Ensemble machine learning classifiers based on computed tomography radiomics for predicting spread through air spaces in lung adenocarcinoma: a multicenter retrospective cohort study with transcriptomic interpretation

Introduction

Lung cancer is the leading cause of cancer-related mortality worldwide, accounting for an estimated 1.8 million deaths annually and approximately 18% of all cancer-related deaths (1,2). Lung adenocarcinoma (LUAD) is the most common histological subtype of lung cancer (1,2). Spread through air spaces (STAS), defined as the presence of tumor cells within the air spaces of the surrounding lung parenchyma beyond the edge of the main tumor, was recognized by the World Health Organization in 2015 as a distinct pattern of invasion in LUAD (3,4). STAS positivity is an independent adverse prognostic factor and is associated with increased local recurrence, particularly after sublobar resection, as well as reduced overall survival, even in stage I disease (5-7). Research suggests that wedge resection and segmentectomy may be oncologically inadequate for STAS-positive tumors, while lobectomy may provide improved outcomes (8-10). Thus, preoperative assessment of STAS is clinically important; however, intraoperative frozen-section analysis has limited sensitivity (34–55%) due to artifacts and sampling limitations (11,12). As a result, STAS is typically confirmed on permanent sections postoperatively, limiting its utility in intraoperative decision-making and highlighting the need for reliable preoperative imaging biomarkers.

Computed tomography (CT) is the primary imaging modality for assessing local tumor extent and resectability in lung cancer (13). In current clinical practice, preoperative STAS risk is commonly inferred from clinicoradiological (CR) features, including nodule size, attenuation pattern, lobulation, and pleural indentation (14-16). However, these assessments demonstrate limited interobserver reproducibility and are influenced by reader experience (17). Radiomics enables high-throughput quantification of imaging features and facilitates computational characterization of intratumoral heterogeneity by converting imaging data into quantitative biomarkers relevant to pathological profiling (18,19). Several studies have reported that CT-based radiomics may assist in predicting STAS in LUAD (20-23). More recent studies have further emphasized the value of multicenter validation, peritumoral radiomics, clinical-radiomics integration, deep learning, and explainable machine learning for STAS assessment (21,22). Multicenter studies have also reported improved STAS prediction using tumor and peritumoral radiomic features combined with SHapley Additive exPlanations (SHAP) and machine learning models, particularly in small pulmonary nodules (24,25). Despite rapid advancements in this field, there remains a need for transparent feature engineering, robust external validation, uncertainty estimation, and biological interpretation of imaging-derived signatures.

To address these limitations, we developed predictive models using a large multicenter cohort and validated them in independent external cohorts. We trained multiple tree-based classifiers, including random forest (RF), gradient boosting decision tree (GBDT), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost), and further evaluated ensemble approaches using hard voting, soft voting, and stacking to improve the robustness of preoperative STAS prediction.

Although radiomics primarily captures macroscopic imaging phenotypes, the molecular and microenvironmental basis of radiomics signatures associated with STAS remains incompletely understood. Thus, we extended our investigation using a radiogenomic framework by integrating matched CT images, whole-slide images (WSIs), and RNA sequencing data from The Cancer Genome Atlas lung adenocarcinoma (TCGA-LUAD) project and The Cancer Imaging Archive (TCIA). This integrative analysis aimed to identify transcriptomic programs associated with radiomics-predicted STAS risk, including pathways related to cell adhesion and metabolic regulation, thereby providing mechanistic context for imaging-based prediction and supporting personalized surgical planning. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0563/rc).

Methods

Local multicenter cohort for radiomic modeling

In this retrospective multicenter study, we collected the preoperative CT images and postoperative pathological diagnoses of patients who underwent surgical resection for LUAD between January 2019 and January 2025 at three medical centers: Xiangtan Central Hospital (Center 1), Liuyang Hospital of Traditional Chinese Medicine (Center 2), and the Affiliated Hospital of Southwest Medical University (Center 3). Patients were included in the study if they met all the following inclusion criteria: (I) postoperative pathological confirmation of LUAD; (II) a nodule size of 5–30 mm on preoperative CT; and (III) preoperative CT performed within 14 days before surgery. Patients were excluded from the study if they: (I) had received preoperative chemoradiotherapy; (II) had synchronous multiple LUAD lesions or distant metastases; and/or (III) had CT images unsuitable for analysis because of artifacts, concomitant atelectasis or pneumonia that precluded tumor segmentation, or poor overall image quality. The patient screening and enrollment process for the local cohorts is summarized in Figure 1.

Figure 1 Flowchart illustrating the patient screening and enrollment process. Center 1: Xiangtan Central Hospital; Center 2: Liuyang Hospital of Traditional Chinese Medicine; Center 3: Affiliated Hospital of Southwest Medical University. CT, computed tomography; STAS, spread through air spaces.

Radiogenomic cohort for transcriptomic analysis

To elucidate the biological mechanisms underlying the radiomic signatures associated with STAS, an independent radiogenomic cohort was established using publicly available data. We retrospectively screened TCGA-LUAD and TCIA datasets to identify patients with matched preoperative CT scans, diagnostic WSIs, and raw RNA sequencing data. To establish a rigorous reference standard for STAS, two senior pathologists independently reviewed the digitized WSIs of all included participants. Following consensus review, 24 eligible cases (10 STAS-positive and 14 STAS-negative) with definitive pathological classification were included in the radiogenomic analysis. The radiomics formula and the radiomics score (Rad-score) threshold derived from the local training cohort were directly applied to TCGA-LUAD/TCIA cohort; no feature reselection, coefficient refitting, or threshold recalibration was performed in the public cohort.

Ethical approval

This study was approved by the institutional review boards of the participating centers [Xiangtan Central Hospital (No. 2021-07-009), Liuyang Hospital of Traditional Chinese Medicine (No. 2022-016), and the Affiliated Hospital of Southwest Medical University (No. KY2020147)], and was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Because of the retrospective design and anonymization of patient data, the requirement for informed consent was waived for the local cohorts. TCGA-LUAD and TCIA datasets are publicly accessible, and were used in strict accordance with their respective data access policies and publication guidelines.

CT image acquisition, segmentation, and feature robustness assessment

All CT images from the participating centers were reconstructed with a slice thickness of 1.5 mm or less. Scanner manufacturers, scanner models, tube voltage, tube current-time product, matrix size, section thickness/interval, and reconstruction kernels are summarized in Table S1. The CT acquisition protocols followed approaches reported in previous studies while reflecting heterogeneous real-world imaging practices across centers, thereby enabling evaluation of the models under multicenter imaging variability (26-29).

Lung nodules were manually segmented on thin-section CT images using ITK-SNAP version 3.6.0 with standard lung window settings [window level, –600 Hounsfield units (HU); window width, 1,500 HU]. The three-dimensional region of interest (ROI) encompassed the entire visible tumor volume on all contiguous slices. Both solid and ground-glass components were included when they were considered part of the lesion. Adjacent vessels, bronchi, atelectatic lung, inflammatory consolidation, and non-tumoral calcified or cavitary components were excluded when distinguishable from the tumor boundary. Initial segmentations were performed by a cardiothoracic radiologist with 5 years of experience and were subsequently reviewed and refined by an attending radiologist with more than 10 years of experience. Both radiologists were blinded to pathologic STAS status, molecular results, and model predictions. Disagreements were resolved by consensus after joint review [Appendix 1 (1.1)].

To assess interobserver robustness, 100 cases were randomly selected and independently re-segmented by a third radiologist. Radiomic features were extracted from the two segmentation sets using identical preprocessing and feature extraction parameters. Interobserver agreement was quantified using the intraclass correlation coefficient (ICC) based on a two-way random-effects, absolute-agreement model. ICC values greater than 0.75 indicated good reproducibility.

Radiomics analysis

Patients were classified into STAS-positive and STAS-negative groups according to the pathological gold standard; a detailed description of the pathological assessment is provided in Appendix 1 (1.2). Radiomic analysis was conducted using a structured four-phase computational pipeline: (I) image standardization, (II) feature quantification, (III) dimensionality optimization, and (IV) predictive modeling. All digital imaging and communications in medicine (DICOM) images were isotropically resampled to a voxel size of 1×1×1 mm³ using B-spline interpolation implemented in PyRadiomics version 3.1.0 under Python 3.9 (30,31). Radiomic features were extracted from original, Laplacian of Gaussian (LoG)-filtered (sigma =2, 3, and 4 mm), and wavelet-transformed images. The extracted feature classes included first-order statistics, shape, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size-zone matrix, gray-level dependence matrix, and neighboring gray-tone difference matrix features. A fixed bin width of 25 HU was used for the gray-level discretization of CT intensities within the tumor ROI. No artificial conversion to an 8-bit 0 to 255 intensity scale was performed, and all voxels within the segmented tumor region were retained according to their original HU values after resampling.

All features were standardized using z-score normalization to ensure comparable feature scales and improve numerical stability during model training. Dimensionality optimization was performed only within the training process to avoid information leakage. Near-zero variance features were first removed in an unsupervised manner. When univariable testing was used, two-sided t tests were adjusted for multiple comparisons using the Benjamini-Hochberg false discovery rate (FDR) procedure. Pearson correlation analysis was then applied to identify redundant features with |r| ≥0.90. For each pair of highly correlated features, the feature with lower univariable discrimination, lower segmentation reproducibility when available, or greater redundancy with other features was excluded.

Finally, the retained features were entered into a least absolute shrinkage and selection operator (LASSO) regression model for feature selection and Rad-score construction. In the five-fold cross-validation framework, feature standardization, variance filtering, FDR-adjusted univariable screening, correlation filtering, and LASSO selection were performed within the training folds only. Parameters learned from the training folds were then applied to the corresponding validation fold (Figure S1). The internal test set, external validation set, and TCGA-LUAD/TCIA radiogenomic cohort were not used for feature selection, hyperparameter tuning, threshold determination, or model adaptation.

The final Rad-score was calculated as a linear combination of the selected radiomic features weighted by their LASSO coefficients:

Rad-score=sum(ωj×Fj),j=1,...,m[1]

Machine learning development, evaluation, and interpretability

To predict STAS in stage I LUAD, the Rad-score was integrated with CR features. Six tree-based classifiers were developed, including RF, GBDT, XGBoost, LightGBM, AdaBoost, and CatBoost (32). Three ensemble approaches were evaluated: hard voting based on majority class assignment, soft voting based on weighted probability averaging (33,34), and stacking. For stacking, the six base classifiers comprised the first layer, and logistic regression served as the second-layer meta-classifier. The meta-classifier was trained using cross-validated out-of-fold predicted probabilities from the base models to minimize information leakage and enhance generalizability (Figure 2).

Figure 2 Comprehensive algorithmic architecture of the stacking classifier. Center 1: Xiangtan Central Hospital; Center 2: Liuyang Hospital of Traditional Chinese Medicine; Center 3: Affiliated Hospital of Southwest Medical University. AdaBoost, adaptive boosting; CatBoost, categorical boosting; GBDT, gradient boosting decision tree; LightGBM, light gradient boosting machine; RF, random forest; XGBoost, extreme gradient boosting.

All models were trained on the training set and evaluated on the internal test set and the external validation set. The final model was selected based on the highest area under the receiver operating characteristic curve (AUC) with consistent performance across the two validation sets. Model performance was assessed using accuracy, AUC, precision, recall, and F1 score (35). Uncertainty was quantified using 1,000 bootstrap resamples. Bootstrap 95% confidence intervals (CIs) were calculated for AUC, accuracy, precision, recall, and F1 score in the internal test and external validation sets (36).

Model interpretability was evaluated using TreeSHAP version 0.42.1, and feature contributions were quantified using SHAP values. The SHAP-guided iterative elimination procedure was applied to the final model-level input variables, including the Rad-score and CR features, rather than to the original high-dimensional radiomic features. Accordingly, the optimization curve and its caption refer to model-level variables rather than the original radiomic features. Ablation experiments were performed to compare the CR-only, radiomics-only, and combined stacking models.

Radiogenomic and transcriptomic analysis

To investigate the biological mechanisms associated with radiomics-predicted STAS risk, raw messenger RNA read counts for the 24 TCGA-LUAD cases were obtained from the Genomic Data Commons portal. Transcriptomic profiles were matched with pathologic STAS status and with high- and low-risk groups defined by the fixed training-set Rad-score threshold. Differential expression analysis was performed using DESeq2 in R version 4.3.0. Differentially expressed genes (DEGs) were defined as those with an absolute log₂ fold change >1 and a Benjamini-Hochberg adjusted P value <0.05. Gene Ontology (GO) biological process and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were conducted using clusterProfiler. Gene set enrichment analysis (GSEA) was performed using the MSigDB Hallmark gene sets to identify pathways associated with radiomics-predicted high-risk STAS.

Statistical analysis

Statistical analyses were performed using R version 4.3.0 with the tableone package for descriptive statistics. The distribution of continuous variables was assessed using the Shapiro-Wilk test. Normally distributed data are presented as mean ± standard deviation, and non-normally distributed data are presented as median (interquartile range). Categorical variables are summarized as frequencies (percentages).

For group comparisons, continuous variables were analyzed using one-way analysis of variance when assumptions of normality and homogeneity were met, as assessed by the Levene test. Otherwise, the Kruskal-Wallis test was used. Categorical variables were compared using the Pearson chi-square test or Fisher’s exact test, as appropriate. Statistical significance was defined as a two-sided P value <0.05. The Radiomics Quality Score (RQS) was assessed according to the RQS 1.0 framework (37).

Results

Baseline patient characteristics

A total of 1,206 patients with pathologically confirmed stage I LUAD were included in this study, of whom 384 (31.84%) were STAS-positive. The participants were recruited from three medical centers. Patients from Centers 1 and 2 (n=965) were randomly allocated to the training set (n=675) and internal test set (n=290) at a 7:3 ratio. Patients from Center 3 (n=241) comprised an independent external validation set to assess model generalizability.

Detailed baseline CR characteristics of the study population are summarized in Table 1. The key baseline characteristics were broadly comparable across the training, internal test, and external validation cohorts.

Radiomics analysis

Initially, 1,239 radiomic features were extracted from the CT images of each segmented lung nodule. After feature screening within the training folds, FDR correction for univariable testing, and correlation filtering, the retained candidate features were entered into a LASSO regression model. Five-fold cross-validation identified 19 noncollinear radiomic features with optimal predictive value (Figure S2). These features comprised shape, first-order intensity, and texture categories derived from original images, LofG-filtered images, and wavelet-transformed images, while maintaining low interfeature redundancy (Figure S3). Detailed feature classes, filter types, coefficients, and directions of association are provided in Table S2.

In the randomly selected 100-case reproducibility subset, all 19 selected radiomic features showed high interobserver robustness, with ICCs ranging from 0.895 to 0.933. The Rad-score calculated from repeated segmentations also showed good reproducibility, with an ICC of 0.889 (Table S3). Waterfall and violin plots demonstrated that the STAS-positive samples generally had higher Rad-scores than the STAS-negative samples across the training, internal test, and external validation sets (Figures S4,S5). Therefore, the Rad-score was established as the primary radiomic biomarker for subsequent modeling. The item-level RQS assessment is reported in Table S4.

Machine learning model performance

Receiver operating characteristic (ROC) curve analysis was used to evaluate the diagnostic efficacy of the developed machine learning models (Figure S6). Among the evaluated frameworks, the stacking ensemble classifier, which integrated the Rad-score with CR features, exhibited the most stable performance. It achieved an AUC of 0.932 (95% CI: 0.904–0.960) in the internal test set and 0.883 (95% CI: 0.843–0.923) in the external validation set. The corresponding accuracies were 0.879 (95% CI: 0.844–0.914) and 0.826 (95% CI: 0.781–0.871), respectively. Detailed bootstrap 95% CIs for all models and metrics are shown in Table 2.

Ablation analysis demonstrated that the combined stacking model outperformed the CR-only and radiomics-only models in both validation sets. In the internal test set, the combined model achieved an AUC of 0.932 compared with 0.873 for the CR-only model and 0.903 for the radiomics-only model. In the external validation set, the combined model achieved an AUC of 0.883 compared with 0.848 for the CR-only model and 0.857 for the radiomics-only model (Table 3).

Model interpretation and feature importance

To interpret model predictions, SHAP values were calculated for all final model-level input variables. Using a SHAP-guided iterative feature elimination process modeled as an optimization problem, the stacking classifier maintained robust discrimination across feature subsets (Figure 3).

Figure 3 SHAP-guided optimization of model-level input variables after integration of the Rad-score and CR features. This procedure was performed after construction of the Rad-score and therefore did not refer to the selection of individual radiomic features used to calculate the Rad-score. AdaBoost, adaptive boosting; AUC, area under the receiver operating characteristic curve; CatBoost, categorical boosting; CR, clinicoradiological; GBDT, gradient boosting decision tree; LightGBM, light gradient boosting machine; Rad-score, radiomics score; RF, random forest; SHAP, SHapley Additive exPlanations; XGBoost, extreme gradient boosting.

The optimization curve plateaued at a feature subset size of three, indicating that the stacking classifier achieved optimal performance using the Rad-score, CT density, and nodule size. Representative case-level visualizations further illustrated individualized model predictions (Figure 4). In the low-risk representative case, the predicted probability of STAS was 0.026 with a Rad-score of −0.424. The SHAP waterfall plot showed that the Rad-score, CT density, and nodule size mainly contributed to lowering the predicted risk. Conversely, in the high-risk representative case, the predicted probability of STAS was 0.984, with a Rad-score of 0.616. The SHAP waterfall plot showed that the Rad-score was the predominant contributor to increased predicted risk, followed by CT density and nodule size. As illustrated by the bar charts and rose diagrams (Figure 5), these three features accounted for 68.5%, 14.9%, and 3.1% of the total SHAP contribution, respectively.

Figure 4 Representative individualized predictions with CT images, tumor outlines, and SHAP waterfall plots. (A) Low-risk case with a predicted STAS probability of 0.026 and a Rad-score of −0.424. The SHAP plot shows that the Rad-score, CT density, and nodule size decreased the predicted risk. (B) High-risk case with a predicted STAS probability of 0.984 and a Rad-score of 0.616. The SHAP plot shows that the Rad-score was the main contributor increasing the predicted risk, followed by CT density and nodule size. CT, computed tomography; SHAP, SHapley Additive exPlanations; STAS, spread through air spaces.

Figure 5 Combined line plot and circular chart visualizing SHAP feature contribution rankings for the stacking classifier. CT, computed tomography; SHAP, SHapley Additive exPlanations.

Collectively, these analyses demonstrated that the stacking model primarily relied on the radiomics-derived Rad-score, supplemented by clinically interpretable CT features, to support individualized STAS prediction.

Radiogenomic interpretation of the radiomics-predicted STAS phenotype

Given that the Rad-score was the predominant predictive feature, the fixed Rad-score formula derived from the training set was applied to the 24 TCGA-LUAD/TCIA cases. Patients were stratified into radiomics-predicted high-risk and low-risk groups using the training-set cutoff determined by the maximum Youden index. The concordance between radiomics-predicted risk strata and pathologic STAS status is summarized in Table S5.

The volcano plot and heatmap of the top 50 DEGs (Figure 6A,6B) revealed transcriptomic differences between radiomics-predicted high-risk and low-risk groups. Functional enrichment analyses of these genes highlighted structural and microenvironmental abnormalities. GO analysis (Figure 6C) revealed enrichment in cell junction assembly and positive regulation of cell junction assembly. Consistently, KEGG analysis (Figure 6D) indicated alterations in cell adhesion molecule (CAM) interactions. These findings suggest a biological link between the imaging-derived high-risk signature and reduced tumor cell cohesion.

Figure 6 Transcriptomic alterations associated with radiomics-predicted STAS in LUAD. (A) Volcano plot of DEGs between high-risk and low-risk groups (|log₂FC| >1, adj. P<0.05). Red: upregulated; blue: downregulated; gray: nonsignificant. (B) Heatmap of top 50 DEGs. Color scale: Z-score. Columns: patients; rows: genes. (C) GO biological process enrichment. Bar length: gene count; color: −log₁₀(P). (D) KEGG pathway enrichment. Dot size: gene ratio; color: −log₁₀(P). (E) GSEA of hallmark gene sets. X-axis: hallmark gene set; Y-axis: NES. High-risk tumors show enrichment of MYC targets, E2F targets, G2M checkpoint, and glycolysis. DEGs, differentially expressed genes; GO, Gene Ontology; GSEA, gene set enrichment analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; NES, normalized enrichment score; STAS, spread through air spaces.

Global GSEA using Hallmark gene sets (Figure 6E) suggested an aggressive biological phenotype in radiomics-predicted high-risk tumors, including enrichment of MYC targets, E2F targets, the G2M checkpoint, and glycolysis. Collectively, this radiogenomic analysis suggests that the imaging-derived high-risk STAS signature reflects an invasive tumor microenvironment characterized by hyperproliferation, metabolic reprogramming, and impaired cellular adhesion.

Discussion

This multicenter study demonstrated that a stacking ensemble classifier integrating the Rad-score and CR features achieved robust performance in the preoperative prediction of STAS in stage I LUAD. Using SHAP-based optimization, the Rad-score, CT density, and nodule size were identified as the primary contributors to the model’s predictive power.

Several CT-based radiomics and machine learning studies have investigated preoperative STAS prediction (20-25). Early radiomics studies showed promising performance but were often limited by modest sample sizes, single-center designs, and limited external validation (20-23). More recent studies have improved model validation and interpretability. Bao et al. developed a multicenter intratumoral and peritumoral CT radiomics-clinical model for STAS prediction in LUAD (21), while Zhang et al. reported a multicohort deep learning-clinical model with SHAP-based interpretation (22). Wang et al. incorporated tumor and peritumoral regions and reported external AUC values ranging from 0.807 for a radiomics model to 0.894 for a combined model (24). Zhang et al. developed an XGBoost model that achieved AUC values of 0.931 in the internal validation cohort and 0.904 in the external validation cohort (25). In this context, our study contributes a large multicenter CT radiomics cohort, a stacking ensemble framework, bootstrap uncertainty estimation, segmentation reproducibility analysis, and an exploratory radiogenomic interpretation of the Rad-score.

SHAP-based interpretability analysis showed that the Rad-score was the dominant predictor, highlighting the value of radiomics for quantifying intratumoral heterogeneity that is not captured by routine visual assessment on CT. The newly added coefficient table shows that several features reflecting texture heterogeneity and gray-level nonuniformity contributed positively to the Rad-score, whereas selected features related to more homogeneous intensity distribution or lower complexity showed negative coefficients. Ablation analyses further supported the complementary value of radiomics and CR features, as the combined stacking model consistently outperformed the CR-only and radiomics-only models in both validation sets.

A key contribution of this study is the radiogenomic interpretation of the Rad-score. By transferring the fixed training-set Rad-score formula and threshold to TCGA-LUAD/TCIA cohort, we showed that radiomics-predicted high- and low-risk groups were concordant with pathologic STAS status in most cases. The transcriptomic findings further indicated dysregulation of cell junction assembly and CAM pathways, providing molecular support for the hypothesis that reduced cell-to-cell cohesion facilitates air-space dissemination. Enrichment of MYC, E2F, the G2M checkpoint, and glycolytic programs suggests an aggressive phenotype characterized by enhanced proliferation and metabolic reprogramming (38-40). The study was strengthened by multicenter validation, feature reduction, the integration of radiomic and CR features, biological interpretation, reporting of discrimination statistics with CIs, and external validation. However, limitations remained, including the lack of prospective registration, phantom-based analysis of scanner robustness, formal cost-effectiveness analysis, and full open sharing of patient-level imaging data (37).

Several limitations warrant consideration. First, the retrospective multicenter design introduced variability in the CT acquisition protocols, although the scanner parameters were summarized and external validation was performed. Second, interobserver reproducibility was evaluated in a 100-case subset rather than in the full cohort. Third, the study focused on conventional intratumoral radiomics and did not explicitly model heterogeneous tumor habitats. Future studies should integrate habitat-based subregion analysis (41,42), ITH score-like descriptors, and other local and global heterogeneity metrics to better characterize intratumoral spatial complexity (43-45). Fourth, peritumoral radiomics was not included. Peritumoral regions may capture complementary information related to tumor invasion, host response, and microenvironmental remodeling; thus, future studies should evaluate peritumoral habitats or combined intratumoral and peritumoral radiomics. Fifth, the radiogenomic analysis was based on a small public cohort of 24 TCGA-LUAD/TCIA cases with matched CT, WSI, and RNA sequencing data; therefore, these transcriptomic findings should be interpreted as exploratory. Sixth, genomic or transcriptomic data were not routinely collected in the local internal and external validation cohorts, as this was a retrospective imaging-pathology study. Finally, prospective validation and outcome analyses are needed before the model can be used to guide surgical decision-making.

Conclusions

We developed and externally validated a stacking ensemble classifier that integrates CT radiomics and CR features for the preoperative prediction of STAS in stage I LUAD. This biologically interpretable framework provides a data-driven basis for personalized surgical planning, and the exploratory radiogenomic findings support a link between imaging-derived high-risk signatures, reduced cellular adhesion, and metabolic reprogramming.

Acknowledgments

We thank all collaborators and participating centers for their contributions to this study.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0563/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0563/dss

Funding: This work was supported by the President Foundation of Nanfang Hospital, Southern Medical University (No. 2024A009).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0563/coif). All authors report that the submitted work was supported by the President Foundation of Nanfang Hospital, Southern Medical University (No. 2024A009), as disclosed in the manuscript funding statement, with no personal payments reported on this form. The authors have no other 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 institutional review boards of Xiangtan Central Hospital (No. 2021-07-009), Liuyang Hospital of Traditional Chinese Medicine (No. 2022-016), and the Affiliated Hospital of Southwest Medical University (No. KY2020147). The requirement for informed consent was waived because of the retrospective design and use of anonymized clinical data.

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, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. 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. Barta JA, Powell CA, Wisnivesky JP. Global Epidemiology of Lung Cancer. Ann Glob Health 2019;85:8. [Crossref] [PubMed]
3. Kadota K, Nitadori JI, Sima CS, Ujiie H, Rizk NP, Jones DR, Adusumilli PS, Travis WD. Tumor Spread through Air Spaces is an Important Pattern of Invasion and Impacts the Frequency and Location of Recurrences after Limited Resection for Small Stage I Lung Adenocarcinomas. J Thorac Oncol 2015;10:806-14. [Crossref] [PubMed]
4. Ibuki E, Yoshida C, Kadota K. Clinicopathological Significance of Spread Through Air Spaces in Lung Cancer. Pathol Int 2025;75:489-503. [Crossref] [PubMed]
5. Han YB, Kim H, Mino-Kenudson M, Cho S, Kwon HJ, Lee KR, Kwon S, Lee J, Kim K, Jheon S, Lee CT, Lee JS, Kook W, Chung JH. Tumor spread through air spaces (STAS): prognostic significance of grading in non-small cell lung cancer. Mod Pathol 2021;34:549-61. Erratum in: Mod Pathol 2021;34:1038.
6. Liu H, Yin Q, Yang G, Qie P. Prognostic Impact of Tumor Spread Through Air Spaces in Non-small Cell Lung Cancers: a Meta-Analysis Including 3564 Patients. Pathol Oncol Res 2019;25:1303-10. [Crossref] [PubMed]
7. Khalil HA, Shi W, Mazzola E, Lee DN, Norton-Hughes E, Dolan D, Corman S, White A, Sholl LM, Swanson SJ. Analysis of recurrence in lung adenocarcinoma with spread through air spaces. J Thorac Cardiovasc Surg 2023;166:1317-1328.e4. [Crossref] [PubMed]
8. Dai ZY, Shen C, Wang X, Wang FQ, Wang Y. Could less be enough: sublobar resection vs lobectomy for clinical stage IA non-small cell lung cancer patients with visceral pleural invasion or spread through air spaces. Int J Surg 2025;111:2675-85. [Crossref] [PubMed]
9. Eguchi T, Kameda K, Lu S, Bott MJ, Tan KS, Montecalvo J, Chang JC, Rekhtman N, Jones DR, Travis WD, Adusumilli PS. Lobectomy Is Associated with Better Outcomes than Sublobar Resection in Spread through Air Spaces (STAS)-Positive T1 Lung Adenocarcinoma: A Propensity Score-Matched Analysis. J Thorac Oncol 2019;14:87-98. [Crossref] [PubMed]
10. Pan H, Chen H, Kong W, Ge Z, Zhang J, Li W, et al. Lobar versus sub-lobar resection for clinical stage IA (≤ 2 cm) lung neuroendocrine tumor with spread through air spaces: a multi-center study. Updates Surg 2025;77:2601-10. [Crossref] [PubMed]
11. Cao H, Zheng Q, Deng C, Fu Z, Shen X, Jin Y. eta l. Prediction of Spread Through Air Spaces (STAS) By Intraoperative Frozen Section for Patients with cT1N0M0 Invasive Lung Adenocarcinoma: A Multi-Center Observational Study (ECTOP-1016). Ann Surg 2024; Epub ahead of print. [Crossref]
12. Almeida GL, Pinto BM, Pinto VM, Tregnago AC, Almeida RF, Pinto Filho DR. Tumor spread through air spaces in lung cancer: prospective analysis of the accuracy of intraoperative frozen section examination. J Bras Pneumol 2024;50:e20240165. [Crossref] [PubMed]
13. Chan HP, Hadjiiski L, Zhou C, Sahiner B. Computer-aided diagnosis of lung cancer and pulmonary embolism in computed tomography-a review. Acad Radiol 2008;15:535-55. [Crossref] [PubMed]
14. Yin Q, Wang H, Cui H, Wang W, Yang G, Qie P, Xun X, Han S, Liu H. Meta-analysis of association between CT-based features and tumor spread through air spaces in lung adenocarcinoma. J Cardiothorac Surg 2020;15:243. [Crossref] [PubMed]
15. Gu Y, Zheng B, Zhao T, Fan Y. Computed Tomography Features and Tumor Spread Through Air Spaces in Lung Adenocarcinoma: A Meta-analysis. J Thorac Imaging 2023;38:W19-29. [Crossref] [PubMed]
16. Tasnim S, Raja S, Mukhopadhyay S, Blackstone EH, Toth AJ, Barron JO, Raymond DP, Bribriesco AC, Schraufnagel DP, Murthy SC, Sudarshan M. Preoperative predictors of spread through air spaces in lung cancer. J Thorac Cardiovasc Surg 2024;168:660-669.e4. [Crossref] [PubMed]
17. Hoye J, Solomon J, Sauer TJ, Robins M, Samei E. Systematic analysis of bias and variability of morphologic features for lung lesions in computed tomography. J Med Imaging (Bellingham) 2019;6:013504. [Crossref] [PubMed]
18. Zuo Z, Zhang G, Lin S, Xue Q, Qi W, Zhang W, Fan X. Radiomics Nomogram Based on Optimal Volume of Interest Derived from High-Resolution CT for Preoperative Prediction of IASLC Grading in Clinical IA Lung Adenocarcinomas: A Multi-Center, Large-Population Study. Technol Cancer Res Treat 2024;23:15330338241300734. [Crossref] [PubMed]
19. Zuo Z, Li Y, Peng K, Li X, Tan Q, Mo Y, Lan Y, Zeng W, Qi W. CT texture analysis-based nomogram for the preoperative prediction of visceral pleural invasion in cT1N0M0 lung adenocarcinoma: an external validation cohort study. Clin Radiol 2022;77:e215-e221. [Crossref] [PubMed]
20. Tao J, Liang C, Yin K, Fang J, Chen B, Wang Z, Lan X, Zhang J. 3D convolutional neural network model from contrast-enhanced CT to predict spread through air spaces in non-small cell lung cancer. Diagn Interv Imaging 2022;103:535-44. [Crossref] [PubMed]
21. Bao CZ, Zhang R, Deng SY, Liu ZW, Chen DH, Sun JS, Hu QG. Joint model based on intratumoral and peritumoral computed tomography radiomics integrated with clinical features for predicting the spread through air spaces in lung adenocarcinoma: a multicenter study. Quant Imaging Med Surg 2025;15:10020-36. [Crossref] [PubMed]
22. Zhang R, Sun JS, Liu ZW, Lin YY, Cai XY, Kang LJ, Guo BL, Song YX, Bao CZ, Chen K, Sun Y, Chen YF, Cai ZP, Xie H, Chen HX, Yang SM, Hu QG. A deep learning-clinical combined model with SHapley Additive exPlanations (SHAP) method for assessing the tumor spread through air spaces in lung adenocarcinoma: a multicohort retrospective study. Quant Imaging Med Surg 2025;15:8833-49. [Crossref] [PubMed]
23. Chen D, She Y, Wang T, Xie H, Li J, Jiang G, Chen Y, Zhang L, Xie D, Chen C. Radiomics-based prediction for tumour spread through air spaces in stage I lung adenocarcinoma using machine learning. Eur J Cardiothorac Surg 2020;58:51-8. [Crossref] [PubMed]
24. Wang Y, Liu X, Zhao X, Wang Z, Li X, Sun D. A radiomics-based machine learning model and SHAP for predicting spread through air spaces and its prognostic implications in stage I lung adenocarcinoma: a multicenter cohort study. Cancer Imaging 2025;25:115. [Crossref] [PubMed]
25. Zhang Z, Zhao Y, Ma YJ, Chen CQ, Li ZY, Wang YK, Zhang SJ, Li HM, Li Y, Tian Y, Tian H. Prediction of STAS in lung adenocarcinoma with nodules ≤ 2 cm using machine learning: a multicenter retrospective study. BMC Cancer 2025;25:417. [Crossref] [PubMed]
26. Zuo Z, Fan X, Zeng Y, Qi W, Liu W, Li W, Liang Q. Topologically distinct 2D and 3D intratumoral heterogeneity scores for preoperatively predicting invasiveness in stage I lung adenocarcinoma: A multicenter study. PLOS Digit Health 2026;5:e0001246. [Crossref] [PubMed]
27. Zuo Z, Zeng Y, Deng J, Lin S, Qi W, Fan X, Feng Y. Intratumoral heterogeneity score enhances invasiveness prediction in pulmonary ground-glass nodules via stacking ensemble machine learning. Insights Imaging 2025;16:209. [Crossref] [PubMed]
28. Zuo Z, Fan X, Zeng Y, Qi W, Liu W, Zhang J. Multiperspective tumor heterogeneity metrics for preoperative prediction of IASLC grading in clinical stage IA lung adenocarcinomas: A multicenter study. Comput Methods Programs Biomed 2026;274:109137. [Crossref] [PubMed]
29. Li Y, Ding J, Wu K, Qi W, Lin S, Chen G, Zuo Z. Ensemble Machine Learning Classifiers Combining CT Radiomics and Clinical-Radiological Features for Preoperative Prediction of Pathological Invasiveness in Lung Adenocarcinoma Presenting as Part-Solid Nodules: A Multicenter Retrospective Study. Technol Cancer Res Treat 2025;24:15330338251351365. [Crossref] [PubMed]
30. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts HJWL. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
31. Long D, Zuo Z, Zhou H, Qi W, Zhang S, Deng J, Yang Z. Preoperative Ternary Classification of Pulmonary Ground-Glass Nodules (AIS/MIA/IAC): ResNet-10 Outperforms Radiomics and Clinicoradiographic Models in Multicenter Study. Technol Cancer Res Treat 2026;25:15330338261423265. [Crossref] [PubMed]
32. Jovel J, Greiner R. An Introduction to Machine Learning Approaches for Biomedical Research. Front Med (Lausanne) 2021;8:771607. [Crossref] [PubMed]
33. Ling T, Zuo Z, Wu L, Ma J, Wang T, Huang M. Predicting neoadjuvant chemotherapy response in locally advanced gastric cancer using a machine learning model combining radiomics and clinical biomarkers. Digit Health 2025;11:20552076251341740. [Crossref] [PubMed]
34. Zheng H, Chen W, Qi W, Liu H, Zuo Z. Enhancing the prediction of the invasiveness of pulmonary adenocarcinomas presenting as pure ground-glass nodules: Integrating intratumor heterogeneity score with clinical-radiological features via machine learning in a multicenter study. Digit Health 2024;10:20552076241289181. [Crossref] [PubMed]
35. Obuchowski NA, Bullen JA. Receiver operating characteristic (ROC) curves: review of methods with applications in diagnostic medicine. Phys Med Biol 2018;63:07TR01. [Crossref] [PubMed]
36. Fan X, Deng J, Feng Y, Qi W, Lin S, Zeng Y, Zuo Z. Robustness and Accuracy of Radiomics Models for Classifying IASLC Grading in Lung Adenocarcinomas: A Comprehensive Analysis of a Large Multicenter CT Database. Technol Cancer Res Treat 2026;25:15330338261429796. [Crossref] [PubMed]
37. Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J, Sanduleanu S, Larue RTHM, Even AJG, Jochems A, van Wijk Y, Woodruff H, van Soest J, Lustberg T, Roelofs E, van Elmpt W, Dekker A, Mottaghy FM, Wildberger JE, Walsh S. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
38. Meng Y, Wang Y, Liu L, Wu R, Zhang Q, Chen Z, Yao Y, Li X, Gong Y, Li H, Wang Z, Liu H. Immunohistochemistry identifies E-cadherin, N-cadherin and focal adhesion kinase (FAK) as predictors of stage I non-small cell lung carcinoma spread through the air spaces (STAS), and the combinations as prognostic factors. Transl Lung Cancer Res 2024;13:1450-62. [Crossref] [PubMed]
39. Wang W, Zhou W, Fan J, Jiang T, Yang G, Song C, Xu S, Luo H, Liu H. Spatial deciphering of the transcriptomic heterogeneity of tumor spread through air spaces in lung cancer. Front Pharmacol 2025;16:1567527. [Crossref] [PubMed]
40. Zhu Z, Li Y, Wang P, Liang S, Zhang H, Yang X, Chen K, Wang J, Wen Z, Lei J, Shang J, Wang Y, Song Y, Song L, Ma J, Jin Z, Meng W. Machine learning-based quantitative prediction of spread through air spaces in primary lung adenocarcinoma using intratumoural heterogeneity scores. Int J Surg 2025;112:9213-24. [Crossref] [PubMed]
41. Zuo Z, Deng J, Ge W, Zhou Y, Liu H, Zhang W, Zeng Y. Quantifying intratumoral heterogeneity within sub-regions to predict high-grade patterns in clinical stage I solid lung adenocarcinoma. BMC Cancer 2025;25:51. [Crossref] [PubMed]
42. Ge W, Fan X, Zeng Y, Yang X, Zhou L, Zuo Z. Exploring habitats-based spatial distributions: improving predictions of lymphovascular invasion in invasive breast cancer. Acad Radiol 2024;31:4317-28. [Crossref] [PubMed]
43. Zhao J, Zhou M, Tan Y, Wei X, Jiang W, Tong L, Yang Z, You H, Chen R, Gong L, Deng J, Zhou Y, Zuo Z, Xu S. Preoperative CT-based topologically distinct intratumoral heterogeneity scores for predicting intratumoral tertiary lymphoid structures and outcomes in hepatocellular carcinoma: A multicenter study. Eur J Surg Oncol 2026;52:111756. [Crossref] [PubMed]
44. Zhou Y, Zuo Z, Zhao J, Tan Y, Deng J, Wei X, Li H, Gong L, Liu L, Zhong L. Development and Validation of Time-to-Event Machine Learning Models for Predicting Disease-Free Survival in Patients with Locally Advanced Colorectal Cancer: A Multicenter Cohort Study. Ann Surg Oncol 2026;33:1288-300. [Crossref] [PubMed]
45. Qi H, Zuo Z, Lin S, Chen Y, Li H, Hu D, Zhou Y, Qi W, Chen H. Assessment of intratumor heterogeneity for preoperatively predicting the invasiveness of pulmonary adenocarcinomas manifesting as pure ground-glass nodules. Quant Imaging Med Surg 2025;15:272-86. [Crossref] [PubMed]