Artificial intelligence-based density proportion analysis in predicting the invasiveness of neoplastic ground-glass nodules

Introduction

Recently, the popularization of lung cancer screening and the widespread application of low-dose computed tomography (LDCT) in the chest have led to the increasing detection of ground-glass nodules (GGNs) (1,2). GGNs can be further divided into pure GGNs (pGGNs) and part-solid nodules (PSNs) based on the absent or present of solid component (3), and have a variety of properties. Pathologically, most neoplastic GGNs are adenocarcinomas with different degrees of invasiveness, including atypical adenomatous hyperplasia (AAH), adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IAC) (4,5). In clinical settings, preinvasive lesions (AAH and AIS) can be managed through follow-up in accordance with the recommendations of the Fleischner Society and the Lung Imaging Reporting and Data System (Lung-RADS) guidelines, or through limited resections (e.g., wedge resection or segmental resection). In contrast, invasive lesions (ILs) like MIA and IAC require more aggressive interventions, with sublobar resection typically recommended for patients with MIA and lobectomy for those with IAC (6-9). Therefore, accurately evaluating the invasiveness of neoplastic GGNs is particularly important for directing further treatment.

The computed tomography (CT) features, including quantitative and qualitative indicators, were typically used for evaluating the invasiveness of GGNs in previous studies, and have revealed the positive correlations between them (10-12). Among them, the density was more frequently used than others, and both the two-dimensional (2D) and three-dimensional (3D) mean CT values of nodules were effective for further evaluations. Although previous results indicate that mean CT values increase with increasing invasiveness of GGNs, different studies have shown inconsistent density thresholds for distinguishing the invasiveness of varied GGNs (13-16). The main reason for the difference could be that the mean CT value neglected the heterogeneity of GGNs in density. Additionally, the components with higher density in GGNs have a closer relationship with their invasiveness (17-19). Therefore, the application of density in evaluating the invasiveness of GGNs should be further studied.

Currently, artificial intelligence (AI) techniques have been widely used in the detection and evaluation of pulmonary nodules, which can extract the overall pixel values of nodules. Based on the overall pixel values, the proportion of internal components with density higher than any determined reference value can be easily obtained. Such information could be used for quantitatively reflecting the heterogeneity of GGNs, which in turn, may improve the value of density in assessing their invasiveness. Compared with prior approaches (density histograms, radiomics, and AI-derived attenuation features) that focus on the global quantitative characterization of nodules (20,21), the AI-based density proportion analysis uniquely quantifies the ratio of pathologically meaningful density fractions, directly linking radiological quantification to histological invasiveness. Furthermore, this method is less complex than radiomic models, rendering it more accessible for routine clinical practice with superior clinical interpretability. Thus, this study aimed to verify the value of AI-based density analysis in predicting the invasiveness of neoplastic GGNs. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1020/rc).

Methods

Patients selection

This single-center retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of The First Affiliated Hospital of Chongqing Medical University (No. 2019-062) and individual consent for this retrospective analysis was waived. We searched the electronic medical record system for patients who underwent surgical resection in thoracic surgery due to pulmonary ground-glass opacities between January 2019 and May 2023. A total of 1,024 patients were identified, and all GGNs were confirmed by pathological examination following surgical resection. Two radiologists (with more than 5 years of post-training experience) manually reviewed the patients’ images at the picture archiving and communication system (Carestream Vue PACS) workstation. The patients who met the following inclusion criteria were included: (I) lesions presenting as GGNs (maximum diameter ≤3 cm) on CT images; (II) patients with complete clinical, CT, and pathological data; (III) the GGNs were confirmed as neoplastic lesions; (IV) all the examinations were performed by the same brand of CT scanners; and (V) no needle biopsy or treatment was performed for lesions prior to CT examination. Patients with the following conditions were excluded: (I) the interval between chest CT examination and surgery was >1 month (n=16); (II) absence of thin-section CT images (≤1.0 mm) (n=11); (III) patients had severe and diffuse lung diseases (emphysema, interstitial lung disease, etc.) (n=25); and (IV) their CT images had artifacts or poor image quality, which affected evaluation (n=9). Finally, a total of 996 nodules in 963 patients were further analyzed and divided into training and validation cohorts with a ratio of 7:3. In the training and validation cohorts, there were 687 (247 AISs, 231 MIAs, and 209 IACs) and 309 (105 AISs, 103 MIAs, and 101 IACs) neoplastic GGNs in 654 and 309 patients, respectively.

CT examinations

CT examinations were performed using the following scanners: Somatom Definition Flash (Siemens Healthineers, Erlangen, Germany), Somatom Perspective (Siemens Healthineers), and Somatom Force (Siemens Healthineers). All patients were examined in a supine position with raised upper limbs, followed by performing a breath-holding exercise before image acquisition. CT examination was performed from the thoracic inlet to the costophrenic angle at the end of inspiration during a single breath-hold. The scanning parameters were as follows: tube voltage, 120 kVp; tube current, 50–130 mAs (using automatic current modulation technology); scanning slice thickness, 5 mm; rotation time, 0.5 seconds; pitch, 1.0–1.1; collimation, 0.60 mm; and matrix, 512×512. Images were reconstructed at a slice thickness and slice interval of 1.00 mm using iterative reconstruction with a medium-sharpness algorithm. Both mediastinal window (width, 350–400 HU; level, 20–40 HU) and lung window (width, 1,200–1,600 HU; level, −500 to −700 HU) were obtained.

Data acquisition and image interpretation

Quantitative data of GGNs were acquired using the lung nodule AI-assisted diagnosis system (InferRead CT Lung, InferVision Medical Health, Beijing, China), which is widely used for the detection, diagnosis, deep learning analysis, and radiomics assessment of pulmonary nodules (22-27). This system performs multiple tasks, including nodule detection, malignant tumor prediction, nodule segmentation, and automatic histogram analysis. The underlying algorithms involve Faster RCNN, ResNet, and UNet models. Faster RCNN is mainly responsible for the detection and localization of pulmonary nodules. ResNet mainly classifies the benign and malignant pulmonary nodules, while UNet achieves precise segmentation of pulmonary nodules. Objective and subjective assessments were both implemented for algorithmic validation of its ability in 3D reconstruction. The Sørensen-Dice coefficient (Dice coefficient) was used to objectively evaluate the similarity between the AI-3D output and manual reconstruction. The overall average Dice coefficient reached 0.88 [95% confidence interval (CI): 0.83–0.92] for bronchi segmentation, and 0.88 (95% CI: 0.83–0.91) and 0.91 (95% CI: 0.86–0.94) for pulmonary vessels segmentations on non-contrast and contrast-enhanced chest CT scans, respectively. In the subjective assessment, 94.35% for bronchi, 96.12% for vessels on non-contrast CT, and 93.22% for vessels on contrast-enhanced CT scans met the designated quality criteria (28). In this study, the acquisition of density histograms was achieved through the UNet model, which accurately segmented lesions and obtained a corresponding region of interest (ROI), in which the AI system extracted density features and performed quantitative analysis. In the whole calculation process, this system automatically divides the range of the nodules and calculates the number of voxels corresponding to each CT value in the whole GGN. Each CT value and the corresponding number of voxels are stored as a LIST, and the LIST of the whole nodule is stored as a DICTIONARY. Based on the DICTIONARY, the density histograms of GGNs are automatically obtained, and then the proportions of the components with higher density in lesions at density reference values (≥−800, ≥−750, ≥−700, ≥−650, ≥−600, ≥−550, ≥−500, ≥−450, ≥−400, ≥−350, ≥−300, ≥−250, and ≥−200 HU) were acquired by manually adjusting the bar (Figure 1).

Figure 1 A schematic figure shows the process of AI-based nodule density histogram acquisition, and the proportions of the components in lesions with higher density than the determined thresholds were obtained by manually adjusting the bar. The part in red in the GGN indicates the component with density higher than the corresponding threshold. AI, artificial intelligence; CT, computed tomography; GGN, ground-glass nodule.

At the same time, the CT images and clinical information of the lesions were reviewed independently by two additional thoracic radiologists (with more than 10 years of post-training experience) who were blinded to the pathological results. The following indicators were evaluated: size (the average of the maximum diameter of the nodule and the diameter perpendicular to the maximum diameter), distribution in lobes (right upper, middle and lower lobe; and left upper and lower lobe), shape [irregular vs. regular (round or oval)], boundary (well-defined vs. ill-defined), air bronchogram sign (present vs. absent), vacuole sign (present vs. absent), vascular changes (present vs. absent), lobulation sign (present vs. absent), spiculation sign (present vs. absent), and pleural indentation sign (present vs. absent). Air bronchogram sign was defined as visible air-filled bronchi detected in the nodule (29). Vacuole sign was defined as a round or oval air density within a nodule (30), whereas spiculation sign refers to multiple fine linear strands extending from a structure into the surrounding lung parenchyma in a stellate manner and lobulation sign refers to a lobule-like and often asymmetric protrusion at the margins of a structure (31), while pleural indentation sign was defined as a linear strand radiating from the nodule and extending distal to the pleural surface (32). In addition, vascular changes included dilation and distortion of involved vessels. Vascular dilation was observed when the diameter of the vascular segment within lesions was larger than the proximal segment (before entering lesions) or other vessels at the same branch level, whereas vascular distortion was observed when the vessel deviated from its normal route (33). Consensus was reached through consultation and discussion, just in case there were differences in opinion between the two radiologists.

Statistical methods

All data were processed by SPSS 25 software (IBM Corp., Armonk, NY, USA) and MedCalc software (MedCalc Software Ltd., Ostend, Belgium). Clinical data and various CT features were statistically analyzed for each patient. Continuous variables were represented by mean ± standard deviation. The Kruskal-Wallis test was used for age and nodule size, whereas Pearson’s Chi-squared test was used for sex, nodule shape, location, boundary, lobulation, spiculation, vacuole sign, vascular changes, air bronchogram sign, and pleural indentation sign. Given the substantial density heterogeneity of GGNs, with CT attenuation values mainly ranging from −800 to −200 HU in this study, excessively large density intervals may fail to capture meaningful characteristic variations, while overly narrow intervals would result in numerous subgroups with potentially overlapping outcomes. Based on these considerations, we selected a 50 HU interval for the final analysis. Then, the optimal density thresholds and the corresponding cutoff values for the proportions of the components in lesions with density higher than the thresholds for determining ILs and IACs were determined via receiver operating characteristic (ROC) curve analysis and comparing the area under the curves (AUCs) and Youden indexes. The Delong test was performed to assess the statistical difference between the traditional morphological features model and the combined model of the optimal density indicators and traditional morphological features in determining ILs and IACs, and compare the diagnostic efficacy between density threshold proportion and radiological features of GGNs for predicting ILs and IACs. To select clinical characteristics and morphological features, we first conducted univariate logistic analyses of the clinical characteristics and CT features to predict ILs and IACs. Subsequently, all statistically significant features (P<0.05) were included in the multivariate logistic regression models to identify independent factors associated with the invasiveness of ILs and IACs, respectively. P value less than 0.05 was considered statistically significant.

Results

Patients’ clinical data and CT features of GGNs in the training cohort

The patients’ clinical data and the CT features of GGNs are summarized in Table 1. The results revealed statistical significance in sex (P=0.01), shape (P=0.007), and vascular change (P=0.001), along with age, size, lobulation sign, spiculation sign, air bronchogram sign, and pleural indentation sign (P<0.001) between the three groups. The prevalence of AIS, MIA, and IAC was higher in females. The IAC corresponds to older age and larger size, as well as higher incidences of irregular shape, lobulation, spiculation, vacuole sign, air bronchogram, vascular change, and pleural indentation, followed by MIA and AIS.

The density threshold for distinguishing ILs and IACs

In the training cohort, the AUCs and cutoff values for proportions of the components in GGNs with higher density at different density thresholds for determining ILs and IACs are shown in Tables 2,3, respectively. For determining ILs, a higher AUC was obtained at the optimal density threshold of ≥−350 HU, and the corresponding cutoff value for the proportion of components with density higher than −350 HU was 17.22% (AUC: 0.801; 95% CI: 0.769–0.830; sensitivity: 51.59%; specificity: 93.52%; P<0.001), respectively. For determining IACs, a higher AUC was obtained at the optimal density threshold of ≥−250 HU, and the corresponding cutoff value for the proportion of components with density higher than −250 HU was 5.64% (AUC: 0.882; 95% CI: 0.855–0.905; sensitivity: 85.65%; specificity: 76.15%; P<0.001). In the validation cohort, these predictive parameters were also effective in determining ILs (AUC: 0.814; 95% CI: 0.766–0.856; sensitivity: 62.75%; specificity: 100%) and IACs (AUC: 0.885; 95% CI: 0.844–0.918; sensitivity: 100%; specificity: 76.92%), respectively (each P<0.001).

The cross-validation of the optimal density threshold for determining ILs and IACs

The 687 neoplastic GGNs in the training cohort were randomly divided into five groups, and a five-fold cross-validation for determining the optimal density threshold in predicting ILs and IACs was conducted, and the results are shown in Tables 4,5. For predicting ILs, four out of five groups had an optimal density threshold of ≥−350 HU. Moreover, for predicting IACs, all five groups had an optimal density threshold of ≥−250 HU. The AUCs of the models for predicting ILs (range, 0.790–0.810) and IACs (range, 0.879–0.889) in these training cohorts were comparable to those in the overall training cohort (0.801 and 0.882) but slightly higher than those in the validation cohort (0.707–0.747 for ILs and 0.745–0.814 for IACs).

Univariate logistic analysis of the patients’ clinical characteristics and CT features for predicting ILs and IACs in the training cohort

The univariate logistic analysis of the clinical characteristics and CT features for predicting ILs and IACs were conducted, and the results are shown in Tables 6,7. Compared with patients with AIS, those with ILs were more likely to be male and older; similarly, patients with IAC were more predominantly male and older than those with AIS/MIA (each P<0.05) (Tables 6,7). Compared to AIS, ILs were larger and exhibited a greater prevalence of irregular shape, lobulation sign, spiculation sign, air bronchogram sign, vascular change, and pleural retraction sign, and a higher proportion of the component in lesions with density ≥−350 HU (each P<0.05) (Table 6). Similarly, compared to AIS/MIAs, IACs were also larger and exhibited a greater prevalence of irregular shape, lobulation sign, spiculation sign, air bronchogram sign, vascular change, and pleural retraction sign, and a higher proportion of the component in lesions with density ≥−250 HU (each P<0.05) (Table 7).

Multivariate logistic regression models for distinguishing ILs and IACs

Two multivariate logistic regression models were performed based on the morphological features of nodules and the combination of acquired density indicators and morphological features for distinguishing ILs and IACs, respectively. The results are shown in Tables 8,9. In the two models, the density indicators were all independent predictors for ILs [cutoff value >17.22% for threshold ≥−350 HU; odds ratio (OR) =10.349; P<0.001] and IACs (cutoff value >5.64% for threshold ≥−250 HU; OR =10.870; P<0.001). After incorporating these independent density indicators, the AUCs of the models for predicting ILs and IACs based solely on morphological features increased from 0.794 to 0.849 and from 0.843 to 0.902, respectively (both P<0.001) (Figures 2,3).

Figure 2 Models A and B are the ROC curves of the ILs predictive model established by the traditional morphological features (Model A) and the combination of the traditional morphological features and optimal density indicator (Model B). The AUC increased from 0.794 (Model A) to 0.849 (Model B). AUC, area under the curve; CI, confidence interval; IAC, invasive adenocarcinoma; IL, invasive lesion (MIA and IAC); MIA, minimally invasive adenocarcinoma; ROC, receiver operating characteristic.

Figure 3 Models A and B are the ROC curves of the IACs predictive model established by the traditional morphological features (Model A) and the combination of the traditional morphological features and optimal density indicator (Model B). The AUC increased from 0.843 (Model A) to 0.902 (Model B). AUC, area under the curve; CI, confidence interval; IAC, invasive adenocarcinoma; ROC, receiver operating characteristic.

Comparison of diagnostic efficacy between density threshold proportion and radiological features of GGNs for predicting ILs and IACs

Tables 10,11 show the comparison of diagnostic efficacy between density threshold proportion and radiological features of GGNs for predicting ILs and IACs. The AUC values of density threshold proportion (≥−350 HU) and size for predicting ILs are similar (0.801 vs. 0.820, P=0.32), and which are significantly higher than those of other CT features (each P<0.05). Additionally, the AUC values of density threshold proportion (≥−250 HU) and size for predicting IACs are also similar (0.882 vs. 0.869, P=0.45), and which are significantly higher than those of other radiological features (each P<0.05).

Discussion

This retrospective study explored the value of AI-based component analysis in predicting the invasiveness of neoplastic GGNs. The results indicated that AI-based density threshold proportion analysis could effectively identify ILs and IACs, and its predictive performance was comparable with that of nodule size. For determining the ILs and IACs, the optimal density thresholds and the corresponding cutoff values of the proportions of components with density higher than the threshold were ≥−350 HU and 17.22% for ILs and ≥−250 HU and 5.64% for IACs in the training cohort, which were reproducible in the validation cohort and cross-validation. Upon combining these independent density indicators, results revealed that the ability of the traditional morphological features of nodules improved significantly in evaluating ILs and IACs.

Previous studies frequently adopted both quantitative and qualitative CT features for evaluating the invasiveness of GGNs and revealed their correlations (10,17,34-38). It was revealed that the lesion size was primarily associated with the invasiveness of GGNs rather than other radiological features (39-42). In this study, the density indicator was confirmed as a new independent predictor for ILs or IACs, which could significantly improve the ability of the traditional morphological features in evaluating the invasiveness of neoplastic GGNs. This finding can be attributed to the concurrent increase in density, size, and invasiveness of GGNs as they progress from AIS to IAC (29,43-46). Therefore, when evaluating the invasiveness of neoplastic GGNs, it is crucial to prioritize the density indicator and size over other morphological features. While the predictive ability of the density indicator in this study did not significantly outperform that of size, they are not mutually substitutable due to their different focuses. On the contrary, they should be complementary, because the smaller GGNs may also contain solid components or present as heterogeneous lesions, and their invasiveness may be underestimated by size. Additionally, density proportion analysis offers a distinct advantage in follow-up scenarios where the small GGNs demonstrate no substantial size changes but exhibit increased density, or vice versa.

Li et al. (10) reported that the cutoff values of 2D average CT values of lesions for determining ILs and IACs were −548 HU (AUC: 0.728) and −364.89 HU (AUC: 0.874) in 216 GGNs, respectively. Meanwhile, Zhang et al. (47) reported that cutoff values for ILs and IACs were −536.5 HU (AUC: 0.675) and −493.5 HU (AUC: 0.589) in 657 GGNs, respectively. Additionally, 3D density was also used for distinguishing GGNs with different invasiveness (15,21). Kitazawa et al. (15) found that a 3D average CT value of −489 HU was the threshold for diagnosing IAC (AUC: 0.838). Although the 2D or 3D average density could effectively reflect the overall condition, they weakened the heterogeneity of GGNs in density. This limitation is particularly notable in PSNs, where the average density may mask significant density variations, thereby failing to accurately represent the presence and amount of internal solid components that typically indicate invasiveness. In the present study, the density heterogeneity of GGNs was considered, and the higher AUCs indicated that the proportion of the components with different density was more closely associated with the invasiveness.

With the popularity of AI in recent years, AI-based density analysis reflecting heterogeneity has been used in determining GGN invasiveness (18,21,48). A study similar to ours analyzed 48 pGGNs and found that a CT volume density proportion >−300 HU with cutoff value of 5.41% optimally predicted IACs (AUC: 0.933; sensitivity: 85%; specificity: 95%) (48). The study also confirmed the feasibility of the method in the present study for evaluating the invasiveness of GGN. In particular, the pGGNs and PSNs with different degree of invasiveness were evaluated in the present study, and although this method proved effective, its predictive ability was slightly lower. This difference may be attributed to the fact that the heterogeneity in density of PSNs is more complex than that of pGGNs, while not all of the solid components in the PSNs represent invasiveness because some of them are fibrosis, collapsed alveoli, or infiltration of inflammatory cells (49,50). Thus, further study is needed to determine whether this AI-based method is applicable for evaluating PSNs.

The results also revealed that the optimal density threshold for predicting IAC was higher than that for ILs, which confirmed that the degree of invasiveness of GGNs was related to their density. However, the AUCs of some thresholds in predicating ILs or IACs were similar, indicating that not all invasive components in lesions had the highest density. Additionally, this method was more effective in predicting IACs than ILs, and its sensitivity and specificity were respectively lower when distinguishing ILs and IACs. There are two main reasons for these findings. First, as a transitional state in the progression from preinvasive lesions to IAC, MIA possesses some of the characteristics found in both; thus, someone more similar to preinvasive lesions could not be discriminated (low sensitivity for ILs). Second, vessels in the GGNs cannot be accurately removed; thus, the density of internal vessels was seen as a part of the nodule, but the large vessels were not the indicators of IAC (low specificity for IACs). In clinical practice, the density index could be used for accurately identifying the GGNs with invasiveness. However, it is important to note that GGNs predicted as non-ILs should not be automatically classified as AIS. Additionally, while the density index can effectively detect IACs, caution is warranted when assessing the likelihood of IAC in GGNs that have lower density but surround large vessels. In such cases, it is essential to employ other indicators to ensure a comprehensive evaluation.

Currently, the lung nodule AI-assisted diagnostic system has become widely adopted in clinical practice. It significantly reduces image interpretation time, enhances radiologists’ work efficiency, and improves the detection rate of pulmonary nodules. Additionally, the system could automatically analyze lesions, providing reliable diagnostic information to support radiologists. This study demonstrates that the AI analysis system could rapidly perform density analysis for the detected GGNs, offering additional crucial information for further evaluating the invasiveness of neoplastic GGNs. However, these AI-derived density parameters should not be used as standalone diagnostic tools due to the low sensitivity in some conditions; instead, clinicians should interpret them as complementary adjuncts to comprehensive radiological assessment. Compared with existing methods, the predictive ability of density proportion (AUC: 0.801 for ILs, 0.882 for IACs) was comparable to that of radiomics (AUC: 0.89 for ILs, 0.87 for IACs) as reported in previous studies (51,52). Compared with the present method, radiomics is technically complex, requires strict data standardization, confronts challenges related to poor model reproducibility and interpretability, and thus poses difficulties for clinical translation. For GGN management, density proportion analysis can serve as a valuable supplement to current guidelines [e.g., the Fleischner Society and International Association for the Study of Lung Cancer (IASLC) guidelines], which primarily rely on nodule size, solid component size, and longitudinal follow-up changes (53,54). The consolidation-to-tumor ratio (CTR) and volume doubling time (VDT) are conventional metrics for evaluating solid components and volume change of GGN (55), but the CTR is often limited in accurately quantifying diffusely distributed solid components, and VDT may not effectively detect progression of GGNs in density. In contrast, the density proportion method quantifies GGN density via predefined thresholds and offers potential advantages over CTR and VDT in specific clinical scenarios. Despite these advantages, integrating AI analysis system into existing PACS is necessary for improving efficiency in large-scale nodule assessments. Additionally, maintaining such an AI system can be costly, which may limit its adoption in smaller medical institutions with relatively low demand.

There are several limitations in this study. First, we analyzed all types of GGNs collectively without comparing the performance of this method in evaluating the invasiveness of pGGNs and PSNs individually. Given the inherent differences in density and solid component characteristics between pGGNs and PSNs, distinct density thresholds might be required when applying assessment criteria to each subtype. However, due to overlapping density features between the two subtypes, we did not explore subtype-specific density proportion thresholds individually. To avoid unnecessary complexity in clinical evaluation, we opted for a combined analysis of pGGNs and PSNs. This approach still yielded high predictive performance while offering greater simplicity and practical utility for routine clinical application. Second, the GGNs with large vessels, air bronchogram, and vacuole were not excluded. This may inflate measured density proportion, particularly in GGNs with traversing vascular structures, as high-attenuation vascular pixels could be misclassified as pathological solid components, leading to an overestimated higher-density fraction and potential bias in threshold-based invasiveness assessment. However, this impact is likely mitigated by the relatively small vascular volume relative to overall nodule size in most cases, and consistent vascular inclusion across lesions may reduce systematic bias in intergroup comparisons. Future studies could integrate vascular exclusion algorithms to refine segmentation, with sensitivity analyses excluding vessels to further quantify the specific impact on threshold performance. Third, this is a retrospective study that was conducted on a single center. Although the results were derived from a larger sample size, the model was trained and validated solely in a single center with similar imaging conditions, which may restrict the generalizability of the results across different scanners or patient populations. To address this, we plan to conduct a prospective multi-center validation study to verify the model’s robustness, thereby enhancing its clinical translatability.

Conclusions

In conclusion, AI-based density analysis is effective for predicting the invasiveness of neoplastic GGNs, especially when visual solid component assessment is questionable. For determining the ILs and IACs, the optimal density thresholds and cutoff values for the proportions of components with density higher than the threshold were ≥−350 HU and 17.22% for ILs and ≥−250 HU and 5.64% for IACs. These density indicators could significantly improve the ability of morphological CT features in evaluation. The present findings may provide key information for managing GGNs in clinical practice.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1020/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1020/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1020/prf

Funding: This work was supported by the Project of Chongqing Natural Science Foundation (No. CSTB2024NSCQ-MSX0655), and the Senior Medical Talents Program of Chongqing for Young and Middle-Aged from Chongqing Health Commission (to Z.G.C.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1020/coif). The 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 the Ethics Committee of The First Affiliated Hospital of Chongqing Medical University (No. 2019-062) and individual consent for this retrospective analysis was waived.

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. Meng Q, Liu T, Peng H, et al. Construction and validation of a risk stratification model based on Lung-RADS(®) v2022 and CT features for predicting the invasive pure ground-glass pulmonary nodules in China. Insights Imaging 2025;16:68. [Crossref] [PubMed]
2. Hsu WC, Huang PC, Pan KT, et al. Predictors of Invasive Adenocarcinomas among Pure Ground-Glass Nodules Less Than 2 cm in Diameter. Cancers (Basel) 2021;13:3945. [Crossref] [PubMed]
3. Yu Y, Wang N, Huang N, et al. Determining the invasiveness of ground-glass nodules using a 3D multi-task network. Eur Radiol 2021;31:7162-71. [Crossref] [PubMed]
4. Borghesi A, Michelini S, Golemi S, et al. What's New on Quantitative CT Analysis as a Tool to Predict Growth in Persistent Pulmonary Subsolid Nodules? A Literature Review. Diagnostics (Basel) 2020;10:55. [Crossref] [PubMed]
5. Nicholson AG, Tsao MS, Beasley MB, et al. The 2021 WHO Classification of Lung Tumors: Impact of Advances Since 2015. J Thorac Oncol 2022;17:362-87. [Crossref] [PubMed]
6. Hammer MM, Hatabu H. Subsolid pulmonary nodules: Controversy and perspective. Eur J Radiol Open 2020;7:100267. [Crossref] [PubMed]
7. 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]
8. Shi L, Zhao J, Peng X, et al. CT-based radiomics for differentiating invasive adenocarcinomas from indolent lung adenocarcinomas appearing as ground-glass nodules: Asystematic review. Eur J Radiol 2021;144:109956. [Crossref] [PubMed]
9. Mei K, Feng Z, Liu H, et al. Preoperative prediction of pulmonary ground-glass nodule infiltration status by CT-based radiomics combined with neural networks. BMC Cancer 2025;25:659. [Crossref] [PubMed]
10. Li X, Zhang W, Yu Y, et al. CT features and quantitative analysis of subsolid nodule lung adenocarcinoma for pathological classification prediction. BMC Cancer 2020;20:60. [Crossref] [PubMed]
11. Fang W, Zhang G, Yu Y, et al. Identification of pathological subtypes of early lung adenocarcinoma based on artificial intelligence parameters and CT signs. Biosci Rep 2022;42:BSR20212416. [Crossref] [PubMed]
12. Kou J, Gu X, Kang L. Correlation Analysis of Computed Tomography Features and Pathological Types of Multifocal Ground-Glass Nodular Lung Adenocarcinoma. Comput Math Methods Med 2022;2022:7267036. [Crossref] [PubMed]
13. Chu ZG, Li WJ, Fu BJ, et al. CT Characteristics for Predicting Invasiveness in Pulmonary Pure Ground-Glass Nodules. AJR Am J Roentgenol 2020;215:351-8. [Crossref] [PubMed]
14. Yan B, Jiang Y, Fu S, et al. PMILACG Model: A Predictive Model for Identifying Invasiveness of Lung Adenocarcinoma Based on High-Resolution CT-Determined Ground Glass Nodule Features. Tohoku J Exp Med 2025;265:13-20. [Crossref] [PubMed]
15. Kitazawa S, Saeki Y, Kobayashi N, et al. Three-dimensional mean CT attenuation value of pure and part-solid ground-glass lung nodules may predict invasiveness in early adenocarcinoma. Clin Radiol 2019;74:944-9. [Crossref] [PubMed]
16. Ichinose J, Kawaguchi Y, Nakao M, et al. Utility of Maximum CT Value in Predicting the Invasiveness of Pure Ground-Glass Nodules. Clin Lung Cancer 2020;21:281-7. [Crossref] [PubMed]
17. Wang H, Weng Q, Hui J, et al. Value of TSCT Features for Differentiating Preinvasive and Minimally Invasive Adenocarcinoma From Invasive Adenocarcinoma Presenting as Subsolid Nodules Smaller Than 3 cm. Acad Radiol 2020;27:395-403. [Crossref] [PubMed]
18. Gao J, Qi Q, Li H, et al. Artificial-intelligence-based computed tomography histogram analysis predicting tumor invasiveness of lung adenocarcinomas manifesting as radiological part-solid nodules. Front Oncol 2023;13:1096453. [Crossref] [PubMed]
19. Yu Y, Cheng JJ, Li JY, et al. Determining the invasiveness of pure ground-glass nodules using dual-energy spectral computed tomography. Transl Lung Cancer Res 2020;9:484-95. [Crossref] [PubMed]
20. Chen R, Zhang H, Huang X, et al. CT Radiomics-based machine learning approach for the invasiveness of pulmonary ground-glass nodules prediction. Eur J Radiol Open 2025;15:100680. [Crossref] [PubMed]
21. Zhang H, Wang D, Li W, et al. Artificial intelligence system-based histogram analysis of computed tomography features to predict tumor invasiveness of ground-glass nodules. Quant Imaging Med Surg 2023;13:5783-95. [Crossref] [PubMed]
22. Liu K, Li Q, Ma J, et al. Evaluating a Fully Automated Pulmonary Nodule Detection Approach and Its Impact on Radiologist Performance. Radiol Artif Intell 2019;1:e180084. [Crossref] [PubMed]
23. Yang K, Liu J, Tang W, et al. Identification of benign and malignant pulmonary nodules on chest CT using improved 3D U-Net deep learning framework. Eur J Radiol 2020;129:109013. [Crossref] [PubMed]
24. Ni XQ, Yin HK, Fan GH, et al. Differentiation of pulmonary sclerosing pneumocytoma from solid malignant pulmonary nodules by radiomic analysis on multiphasic CT. J Appl Clin Med Phys 2021;22:158-64. [Crossref] [PubMed]
25. Li K, Liu K, Zhong Y, et al. Assessing the predictive accuracy of lung cancer, metastases, and benign lesions using an artificial intelligence-driven computer aided diagnosis system. Quant Imaging Med Surg 2021;11:3629-42. [Crossref] [PubMed]
26. Chen X, Qi Q, Sun Z, et al. Total nodule number as an independent prognostic factor in resected stage III non-small cell lung cancer: a deep learning-powered study. Ann Transl Med 2022;10:33. [Crossref] [PubMed]
27. Huo JW, Luo TY, Diao L, et al. Using combined CT-clinical radiomics models to identify epidermal growth factor receptor mutation subtypes in lung adenocarcinoma. Front Oncol 2022;12:846589. [Crossref] [PubMed]
28. Chen X, Dai C, Peng M, et al. Artificial intelligence driven 3D reconstruction for enhanced lung surgery planning. Nat Commun 2025;16:4086. [Crossref] [PubMed]
29. Li WJ, Lv FJ, Tan YW, et al. Benign and malignant pulmonary part-solid nodules: differentiation via thin-section computed tomography. Quant Imaging Med Surg 2022;12:699-710. [Crossref] [PubMed]
30. Fu G, Yu H, Liu J, et al. Arc concave sign on thin-section computed tomography:A novel predictor for invasive pulmonary adenocarcinoma in pure ground-glass nodules. Eur J Radiol 2021;139:109683. [Crossref] [PubMed]
31. Bankier AA, MacMahon H, Colby T, et al. Fleischner Society: Glossary of Terms for Thoracic Imaging. Radiology 2024;310:e232558. [Crossref] [PubMed]
32. Wang Z, Zhu W, Lu Z, et al. Invasive adenocarcinoma manifesting as pure ground glass nodule with different size: radiological characteristics differ while prognosis remains the same. Transl Cancer Res 2021;10:2755-66. [Crossref] [PubMed]
33. Zhang P, Li T, Tao X, et al. HRCT features between lepidic-predominant type and other pathological subtypes in early-stage invasive pulmonary adenocarcinoma appearing as a ground-glass nodule. BMC Cancer 2021;21:1124. [Crossref] [PubMed]
34. He S, Chen C, Wang Z, et al. The use of the mean computed-tomography value to predict the invasiveness of ground-glass nodules: A meta-analysis. Asian J Surg 2023;46:677-82. [Crossref] [PubMed]
35. Yang Y, Xu J, Wang W, et al. Meta-analysis of the correlation between CT-based features and invasive properties of pure ground-glass nodules. Asian J Surg 2023;46:3405-16. [Crossref] [PubMed]
36. Qi L, Xue K, Li C, et al. Analysis of CT morphologic features and attenuation for differentiating among transient lesions, atypical adenomatous hyperplasia, adenocarcinoma in situ, minimally invasive and invasive adenocarcinoma presenting as pure ground-glass nodules. Sci Rep 2019;9:14586. [Crossref] [PubMed]
37. Li M, Zhu L, Lv Y, et al. Thin-slice computed tomography enables to classify pulmonary subsolid nodules into pre-invasive lesion/minimally invasive adenocarcinoma and invasive adenocarcinoma: a retrospective study. Sci Rep 2023;13:6999. [Crossref] [PubMed]
38. Yang Y, Xu J, Wang W, et al. A nomogram based on the quantitative and qualitative features of CT imaging for the prediction of the invasiveness of ground glass nodules in lung adenocarcinoma. BMC Cancer 2024;24:438. [Crossref] [PubMed]
39. Fu F, Zhang Y, Wang S, et al. Computed tomography density is not associated with pathological tumor invasion for pure ground-glass nodules. J Thorac Cardiovasc Surg 2021;162:451-459.e3. [Crossref] [PubMed]
40. Nelson DB, Godoy MCB, Benveniste MF, et al. Clinicoradiographic Predictors of Aggressive Biology in Lung Cancer With Ground Glass Components. Ann Thorac Surg 2018;106:235-41. [Crossref] [PubMed]
41. Zuo Z, Wang P, Zeng W, et al. Measuring pure ground-glass nodules on computed tomography: assessing agreement between a commercially available deep learning algorithm and radiologists' readings. Acta Radiol 2023;64:1422-30. [Crossref] [PubMed]
42. Zhao H, Zhang X, Xiao L, et al. CT combined with 3D-CTBA to analyze the vasculature around pulmonary ground glass nodules and its value in diagnosis. J Cardiothorac Surg 2025;20:427. [Crossref] [PubMed]
43. Inamura K. Clinicopathological Characteristics and Mutations Driving Development of Early Lung Adenocarcinoma: Tumor Initiation and Progression. Int J Mol Sci 2018;19:1259. [Crossref] [PubMed]
44. Luo W, Ren Y, Liu Y, et al. Imaging diagnostics of pulmonary ground-glass nodules: a narrative review with current status and future directions. Quant Imaging Med Surg 2024;14:6123-46. [Crossref] [PubMed]
45. 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]
46. Gao C, Li J, Wu L, et al. The Natural Growth of Subsolid Nodules Predicted by Quantitative Initial CT Features: A Systematic Review. Front Oncol 2020;10:318. [Crossref] [PubMed]
47. Zhang Y, Wang C, Cheng X, et al. Analysis of the relevance between computed tomography characterization and pathology of pulmonary ground-glass nodules with different pathology types. Turk Gogus Kalp Damar Cerrahisi Derg 2023;31:95-104. [Crossref] [PubMed]
48. Oh J, Piao Z, Cho HJ, et al. CT-based three-dimensional invasiveness analysis of adenocarcinoma presenting as pure ground-glass nodules. Transl Cancer Res 2023;12:765-73. [Crossref] [PubMed]
49. Woo JH, Kim JH, Jeong DY, et al. Differentiation Between Invasive Adenocarcinoma and Focal Interstitial Fibrosis among Persistent Pulmonary Part-solid Nodules: With Emphasis on the CT Morphologic Analysis. J Thorac Imaging 2024;39:335-41. [Crossref] [PubMed]
50. Xu L, Lin S, Zhang Y. Differentiation of adenocarcinoma in situ with alveolar collapse from minimally invasive adenocarcinoma or invasive adenocarcinoma appearing as part-solid ground-glass nodules (≤ 2 cm) using computed tomography. Jpn J Radiol 2022;40:29-37. [Crossref] [PubMed]
51. Chen X, Yao B, Li J, et al. Feasibility of Using High-Resolution Computed Tomography Features for Invasiveness Differentiation of Malignant Nodules Manifesting as Ground-Glass Nodules. Can Respir J 2022;2022:2671772. [Crossref] [PubMed]
52. Bin J, Wu M, Huang M, et al. Predicting invasion in early-stage ground-glass opacity pulmonary adenocarcinoma: a radiomics-based machine learning approach. BMC Med Imaging 2024;24:240. [Crossref] [PubMed]
53. 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]
54. Rami-Porta R, Nishimura KK, Giroux DJ, et al. The International Association for the Study of Lung Cancer Lung Cancer Staging Project: Proposals for Revision of the TNM Stage Groups in the Forthcoming (Ninth) Edition of the TNM Classification for Lung Cancer. J Thorac Oncol 2024;19:1007-27. [Crossref] [PubMed]
55. Cui S, Qi L, Tan W, et al. Development and validation of growth prediction models for multiple pulmonary ground-glass nodules based on CT features, radiomics, and deep learning. Transl Lung Cancer Res 2025;14:1929-44. [Crossref] [PubMed]