and MDT of multiple GGNs were more than 3 years, suggesting that the follow-up time of multiple GGNs can be extended. Three-yearly follow-up is recommended to avoid wasting medical resources. NCCN guidelines (9) recommend surgical resection for pure GGN with a diameter of ≥20 mm or mixed GGN with a solid component ≥4 mm at follow-up. In this study, 21 GGNs met these criteria, although follow-up was still done. This may indicate that the surgical time can be appropriately delayed within a safe range.

Notably, multiple GGNs do not always exhibit the largest GGN growth. In this study, the growth of the largest GGN occurred in 60.3% of patients with multiple GGNs, whereas the remaining 39.7% did not have the largest GGN growth. For multiple GGNs, surgical strategies should consider the malignant potential of each nodule comprehensively. Lobulation, vacuole, initial volume, and mass were significant predictors of GGN growth, consistent with previous studies (14,30). Patients with GGNs are more likely to be never smokers than those with solid stage I adenocarcinoma (31). A recent study reported that GGNs are increasingly found in never-smokers, with about 60% of GGNs found in never-smokers in the past decades (32). These results indicate that there is no absolute correlation between GGNs and smoking. Accurate prediction of GGN growth is crucial for clinical decision-making and developing personalized surgical strategies.

Our results demonstrate that the Clinical-radiomics model exhibits superior predictive performance compared to the other five prediction models. To the best of our knowledge, numerous studies utilizing radiomic methods to analyze GGNs have been reported (33-35). While most of these studies focus on distinguishing benign from malignant or the invasiveness of pulmonary nodules, there is a paucity of radiomics models aimed at predicting the growth of pulmonary GGNs. One study aimed at predicting the growth trend of pulmonary GGNs constructed a clinical and radiomic combined model that achieved an AUC of 0.782, demonstrating higher predictive accuracy and good clinical utility as evidenced by calibration curves and decision curve analysis (DCA) (18). However, this study predominantly included patients with a single GGN, whereas our study enrolled patients with multiple GGNs. Another retrospective study aimed at predicting the growth of pulmonary GGNs developed and validated a nomogram model integrating radiomics and clinical features (size, location, and age) to predict the growth or long-term stability of GGNs. The highest AUC values were observed for the prediction nomogram model in both the training (AUC =0.843) and validation (AUC =0.824) sets (20). Nevertheless, the radiomics features in that study did not include any size-related features, which our study incorporated into the radiomics model. Ma et al. (21) classified GGNs into a fast-growth group (VDT ≤400 days) and a slow-growth group (VDT >400 days). They developed an integrated radiomics and radiographic features prediction model to predict the growth rate of pulmonary GGNs, achieving AUC values of 0.928 in the training set and 0.905 in the validation set, respectively. These results are comparable to our findings (AUC =0.913 and 0.908, respectively). However, in our study, the DL model demonstrated lower performance across all metrics, including an AUC of 0.735 [95% CI: 0.663–0.807], accuracy (69.6%), sensitivity (59.5%), specificity (75.6%), PPV (58.8%), and NPV (76.1%), compared to the other models. The relatively poor performance of the DL model is consistent with previous research (22). The lower performance of the DL model may be attributed to insufficient training data.

There are some limitations in this study. Nearly half of the GGNs were ≤5 mm in diameter, potentially biasing results. Additionally, the small number of patients with rapid GGN growth who had lesions resected within the 3-year follow-up period were excluded, possibly introducing bias. Furthermore, this study was conducted at a single centre with a relatively small sample size. Future research should involve multicentre prospective studies with larger sample sizes to address these limitations.

Conclusions

In conclusion, a significant proportion of patients with multiple GGNs experienced growth over time. During follow-up, some GGNs met the surgical criteria. Multiple pulmonary GGNs showed indolent biological behaviour, with long median time to growth. Risk factors for predicting GGN growth included lobulation, vacuole, initial volume, and mass. The Clinical-Radiomics model demonstrated the highest accuracy in predicting the growth of multiple GGNs compared to Clinical, Radiomics, DL, and Clinical-DL models.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-1039/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-1039/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-1039/prf

Funding: This work was supported by the Special Research Fund for Central Universities, Peking Union Medical College (grant No. 3332022025), the Fundamental Research Funds for the Central Universities, Peking Union Medical College (grant No. 3332024207), and the National Natural Science Foundation of China (grant No. 81971616).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-24-1039/coif). W.T. and Z.Z. were employed by Beijing Deepwise & League of PhD Technology Co. Ltd., Beijing, China. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by ethics committee of National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College (No. 22/434-3636) and individual consent for this study was waived due to the retrospective nature of the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Fan L, Wang Y, Zhou Y, et al. Lung Cancer Screening with Low-Dose CT: Baseline Screening Results in Shanghai. Acad Radiol 2019;26:1283-91. [Crossref] [PubMed]
2. Kim YW, Kwon BS, Lim SY, et al. Lung cancer probability and clinical outcomes of baseline and new subsolid nodules detected on low-dose CT screening. Thorax 2021;76:980-8. [Crossref] [PubMed]
3. Heuvelmans MA, Walter JE, Peters RB, et al. Relationship between nodule count and lung cancer probability in baseline CT lung cancer screening: The NELSON study. Lung Cancer 2017;113:45-50. [Crossref] [PubMed]
4. Detterbeck FC, Marom EM, Arenberg DA, et al. The IASLC Lung Cancer Staging Project: Background Data and Proposals for the Application of TNM Staging Rules to Lung Cancer Presenting as Multiple Nodules with Ground Glass or Lepidic Features or a Pneumonic Type of Involvement in the Forthcoming Eighth Edition of the TNM Classification. J Thorac Oncol 2016;11:666-80.
5. Wu C, Zhao C, Yang Y, et al. High Discrepancy of Driver Mutations in Patients with NSCLC and Synchronous Multiple Lung Ground-Glass Nodules. J Thorac Oncol 2015;10:778-83. [Crossref] [PubMed]
6. Hattori A, Takamochi K, Oh S, et al. Prognostic Classification of Multiple Primary Lung Cancers Based on a Ground-Glass Opacity Component. Ann Thorac Surg 2020;109:420-7. [Crossref] [PubMed]
7. Zhang Y, Li G, Li Y, et al. Imaging Features Suggestive of Multiple Primary Lung Adenocarcinomas. Ann Surg Oncol 2020;27:2061-70. [Crossref] [PubMed]
8. MacMahon H, Naidich DP, Goo JM, et al. Guidelines for Management of Incidental Pulmonary Nodules Detected on CT Images: From the Fleischner Society 2017. Radiology 2017;284:228-43. [Crossref] [PubMed]
9. Wood DE, Kazerooni EA, Aberle D. National Comprehensive Cancer Network. (NCCN) Clinical Practice Guidelines in Oncology. Guidelines Version 1.2023 Lung Cancer Screening 2022.10.26. Available online: https://nccnchina.org.cn/
10. Lung-RADS 2022 Assessment Categories 2022.11. Available online: https://edge.sitecorecloud.io/americancoldf5f-acrorgf92a-productioncb02-3650/media/ACR/Files/RADS/Lung-RADS/Lung-RADS-2022-Summary.pdf
11. Gould MK, Donington J, Lynch WR, et al. Evaluation of individuals with pulmonary nodules: when is it lung cancer? Diagnosis and management of lung cancer, 3rd ed: American College of Chest Physicians evidence-based clinical practice guidelines. Chest 2013;143:e93S-e120S.
12. Chen H, Kim AW, Hsin M, et al. The 2023 American Association for Thoracic Surgery (AATS) Expert Consensus Document: Management of subsolid lung nodules. J Thorac Cardiovasc Surg 2024;168:631-647.e11. [Crossref] [PubMed]
13. Chang B, Hwang JH, Choi YH, et al. Natural history of pure ground-glass opacity lung nodules detected by low-dose CT scan. Chest 2013;143:172-8. [Crossref] [PubMed]
14. Qi LL, Wu BT, Tang W, et al. Long-term follow-up of persistent pulmonary pure ground-glass nodules with deep learning-assisted nodule segmentation. Eur Radiol 2020;30:744-55. [Crossref] [PubMed]
15. Song YS, Park CM, Park SJ, et al. Volume and mass doubling times of persistent pulmonary subsolid nodules detected in patients without known malignancy. Radiology 2014;273:276-84. [Crossref] [PubMed]
16. Lee HW, Jin KN, Lee JK, et al. Long-Term Follow-Up of Ground-Glass Nodules After 5 Years of Stability. J Thorac Oncol 2019;14:1370-7. [Crossref] [PubMed]
17. de Margerie-Mellon C, Ngo LH, Gill RR, et al. The Growth Rate of Subsolid Lung Adenocarcinoma Nodules at Chest CT. Radiology 2020;297:189-98. [Crossref] [PubMed]
18. Gao C, Yan J, Luo Y, et al. The Growth Trend Predictions in Pulmonary Ground Glass Nodules Based on Radiomic CT Features. Front Oncol 2020;10:580809. [Crossref] [PubMed]
19. Lee JH, Lim WH, Hong JH, et al. Growth and Clinical Impact of 6-mm or Larger Subsolid Nodules after 5 Years of Stability at Chest CT. Radiology 2020;295:448-55. [Crossref] [PubMed]
20. Sun Y, Ma Z, Zhao W, et al. Computed tomography radiomics in growth prediction of pulmonary ground-glass nodules. Eur J Radiol 2023;159:110684. [Crossref] [PubMed]
21. Ma ZJ, Ma ZX, Sun YL, et al. Prediction of subsolid pulmonary nodule growth rate using radiomics. BMC Med Imaging 2023;23:177. [Crossref] [PubMed]
22. Tang Y, Li M, Lin B, et al. Deep learning-assisted development and validation of an algorithm for predicting the growth of persistent pure ground-glass nodules. Transl Lung Cancer Res 2023;12:2494-504. [Crossref] [PubMed]
23. Sato Y, Fujimoto D, Morimoto T, et al. Natural history and clinical characteristics of multiple pulmonary nodules with ground glass opacity. Respirology 2017;22:1615-21. [Crossref] [PubMed]
24. Shimada Y, Maehara S, Kudo Y, et al. Profiles of Lung Adenocarcinoma With Multiple Ground-Glass Opacities and the Fate of Residual Lesions. Ann Thorac Surg 2020;109:1722-30. [Crossref] [PubMed]
25. Xia T, Cai M, Zhuang Y, et al. Risk Factors for The Growth of Residual Nodule in Surgical Patients with Adenocarcinoma Presenting as Multifocal Ground-glass Nodules. Eur J Radiol 2020;133:109332. [Crossref] [PubMed]
26. Shimada Y, Saji H, Otani K, et al. Survival of a surgical series of lung cancer patients with synchronous multiple ground-glass opacities, and the management of their residual lesions. Lung Cancer 2015;88:174-80. [Crossref] [PubMed]
27. Tang EK, Chen CS, Wu CC, et al. Natural History of Persistent Pulmonary Subsolid Nodules: Long-Term Observation of Different Interval Growth. Heart Lung Circ 2019;28:1747-54. [Crossref] [PubMed]
28. Sawada S, Yamashita N, Sugimoto R, et al. Long-term Outcomes of Patients With Ground-Glass Opacities Detected Using CT Scanning. Chest 2017;151:308-15. [Crossref] [PubMed]
29. Lee JH, Park CM, Lee SM, et al. Persistent pulmonary subsolid nodules with solid portions of 5 mm or smaller: Their natural course and predictors of interval growth. Eur Radiol 2016;26:1529-37. [Crossref] [PubMed]
30. Takahashi S, Tanaka N, Okimoto T, et al. Long term follow-up for small pure ground-glass nodules: implications of determining an optimum follow-up period and high-resolution CT findings to predict the growth of nodules. Jpn J Radiol 2012;30:206-17. [Crossref] [PubMed]
31. Kobayashi Y, Ambrogio C, Mitsudomi T. Ground-glass nodules of the lung in never-smokers and smokers: clinical and genetic insights. Transl Lung Cancer Res 2018;7:487-97. [Crossref] [PubMed]
32. Woodard GA, Ding V, Cho C, et al. Comparative genomics between matched solid and lepidic portions of semi-solid lung adenocarcinomas. Lung Cancer 2023;180:107211. [Crossref] [PubMed]
33. Zhang T, Zhang C, Zhong Y, et al. A radiomics nomogram for invasiveness prediction in lung adenocarcinoma manifesting as part-solid nodules with solid components smaller than 6 mm. Front Oncol 2022;12:900049. [Crossref] [PubMed]
34. Liu J, Qi L, Wang Y, et al. Development of a combined radiomics and CT feature-based model for differentiating malignant from benign subcentimeter solid pulmonary nodules. Eur Radiol Exp 2024;8:8. [Crossref] [PubMed]
35. Wang Y, Lyu D, Yu D, et al. Intratumoral and peritumoral radiomics combined with computed tomography features for predicting the invasiveness of lung adenocarcinoma presenting as a subpleural ground-glass nodule with a consolidation-to-tumor ratio ≤50. J Thorac Dis 2024;16:5122-37. [Crossref] [PubMed]