Spread through air spaces is a critical lung cancer feature. The World Health Organization introduced spread through air spaces (STAS) as a distinct lung cancer dissemination pattern in 2015. It refers to microscopic tumor cells -- including micropapillary clusters, solid nests, or single cells -- found in air spaces beyond the tumor's main boundary in lung adenocarcinoma.
STAS has been confirmed as an independent prognostic factor associated with reduced survival and higher recurrence rates. Lobectomy is recommended over sublobar resection for STAS-positive early-stage lung adenocarcinoma because it produces better survival outcomes.
The challenge is that STAS can only be definitively confirmed through postoperative histopathology. No reliable preoperative biopsy or intraoperative frozen section method currently exists to detect it, limiting its impact on surgical planning and treatment decisions.
A non-invasive preoperative tool that accurately predicts STAS status would allow surgeons to choose the appropriate resection extent, define adequate radiotherapy margins, and personalize treatment -- motivating this study's development of a CT-based predictive nomogram.
A dual-center retrospective study. This study enrolled 266 patients with pathologically confirmed lung adenocarcinoma from two medical centers. Center 1 contributed 188 patients divided into a training set (131 patients) and an internal validation set (57 patients) in a 7:3 ratio. Center 2 provided 78 patients as an independent external validation set.
All patients underwent thoracic contrast-enhanced dual-layer spectral detector CT (DLCT) within two weeks before surgery. STAS status was confirmed through postoperative pathological examination, serving as the ground truth for model training and evaluation.
Patients with adenocarcinoma in situ, minimally invasive adenocarcinoma, or special histological subtypes were excluded. Cases with severe CT artifacts, incomplete clinical data, or missing spectral-based images were also excluded to ensure data quality.
Among all 266 patients, the overall STAS-positive rate was 30.5%, consistent with previously published literature reporting rates between 15.7% and 58.4%. Patient characteristics were well-matched across the three cohorts, with the exception of lobulation prevalence in the external validation set.
Spectral CT generates multiple quantitative imaging channels. Dual-layer spectral detector CT produces several image types from a single scan: conventional 120 kVp images, virtual monoenergetic images (VMI) at 40, 65, and 100 keV, iodine density maps, effective atomic number maps, and electron density maps. Each channel provides different tissue contrast and physical property information.
Quantitative parameters measured from these images included the spectral curve slope, normalized iodine density, effective atomic number, and electron density. Two radiologists independently measured these values and averaged them, with disagreements resolved by consensus.
For radiomics, a radiologist manually delineated the 3D tumor region of interest on conventional venous-phase images using ITK-SNAP software. Because all DLCT energy images share the same dimensions, this single delineation was applied to all seven image types. Features including first-order statistics, shape descriptors, and texture metrics were extracted using the Pyradiomics package.
A total of 1,834 radiomic features were initially extracted. Feature selection proceeded in three steps: Mann-Whitney U test filtering (p less than 0.05), Pearson correlation pruning for highly correlated features (above 0.9), and LASSO regression with five-fold cross-validation. Only features with an inter-class correlation coefficient of 0.75 or higher were retained to ensure reproducibility.
Two independent predictors of STAS identified. Univariate logistic regression identified multiple clinical and imaging factors associated with STAS status, including tumor size, consolidation-tumor ratio, nodule type, tumor-lung interface appearance, pleural indentation, effective atomic number, and electron density values.
After multivariate analysis, only two factors remained as independent predictors: tumor-lung interface (OR 3.54; p = 0.035) and electron density value (OR 1.17; p = 0.028). An ill-defined tumor-lung interface and higher electron density both significantly increased the likelihood of STAS.
The clinical-radiological model incorporating these two predictors achieved AUC values of 0.870 in the training set, 0.798 in the internal validation set, and 0.819 in the external validation set -- demonstrating consistent and clinically meaningful predictive performance across institutions.
Electron density was found to outperform consolidation-tumor ratio as an independent predictor, likely because it reflects the physical density of the tumor and provides insights into pathological characteristics. Higher solid components correlate with higher CT values and higher electron density, linking this parameter mechanistically to STAS biology.
Seven radiomics models compared across imaging modalities. Separate radiomics models were built from each of the seven image types. In the training set, AUCs ranged from 0.847 (Zeff-based) to 0.945 (iodine density-based). However, performance in validation sets revealed that training-set rankings did not always generalize, highlighting the importance of external validation.
The VMI 40 keV-based model demonstrated the most consistent performance across all three datasets, achieving AUCs of 0.899 (training), 0.835 (internal validation), and 0.828 (external validation). The iodine density model, despite the highest training AUC of 0.945, dropped to 0.751 in internal validation -- suggesting overfitting.
Low-energy VMI at 40 keV provides superior tissue contrast compared to conventional images because it increases signal-to-noise ratio and enhances iodine attenuation, thereby better capturing tumor angiogenesis and heterogeneity. This mechanistic advantage explains the model's superior generalizability.
The highest-weighted features in the VMI 40 keV model were first-order skewness and Gray-Level Co-occurrence Matrix texture features derived from filtered image transformations. Lower skewness values may serve as imaging biomarkers for STAS by capturing increased lesion density, while GLCM features reflect microstructural patterns linked to tumor aggressiveness.
Combining radiomics with clinical predictors into a nomogram. The final nomogram integrated the VMI 40 keV radiomics signature with the two independent clinical predictors: tumor-lung interface and electron density. This multi-parameter tool was designed to provide a clinician-interpretable probability score for STAS.
The nomogram achieved AUCs of 0.910 (training), 0.868 (internal validation), and 0.848 (external validation). It significantly outperformed the clinical-radiological model in the training set (p = 0.018) and internal validation set (p = 0.046), though the improvement in the external validation set did not reach statistical significance (p = 0.184).
Calibration curves showed good agreement between predicted probabilities and actual STAS status across all three cohorts. The Hosmer-Lemeshow test confirmed no significant deviation between predictions and observations. Decision curve analysis showed that the nomogram provided net clinical benefit across most reasonable risk threshold ranges.
Subgroup analysis showed that the nomogram performed slightly better for part-solid ground-glass nodules (AUC 0.902 in training) than for solid nodules (AUC 0.749). False positives were more common in solid nodules due to high attenuation, while false negatives were more frequent in part-solid lesions with heterogeneous internal architecture.
A non-invasive STAS prediction tool for clinical practice. This study developed and validated a DLCT-based nomogram that integrates conventional CT features, quantitative spectral parameters, and VMI 40 keV radiomic features to preoperatively predict STAS in lung adenocarcinoma. The nomogram demonstrated promising performance and good calibration across two independent institutions.
Key limitations include the relatively small sample size of this retrospective study, which may introduce selection bias. Additionally, all scans were acquired on Philips spectral CT scanners, so performance with other manufacturers' DLCT equipment requires separate validation.
The nomogram's performance was comparable to the VMI 40 keV radiomics model alone, suggesting that the incremental value of adding clinical predictors may be limited when radiomics features already achieve high AUC values. Future large-scale, multi-center studies are needed to confirm external generalizability and subgroup-specific performance.
Despite these limitations, the nomogram represents a practical, clinician-friendly tool that may support preoperative surgical planning by estimating STAS probability, enabling more informed decisions about resection extent and treatment personalization for patients with lung adenocarcinoma.