A radiomics-based nomogram for preoperatively predicting the invasiveness of nodular lung adenocarcinoma: a multicenter study

Introduction

Lung cancer remains the most prevalent malignancy and the leading cause of cancer-related death globally, with approximately 2.5 million new cases (12.4% of all cancers) and 1.8 million deaths (18.7% of all cancer-related deaths) reported in 2022 (1). Lung adenocarcinoma (LUAD), the most common pathological type of lung cancer, accounts for about 40% cases of the malignancy (2,3). With the widespread adoption of low-dose and high-resolution computed tomography (CT), the detection rate of pulmonary nodules has been significantly increased, especially for small LUAD lesions (4,5). According to the fifth edition of the World Health Organization (WHO) Classification of Thoracic Tumors published in 2021, LUAD includes minimally invasive adenocarcinoma (MIA), defined by an invasive component ≤5 mm, and invasive adenocarcinoma (IAC), characterized by an invasive focus greater than 5 mm (6). This difference in the extent of invasion leads to distinct biological behaviors, which correspond to different treatment approaches and prognostic outcomes (7,8). The precise preoperative prediction of LUAD invasiveness is essential for selecting appropriate surgical strategies, thereby preventing overtreatment and paving the way for personalized medicine in LUAD.

With the rapid development of artificial intelligence (AI) technology, substantial progress has been made in various medical fields, including its applications in lung cancer (9-11). Radiomics, an emerging quantitative medical imaging analysis method, enables the extraction of high-throughput features from radiological images, provides comprehensive information for the analysis of lesions, and allows for accurate preoperative prediction of diseases (12-14). Unfortunately, the clinical reproducibility and translation of radiomics models are limited according to the sources of variability such as different scanners and manual segmentation (15-17). Additionally, traditional clinical models tend to show limited diagnostic performance (18-21). Interestingly, a nomogram model in combination with radiomics features and clinical data may attain a more comprehensive and higher prediction efficiency.

Currently, most studies have focused on single-density pulmonary nodules or have grouped all preinvasive lesions together, whereas the application and comparison of individual versus combined radiomics models for predicting invasiveness (MIA vs. IAC) in nodular LUAD remains largely unexplored (22,23). Our study included all clinically common incidental pulmonary nodules, encompassing lesions with a wide range of densities. In addition, we employed a more refined classification strategy by examining the distinct histologic subtypes of LUAD, rather than using the conventional approach that broadly groups atypical adenomatous hyperplasia, adenocarcinoma in situ, and MIA into a single category. This study constructed a combined model by integrating the clinical and radiomics signatures, and validated its performance in both internal validation cohort and external test cohort, providing a basis for treatment decisions and prognosis evaluation in patients with different invasive subtypes of LUAD. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-aw-1239/rc).

Methods

Patient populations

From January 2020 to June 2024, patients with pulmonary nodules found on CT images were recruited from three centers, including Union Hospital (center I), Jinshan Hospital (center II), and The First Affiliated Hospital of Zhengzhou University (center III). The latter two were included as external validation cohorts based on the following pre-specified criteria: (I) they represent geographically distinct populations within Asia, allowing us to test the generalizability of our model across diverse healthcare settings; (II) they possess high-quality, independently curated CT imaging data and pathological confirmation, meeting our rigorous data integrity standards; (III) their demographic profiles differ somewhat from our internal development cohort, which is ideal for robustly assessing the model’s performance in a broader context. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was reviewed and approved by the ethics committees of Union Hospital (No. 2025-0140), Jinshan Hospital (No. 2025-S101), and The First Affiliated Hospital of Zhengzhou University (No. 2025-KY-2136), and the requirement for informed consent was waived for this retrospective study.

The inclusion criteria were as follows: (I) pathologically diagnosed with MIA or IAC; (II) available preoperative non-contrast-enhanced chest CT examinations performed within one month before surgery; (III) the slice thickness of CT images ≤1.5 mm; (IV) CT-confirmed pulmonary nodules with a maximum diameter less than 30 mm; and (V) possessing corresponding clinical data, including age, gender, smoking history, pathological diagnosis, etc. The exclusion criteria included: (I) poor-quality CT images, such as respiratory or motion artifact; (II) incomplete clinical data, like missing key demographic or pathological variables; (III) pulmonary nodules with a maximum diameter less than 5 mm as the British Thoracic Society (BTS) recommends follow-up for pulmonary nodules with a maximum diameter ≥5 mm (24); and (IV) failure in radiomics feature extraction. Patients from center I were randomly assigned to a training cohort and an internal validation cohort at a 7:3 ratio, while patients from center II/III were included in the external test cohort.

As shown in Figure 1, a total of 611 patients were recruited in this study, comprising 203 males and 408 females, aged 23–84 years. Among them, 405 were IAC, and 206 were MIA. The training cohort consisted of 356 patients (MIA, 129; IAC, 227), the validation cohort included 153 patients (MIA, 45; IAC, 108), and the external test cohort contained 102 patients (MIA, 32; IAC, 70). Figure 2 shows the workflow of the whole study. The CT images with corresponding histopathological sections of MIA and IAC are displayed in Figure 3.

Figure 1 Flowchart of the patient recruitment. Center I, Union Hospital; center II, Jinshan Hospital; center III, The First Affiliated Hospital of Zhengzhou University. CT, computed tomography; IAC, invasive adenocarcinoma; MIA, minimally invasive adenocarcinoma.

Figure 2 Pulmonary nodules on CT images and corresponding pathological subtypes. (A-C) The pulmonary nodule (red arrow) and minimally invasive adenocarcinoma pathology (blue arrows); (D-F) The pulmonary nodule (red arrow) and invasive adenocarcinoma pathology (blue arrows). The original magnification of hematoxylin and eosin staining was 100× (B,E) and 40× (C,F), respectively. CT, computed tomography.

Figure 3 Workflow of the whole study. AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; LR, logistic regression; MLP, multi-layer perceptron; RF, random forest; ROC, receiver operating characteristic; ROI, region of interest.

Image acquisition

The lung CT scans were performed using the following scanners at the three centers: SOMATOM Definition AS+ 64-slice CT (Siemens Healthcare, Berlin, Germany), Aquilion ONE 320-slice CT (Canon Medical Systems, Otawara-shi, Tochigi, Japan), and uCT780 80-slice CT (United Imaging Healthcare, Shanghai, China) at center I; Discovery CT750 HD 64-slice CT (GE Healthcare, Boston, MA, USA), iCT 256 128-slice CT (Philips Healthcare, Amsterdam, Netherlands), and uCT780 80-slice CT (United Imaging Healthcare) at center II; SOMATOM Perspective 64-slice CT (Siemens Healthcare) at center III. During scanning, the head-first position was adopted, and patients were instructed to take a deep breath and hold it while scanning was ongoing. The scanning range covered from the lung apex to the area below the diaphragm. The CT parameters for pulmonary scans were as follows: tube voltage of 100–120 kVp, tube current of 100–350 mAs, and matrix size of 512×512. After scanning, the raw data were reconstructed with a thin-section thickness of 1.0–1.5 mm and a 20–30% overlap between slices. Image interpretation relied on the following settings: lung window [window level/window width: −600/1,400 Hounsfield units (HU)] and mediastinal window (window level/window width: 40/400 HU).

Clinical-radiological data acquisition and model construction

Clinical data of patients were collected, including gender, age, smoking history, tumor history (pathologically confirmed history of malignancies except for LUAD), and pathological results (MIA or IAC). Radiological review was performed utilizing the Radiant Digital Imaging and Communications in Medicine (DICOM) Viewer (version 2021.2.2; Medixant). Qualitative data involved the nodule lobe location (left upper lobe, left lower lobe, right upper lobe, right middle lobe, and right lower lobe), density types (pure ground-glass nodules, part-solid nodules, and solid nodules), shape (round/oval, irregular), border (clear, unclear), and features such as spiculation, pleural retraction, lobulation, air bronchogram, and cavitation sign. The data were independently assessed in a double-blind manner by two radiologists with over five years of experience, with discrepancies resolved by discussion. The maximum diameter of the nodule (in millimeters) was measured independently by two radiologists, and its average value was calculated. Intra- and inter-observer agreement was assessed in terms of Cohen’s κ and the intraclass correlation coefficient (ICC) for qualitative and quantitative variables, respectively.

All clinical and radiological features in the training dataset were subjected to univariable logistic regression (LR) analysis to identify variables with P<0.05. The selected variables with statistical differences were subsequently used in a multivariable LR analysis to identify independent risk factors. The optimal combination of independent risk factors was used to construct a clinical prediction model.

Image preprocessing and segmentation

To ensure the consistency of imaging data from different scanners used in the three centers, all chest CT images were first resampled to a uniform voxel size of 1 mm × 1 mm × 1 mm. Window level and window width were standardized (window level/window width: −600/1,400 HU). Two senior radiologists, blinded to the pathological results, manually delineated the region of interest (ROI) using ITK-SNAP (version 3.8.0; www.itksnap.org). ROIs were contoured slice-by-slice on axial CT images, with care taken to avoid surrounding blood vessels, bronchi, and normal lung tissue. Disagreements were resolved through discussion. To ensure reproducibility, 30 randomly selected cases from the three centers were analyzed to estimate the intra- and inter-observer variability using ICC. Specifically, ROIs were re-delineated by the same radiologist on the same cohort after an interval of one month for the intra-observer analysis; ROIs were independently delineated by two radiologists blinded to each other’s annotations for the inter-observer analysis.

Extraction and selection of radiomics features

Having established excellent agreement in ROI segmentation (all ICC >0.8), we proceeded to the radiomics feature extraction section. These extracted radiomics features were categorized into three main types: the first-order features, shape features, and texture features. The texture features were further divided into the following matrices: the neighborhood gray-tone difference matrix (NGTDM), gray-level size zone matrix (GLSZM), gray-level run-length matrix (GLRLM), gray-level dependence matrix (GLDM), and gray-level co-occurrence matrix (GLCM). The first-order features represent the distribution of individual voxel intensities in the ROI histogram. Shape features quantify the geometric size and dimensions of the ROI. Texture features, derived from the various matrices, describe the spatial relationships between voxels.

Following feature extraction, the dimensionality reduction and selection of radiomics features were performed. Firstly, Z-score normalization and statistical testing (the Mann-Whitney U test) were applied to the extracted features, after which features with statistical significance were selected (P<0.05). Secondly, features exhibiting a high degree of correlation (Pearson correlation coefficient >0.90) were removed, with only one representative feature retained from each correlated pair. Thirdly, the minimum redundancy maximum relevance (mRMR) algorithm was utilized to select features that exhibited the strongest correlation with LUAD invasiveness prediction and had the lowest redundancy. Finally, the least absolute shrinkage and selection operator (LASSO) model with 10-fold cross-validation was used to find features with non-zero coefficients, where the optimal penalty parameter λ was selected as the value that gave the most regularized model within one standard error of the minimum cross-validated error.

Radiomics model building and comparison

These selected radiomics features were applied to construct radiomics models through the three conventional machine learning algorithms: LR, random forest (RF), and multi-layer perceptron (MLP), which were selected for comparison based on methodological considerations and common practice in radiomics studies. LR served as an interpretable linear baseline to assess the intrinsic discriminative ability of the extracted features. RF was included as a representative traditional non-linear model with strong capacity for modeling feature interactions and robustness to multicollinearity. MLP was chosen as a basic deep learning architecture to explore higher-level non-linear representations, together spanning linear, traditional non-linear, and deep learning approaches for systematic evaluation. In the modeling phase, these machine learning algorithms were implemented. For reproducibility, three representative models are reported here. The LR model was trained without regularization (penalty=‘none’) and with max_iter=100. The RF classifier was configured with n_estimators=50, max_depth=3, and min_samples_split=4. The MLP classifier employed a feedforward architecture with hidden_layer_sizes= [61, 128, 64, 32] and was trained using stochastic gradient descent (solver=‘sgd’) for up to max_iter=300. A fixed random_state=0 was used for the RF and MLP models to ensure reproducibility. After the construction of radiomics models, the performances of these models were compared and assessed, and the best-performing one was selected for the downstream research.

Development and validation of the combined signature model

A combined prediction model was established by integrating the selected radiomics signature and the clinical signature, with a nomogram generated to visualize the risk scores. The model performance was subsequently evaluated across multiple cohorts, and the diagnostic efficacy was ultimately validated in both internal and external test cohorts.

Statistical analysis

Statistical analysis was conducted by employing SPSS software (version 26.0.0.0; IBM). Categorical data were expressed as counts and percentages, and comparisons were made using the chi-square test, Yates’ continuity correction, or Fisher’s exact test. Continuous data were presented as mean ± standard deviation (SD). The Shapiro-Wilk test was used to assess normality, and Levene’s test was used to analyze variance homogeneity. For continuous data with a normal distribution and homogeneous variance, comparisons were conducted by utilizing the t-test. For all others, the non-parametric Mann-Whitney U test was applied. A P<0.05 was considered statistically significant.

The individual risk probability was visualized by a nomogram, which integrated the radiomics and clinical signatures. Receiver operating characteristic (ROC) curves were used to evaluate the diagnostic performance of all models across different cohorts. Additionally, ROC curve analysis was performed to calculate the area under the curve (AUC) with its 95% confidence interval (CI), accuracy, specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV). The DeLong test was conducted to assess AUC differences. The calibration ability of the model was evaluated against calibration curves, supplemented by the Hosmer-Lemeshow test to display its goodness-of-fit. Decision curve analysis (DCA) was performed to evaluate the clinical utility of models.

Statistical visualizations were performed in Python (version 3.12.4) through the OnekeyAI platform (version 4.9.1). Radiomics feature extraction was performed via PyRadiomics (version 3.0.1). Machine learning implementations were facilitated by using Scikit-learn (version 1.0.2). Statistical assessments were conducted with Statsmodels (version 0.13.2).

As a retrospective model development study, the sample size was determined by the available database. To enhance model robustness and reduce overfitting, we followed the commonly used rule-of-thumb that the number of predictors retained in the final model should not exceed one-tenth of the number of events in the training set [events per variable (EPV) ≥10]. Given 227 events in the training cohort, we limited the final model to at most 22 predictors.

Results

Clinical data analysis and clinical model construction

As shown in Table S1, the assessment of intra- and inter-observer agreement yielded a good consistency (all ICC and Cohen’s κ >0.8). The clinical-radiological characteristics are detailed in Table 1, and nearly all results demonstrated consistent performance across different cohorts, indicating virtually no bias division. Notably, these variables (maximum diameter, density type, spiculation, pleural retraction, and lobulation) consistently exhibited significant differences across all cohorts (P<0.05). As shown in Table 2, univariable LR analysis revealed that gender, age, smoking history, maximum diameter, lobe location, shape, cavitation, density type, lobulation, pleural retraction, air bronchogram, and spiculation were significantly associated with differences between the IAC and MIA groups in the training cohort (P<0.05). Based on these selected variables, multivariable LR analysis identified age, maximum diameter, density type, lobulation, pleural retraction, air bronchogram and spiculation as independent risk factors for the invasiveness of LUAD (P<0.05). The visualized univariable and multivariable analysis results are listed in Figure S1. The matrix of Spearman’s correlation coefficients among these predictors is presented in Figure S2 (all P>0.05), showing that among the seven pairs of features, 85.71% had absolute correlation coefficients below 0.3, with only one pair exhibiting a weak correlation (|r|=0.348). No highly correlated features were retained in the final feature selection process to reduce the risk of multicollinearity. Finally, a clinical prediction model was constructed for further research.

Radiomics model development and comparison

A total of 1,834 radiomics features were initially extracted, with their distribution and statistical analysis shown in Figure S3. The mRMR algorithm identified 30 features, and subsequent LASSO regression (Figure 4) further refined the model by retaining 17 non-zero-coefficient predictors of invasiveness. Collectively, the three representative radiomics models were developed and their performances were evaluated. The predictive performances of LR, RF, and MLP across different cohorts are shown in Table 3, and the corresponding ROC curves are illustrated in Figure 5. In the training cohort, LR, RF, and MLP models demonstrated excellent diagnostic performances, with AUCs of 0.953 (95% CI, 0.933–0.974), 0.964 (95% CI, 0.947–0.981), and 0.957 (95% CI, 0.939–0.975), respectively. For the internal validation part, AUCs of LR, RF, MLP were 0.876 (95% CI, 0.821–0.930), 0.890 (95% CI, 0.841–0.940), and 0.884 (95% CI, 0.832–0.936), respectively. For the external test part, AUCs are 0.873 (95% CI, 0.802–0.944), 0.874 (95% CI, 0.800–0.949), and 0.889 (95% CI, 0.819–0.960), respectively. Importantly, the RF model achieved the highest AUC in the training cohort compared to the LR model (P=0.04, DeLong test) and the MLP model (P=0.03, DeLong test), showing strong discriminative ability. Additionally, the RF model demonstrated consistently strong performance, attaining the highest AUCs in both the training and internal validation cohorts. Moreover, the RF model is more suitable for clinical application due to its superior interpretability and robustness. Figure S4 are the histograms of RF model prediction and their distinguishing performances (via confusion matrices) in each cohort. The Hosmer-Lemeshow test further confirmed that no statistically significant evidence of miscalibration was detected, with a P being 0.055 in the training cohort, 0.278 in the validation cohort, and 0.338 in the test cohort (all P>0.05 against the ideal calibration line). Therefore, the RF model was selected as the best machine learning model for the subsequent studies.

Figure 4 Selection of radiomic features with LASSO regression analysis. The 10-fold cross-validated coefficients (A) and MSE (B); the histogram of non-zero radiomic features (C). LASSO, least absolute shrinkage and selection operator; MSE, mean standard error.

Figure 5 Performance evaluation and comparison of radiomic models. The assessment of ROC curves for different radiomic models in the training (A), validation (B), and test (C) cohorts, respectively. AUC, area under the curve; CI, confidence interval; LR, logistic regression; MLP, multi-layer perceptron; ROC, receiver operating characteristic.

Combined model development and assessment

The combined nomogram is displayed in Figure 6. The predictive performances of the clinical model, radiomics model, and combined model across different cohorts are shown in Table 4, with their ROC, DCA, and calibration curves shown in Figure 7. The combined model exhibited an excellent diagnostic performance in all cohorts. Additionally, the combined model produced the highest AUC compared to the clinical model and radiomics model alone, with the AUC being 0.968 (95% CI, 0.953–0.984), 0.902 (95% CI, 0.856–0.949), and 0.899 (95% CI, 0.839–0.959) in the training, validation, and test cohorts, respectively. Statistical comparisons of the AUCs were performed using the DeLong test. When the combined model was compared to the clinical model, the tests yielded P values of <0.001, 0.34, and 0.005 in the training, validation, and test cohorts, respectively. In comparison with the radiomics model, the corresponding P values were 0.71, 0.57, and 0.07 across the same cohorts. Beyond statistical significance, the combined model offers clinical value by integrating multidimensional information to enable a more comprehensive risk assessment, thereby supporting refined clinical decision-making, such as the identification of intermediate-risk cases, even without a substantial increase in AUC. The DCA curves indicated that the combined model provided more clinical benefits than the other models. Specifically, the DCA demonstrated that across a broad and clinically relevant threshold probability range of 4% to 35%, the prediction strategy based on our model yielded a higher net clinical benefit than strategies of intervening in all patients or in none. At the commonly used clinical decision threshold of 10%, application of the model could reduce approximately 15 unnecessary invasive procedures per 100 patients evaluated while avoiding missed diagnoses. Therefore, the combined model enabled the prediction of LUAD invasiveness with accuracy and robustness.

Figure 6 The nomogram of the combined model in the training cohort.

Figure 7 Comprehensive evaluation of predictive models across cohorts. Comparison of ROC curves among clinical, radiomic, and combined models in the training (A), validation (B), and test (C) cohorts; calibration curves of clinical, radiomic, and combined models in the training (D), validation (E), and test (F) cohorts; DCA curves of clinical, radiomic, and combined models in the training (G), validation (H), and test (I) cohorts, respectively. AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; ROC, receiver operating characteristic.

Discussion

Predicting the invasiveness of LUAD (MIA and IAC) with preoperative CT examinations plays a critical role in guiding treatment strategies and evaluating prognosis (25). This study constructed a nomogram by combining clinical signature and radiomics signature, and it achieved higher predictive accuracy and obtained better clinical utility compared to the standalone clinical model, improving the work of radiologists and clinicians.

Pulmonary nodules have a malignancy rate of up to 12% (26). During clinical practice, radiologists are increasingly encountering pulmonary nodules incidentally detected by CT in asymptomatic patients. The pathological heterogeneity of LUAD results in a wide variety of imaging presentations. Given that assessing the invasiveness of LUAD presenting as pulmonary nodules based on CT radiological features is a routine task for radiologists, we therefore selected common imaging features for the construction of a clinical prediction model. The clinical model demonstrated only a relatively moderate diagnostic performance. Radiomics, as a quantitative tool, allows for high-throughput extraction of data from the entire lesion, thereby rendering it possible to more comprehensively and objectively assess a lesion (27). Among common radiomics features, geometric features provide a three-dimensional representation of the tumor, intensity features reflect the voxel intensity within the tumor, and texture features capture the patterns and higher-order spatial distribution of intensities (28). Radiomics enable the detection of subtle density changes that cannot be spotted by the naked eye (28). In this study, we chose three representative machine learning models: the linear model (LR), tree-based model (RF), and deep learning model (MLP). The modelling and diagnostic effects of these classifiers were comprehensively explored. The results showed that all three models had excellent predictive performances. Comparatively, the RF model performed best. Of note, the combination of radiomics and clinical signatures led to a more excellent predictive performance than conventional clinical model alone, as indicated by a significant elevation of the AUC from 0.905 (95% CI, 0.872–0.937) to 0.968 (95% CI, 0.953–0.984), from 0.870 (95% CI, 0.814–0.926) to 0.902 (95% CI, 0.856–0.949), and from 0.880 (95% CI, 0.814–0.947) to 0.899 (95% CI, 0.839–0.959) in the training, validation, and test cohorts, respectively (all P<0.05).

Notably, these radiomics features selected as predictors lacked morphological traits. In contrast, the clinical model incorporated predictors including spiculation, lobulation, and pleural retraction, which represented critical morphological characteristics. Additionally, precise delineation of spiculated margins and other subtle morphological features during ROI segmentation remains technically challenging. The integration of these complementary datasets effectively compensates for the limitations of each separate model, thereby enabling the development of an optimized and robust model. This combined model comprehensively differentiated between IAC and MIA with enhanced accuracy.

In recent years, an increasing number of studies tried to use AI to predict LUAD invasiveness. For example, some researchers employed radiomics or deep learning models based on nodule density types (e.g., pure ground-glass or part-solid) to predict the invasiveness of LUAD (29-33). Nonetheless, focusing on a single nodule density type is neither comprehensive nor clinically practical, given that various density nodules are frequently encountered. Additionally, the general inclusion of atypical adenomatous hyperplasia, adenocarcinoma in situ, MIA, and IAC is not conducive to precise differentiation. Some other studies evaluated the performance of deep learning or radiomics models in predicting LUAD invasiveness based on intratumoral versus peritumoral regions (25,29,34). Furthermore, several researchers improved the invasiveness assessment of LUAD using quantitative intratumoral heterogeneity scores (35,36). Despite continuous advancements in the research methodologies mentioned above, their predictive power and clinical applicability remain limited. This study established a combined model by integrating clinical-radiological characteristics with radiomics features, visualized by a nomogram, which demonstrated good calibration, excellent diagnostic performance, and more clinical utility.

Although our model demonstrated strong discriminative performance with a significant improvement in AUC compared to the clinical model, its generalizability across different patient cohorts warrants careful consideration, particularly with respect to potential overfitting. To mitigate this risk, several strategies were implemented. First, the external validation cohort was strictly independent of the training cohort in terms of patient source, imaging scanner, and acquisition period, thereby simulating real-world deployment. Second, the model exhibited consistently high performance in both the internal validation cohort (AUC: 0.902; 95% CI, 0.856–0.949) and the external test cohort (AUC: 0.899; 95% CI, 0.839–0.959), with only a minimal AUC difference, suggesting robust generalization rather than cohort-specific overfitting. Collectively, these findings support the robustness and cross-cohort generalizability of the model.

This study is subject to some limitations. First, as a retrospective study, it failed to account for factors such as selection bias, limited population representativeness, and evolving data over time. And it warrants further validation in larger, prospective multi-center cohorts. Second, as a retrospective model development and validation study, we did not conduct a direct, head-to-head performance comparison between our combined model and the latest end-to-end deep learning models. This is primarily because the prerequisites for a fair comparison could not be met within the current research framework, which we will strive to address in future studies. Third, some manual errors were inevitable during manual drawing of ROIs, which may introduce unknown variance into the stability of the extracted feature and could affect the absolute performance metrics of the model. Future studies will employ semi-automated segmentation tools to enhance robustness. Fourth, a key limitation of this study is the relatively small size of the external validation cohort, which may have led to less stable estimates of model performance and limited the feasibility of meaningful subgroup analyses. Nevertheless, this independent validation provides initial evidence of model generalizability. Further validation in larger, prospective, multi-center cohorts is warranted to more accurately characterize performance and assess robustness across clinically relevant subgroups. Fifth, although image resampling and standardized window width and level settings were applied in the radiomics workflow to reduce inter-scanner variability, residual differences inevitably remained due to inconsistencies in image acquisition parameters. Therefore, future prospective studies using predefined and unified acquisition protocols are warranted to further validate our findings. Finally, while the study by Wu et al. (37) proposed establishing a risk prediction model tailored for Asians, our research did not incorporate factors, such as gender or family history of lung cancer, into the inclusion/exclusion criteria. Future studies will place greater emphasis on conducting in-depth investigations based on the unique characteristics of lung cancer screening in Asian populations (37). Additionally, Wu et al. (38) highlighted the problem of overdiagnosis in lung cancer screening using low-dose computed tomography (LDCT) and suggested chest X-ray (CXR) as an initial screening step prior to LDCT. A limitation of the present study is that information on whether patients had undergone prior CXR screening was not available and, therefore, not incorporated into our analysis. Consequently, the potential impact of this screening pathway on model performance could not be assessed. Future work will address this limitation by incorporating comprehensive screening history variables, including prior CXR examinations, to further enhance the model’s generalizability and clinical applicability.

Conclusions

In conclusion, with the integration of clinical and radiomics signatures, our combined model can be employed for the accurate preoperative prediction of the invasiveness of LUAD (IAC and MIA). This robust model demonstrated superior diagnostic accuracy and significant clinical applicability, thereby allowing for reliable noninvasive preoperative prediction. This advancement has laid a foundation for the future development of more precise individualized therapeutic strategies.

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-aw-1239/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-aw-1239/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-aw-1239/prf

Funding: This study was supported by the grants of the National Natural Science Foundation of China (No. 82372069), the National Key R&D Program of China (Nos. 2023YFC2413500, 2024YFC2417805), and the Outstanding Youth Foundation of Hubei Province, China (No. 2023AFA107).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-aw-1239/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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was reviewed and approved by the ethics committees of Union Hospital (No. 2025-0140), Jinshan Hospital (No. 2025-S101), and The First Affiliated Hospital of Zhengzhou University (No. 2025-KY-2136). Informed consent was waived because of the retrospective nature.

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. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Jonas DE, Reuland DS, Reddy SM, et al. Screening for Lung Cancer With Low-Dose Computed Tomography: Updated Evidence Report and Systematic Review for the US Preventive Services Task Force. JAMA 2021;325:971-87. [Crossref] [PubMed]
3. Xiao Z, Long Q, Liao J, et al. The role of SLC2A1 in lung adenocarcinoma: From tumorigenesis to patient survival. PLoS One 2025;20:e0324043. [Crossref] [PubMed]
4. Becker N, Motsch E, Trotter A, et al. Lung cancer mortality reduction by LDCT screening-Results from the randomized German LUSI trial. Int J Cancer 2020;146:1503-13. [Crossref] [PubMed]
5. Wen X, Liu MW, Qiu B, et al. CT-based radiomic consensus clustering association with tumor biological behavior in clinical stage IA adenocarcinoma: a retrospective study. Transl Lung Cancer Res 2024;13:1794-806. [Crossref] [PubMed]
6. 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]
7. 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]
8. Jiang Y, Che S, Ma S, et al. Radiomic signature based on CT imaging to distinguish invasive adenocarcinoma from minimally invasive adenocarcinoma in pure ground-glass nodules with pleural contact. Cancer Imaging 2021;21:1. [Crossref] [PubMed]
9. Trujillo JC, Soriano JB, Marzo M, et al. Cost-effectiveness of a machine learning risk prediction model (LungFlag) in the selection of high-risk individuals for non-small cell lung cancer screening in Spain. J Med Econ 2025;28:147-56. [Crossref] [PubMed]
10. Wulaningsih W, Villamaria C, Akram A, et al. Deep Learning Models for Predicting Malignancy Risk in CT-Detected Pulmonary Nodules: A Systematic Review and Meta-analysis. Lung 2024;202:625-36. [Crossref] [PubMed]
11. Prosper AE, Kammer MN, Maldonado F, et al. Expanding Role of Advanced Image Analysis in CT-detected Indeterminate Pulmonary Nodules and Early Lung Cancer Characterization. Radiology 2023;309:e222904. [Crossref] [PubMed]
12. Ye G, Zhang C, Zhuang Y, et al. An advanced nomogram model using deep learning radiomics and clinical data for predicting occult lymph node metastasis in lung adenocarcinoma. Transl Oncol 2024;44:101922. [Crossref] [PubMed]
13. Ye G, Wu G, Zhang C, et al. CT-based quantification of intratumoral heterogeneity for predicting pathologic complete response to neoadjuvant immunochemotherapy in non-small cell lung cancer. Front Immunol 2024;15:1414954. [Crossref] [PubMed]
14. Kim SY, Woo OH. MRI-based Radiomics for Ductal Carcinoma in Situ: Enhancing Its Role in Risk Stratification. Radiology 2025;315:e250085. [Crossref] [PubMed]
15. Whybra P, Zwanenburg A, Andrearczyk V, et al. The Image Biomarker Standardization Initiative: Standardized Convolutional Filters for Reproducible Radiomics and Enhanced Clinical Insights. Radiology 2024;310:e231319. [Crossref] [PubMed]
16. Huisman M, Akinci D'Antonoli T. What a Radiologist Needs to Know About Radiomics, Standardization, and Reproducibility. Radiology 2024;310:e232459. [Crossref] [PubMed]
17. Quinsten AS, Bojahr C, Nassenstein K, et al. Automated Field of View Prescription for Whole-body Magnetic Resonance Imaging Using Deep Learning Based Body Region Segmentations. Invest Radiol 2025; Epub ahead of print. [Crossref]
18. Shin J, Seo N, Baek SE, et al. MRI Radiomics Model Predicts Pathologic Complete Response of Rectal Cancer Following Chemoradiotherapy. Radiology 2022;303:351-8. [Crossref] [PubMed]
19. Li Y, Yang Z, Lv W, et al. Role of combined clinical-radiomics model based on contrast-enhanced MRI in predicting the malignancy of breast non-mass enhancements without an additional diffusion-weighted imaging sequence. Quant Imaging Med Surg 2023;13:5974-85. [Crossref] [PubMed]
20. Shi W, Hu Y, Wang H, et al. The value of relative CT attenuation in predicting invasiveness in patients with T1-stage lung adenocarcinoma. Quant Imaging Med Surg 2025;15:9294-308. [Crossref] [PubMed]
21. Kunihiro Y, Kameda F, Kobayashi T, et al. Qualitative and quantitative CT analyses of the solid component in lung adenocarcinoma for predicting invasiveness. Jpn J Radiol 2025;43:1489-97. [Crossref] [PubMed]
22. Li H, Luo Q, Zheng Y, et al. A nomogram for predicting invasiveness of lung adenocarcinoma manifesting as ground-glass nodules based on follow-up CT imaging. Transl Lung Cancer Res 2024;13:2617-35. [Crossref] [PubMed]
23. 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]
24. Callister ME, Baldwin DR, Akram AR, et al. British Thoracic Society guidelines for the investigation and management of pulmonary nodules. Thorax 2015;70:ii1-ii54. [Crossref] [PubMed]
25. Shang Y, Zeng Y, Luo S, et al. Habitat Imaging With Tumoral and Peritumoral Radiomics for Prediction of Lung Adenocarcinoma Invasiveness on Preoperative Chest CT: A Multicenter Study. AJR Am J Roentgenol 2024;223:e2431675. [Crossref] [PubMed]
26. Khan T, Usman Y, Abdo T, et al. Diagnosis and management of peripheral lung nodule. Ann Transl Med 2019;7:348. [Crossref] [PubMed]
27. Zhang H, Liu K, Ding Y, et al. Predictive value of artificial intelligence-based quantitative CT feature analysis for diagnosing the pathological types of pulmonary nodules. Eur Radiol 2025; Epub ahead of print. [Crossref]
28. Shen J, Du H, Wang Y, et al. A novel nomogram model combining CT texture features and urine energy metabolism to differentiate single benign from malignant pulmonary nodule. Front Oncol 2022;12:1035307. [Crossref] [PubMed]
29. Wang X, Xue H, Ding W, et al. Peritumoral features for assessing invasiveness of lung adenocarcinoma manifesting as ground-glass nodules. Sci Rep 2025;15:14112. [Crossref] [PubMed]
30. Sun Q, Yu L, Song Z, et al. Deep learning and radiomics fusion for predicting the invasiveness of lung adenocarcinoma within ground glass nodules. Sci Rep 2025;15:29285. [Crossref] [PubMed]
31. Qin L, Zhao L, Li XM, et al. A CT-based radiomics nomogram incorporating adipose tissue to differentiate invasive adenocarcinomas among part-solid pulmonary nodules. BMC Cancer 2025;25:1471. [Crossref] [PubMed]
32. Gao Y, Yan C, Cao Z, et al. A computed tomography (CT)-based ternary classification model for predicting the invasiveness of pure ground-glass nodules. Clin Radiol 2025;90:107081. [Crossref] [PubMed]
33. Wan J, Lin X, Wang Z, et al. Dual-layer spectral detector computed tomography multiparameter machine learning model for prediction of invasive lung adenocarcinoma. Transl Lung Cancer Res 2025;14:385-97. [Crossref] [PubMed]
34. Zuo YQ, Liu Q, Li TZ, et al. Computed tomography radiomics of intratumoral and peritumoral microenvironments for identifying the invasiveness of subcentimeter lung adenocarcinomas. BMC Med Imaging 2025;25:331. [Crossref] [PubMed]
35. Zuo Z, Zeng Y, Deng J, et al. Intratumoral heterogeneity score enhances invasiveness prediction in pulmonary ground-glass nodules via stacking ensemble machine learning. Insights Imaging 2025;16:209. [Crossref] [PubMed]
36. Peng W, Qi W, Li Y, et al. Intratumoral Heterogeneity Scores as Predictors of Invasiveness in Lung Adenocarcinoma Presenting as Pure Ground-Glass Nodules: Insights from Explainable Machine Learning-Based Ternary Classification Models. Technol Cancer Res Treat 2025;24:15330338251365985. [Crossref] [PubMed]
37. Wu FZ, Huang YL, Wu CC, et al. Assessment of Selection Criteria for Low-Dose Lung Screening CT Among Asian Ethnic Groups in Taiwan: From Mass Screening to Specific Risk-Based Screening for Non-Smoker Lung Cancer. Clin Lung Cancer 2016;17:e45-56. [Crossref] [PubMed]
38. Wu YJ, Hung YC, Tang EK, et al. Practical Strategy to Mitigate Overdiagnosis in Asian Low-dose Computed Tomography Lung Cancer Screening. Acad Radiol 2026;33:255-65. [Crossref] [PubMed]