A dangerous tumor behavior. Spread through air spaces (STAS) is a pattern of lung cancer invasion where tumor cells migrate beyond the main tumor mass and spread into the surrounding air sacs (alveoli) of the lung. This can occur as micropapillary clusters, solid nests, or individual cells traveling through the airway spaces.
Introduced in 2015. STAS was formally recognized as a pathological category in the 2015 World Health Organization classification of lung tumors. Since then, multiple studies have confirmed that STAS is a significant risk factor for cancer recurrence after surgical removal of early-stage lung adenocarcinoma.
The preoperative problem. STAS can only be definitively diagnosed after surgery, when a pathologist examines the removed tissue under a microscope. This means the surgeon and oncologist often do not know a patient's STAS status when making critical decisions about what type of surgery to perform or what follow-up treatments to recommend.
High recurrence despite surgery. Even after complete surgical removal of stage I lung adenocarcinoma - the earliest, most treatable stage - studies show recurrence rates of 20% to 50%. STAS is a key driver of this high recurrence. A reliable way to predict STAS before surgery would allow physicians to tailor surgical and follow-up treatment strategies to each patient's actual risk level.
Three independent centers. This multicenter retrospective study enrolled 609 patients with pathological stage I lung adenocarcinoma from three hospitals in China: Tianjin Chest Hospital (Center A, which provided both training and one test set), Tianjin Binhai New Area Haibin People's Hospital (Center B), and Qinhuangdao First Hospital (Center C). Using truly independent external institutions strengthens the reliability of validation findings.
Training and testing structure. Patients from Center A treated between 2015 and 2018 (n=226) formed the training set. A separate group from the same center treated in 2019 (n=306) served as Test Set A. Patients from Centers B and C treated between 2019 and 2020 (n=77) formed Test Set B - a genuinely external validation cohort from different institutions with different patient populations and imaging equipment.
Clinical variables collected. Five clinical features were identified from prior research as significantly associated with STAS: maximum tumor diameter, consolidation-to-tumor ratio (CTR, a measure of how solid vs. ground-glass the tumor appears), spiculation (irregular spiky edges), vacuole sign (air pockets within the tumor), and carcinoembryonic antigen (CEA) blood levels. These formed the basis of a clinical comparison model.
Long-term follow-up. Patients from Center A were followed for progression-free survival (PFS) assessment with a median follow-up of 2,330 days in the training set and 1,940 days in Test Set A - up to 10 years of observation. This extended follow-up allowed meaningful survival analysis linking model predictions to actual patient outcomes.
Tumor and peritumoral regions. Rather than analyzing only the tumor itself, the researchers extracted quantitative imaging features from four distinct regions: the tumor and three concentric zones around it extending 3, 6, and 12 voxel units (approximately 3, 6, and 12 mm). This approach tests whether the tissue immediately surrounding a tumor contains useful information about its aggressive behavior.
Massive feature extraction. Using the Pyradiomics software package, 1,133 radiomics features were extracted from each region of interest. These included features capturing first-order statistics, shape characteristics, and various texture measures using gray-level matrices - mathematical tools that quantify patterns of pixel intensity variation that human eyes cannot perceive reliably.
Systematic feature selection. The 1,133 features were first filtered using Spearman correlation analysis to remove redundant features, then ranked by random forest regression for importance. The top 10 features from each of the four regions were selected, yielding 40 final features for model development - a manageable set that still captures information from across the tumor and its surroundings.
Comprehensive model evaluation. A total of 237 model configurations were tested using 15 different machine learning algorithms including LASSO, support vector machines, random forests, gradient boosting, elastic net regression, and neural networks. The final selected model - a ten-fold cross-validated elastic net regression (ENR-Rad) - was chosen based on consistent performance across all three datasets, prioritizing generalizability over mere training set performance.
Radiomics model results. The final radiomics model (ENR-Rad) achieved accuracies of 0.801, 0.866, and 0.831 in the training set and two external test sets respectively. AUC values (a measure of classification ability where 1.0 is perfect and 0.5 is random) were 0.791, 0.829, and 0.807 - consistently strong across all three cohorts, demonstrating robust generalization.
Radiomics vs. clinical model. The clinical model based on traditional imaging features and blood tests (Clinic-LR) performed substantially worse than ENR-Rad in external validation. In Test Set A, the clinical model achieved an AUC of just 0.572 (barely better than chance), while ENR-Rad achieved 0.829 - a statistically significant difference (p less than 0.001). In Test Set B, the gap was similarly large: 0.689 vs. 0.807 (p less than 0.05).
Combined model best overall. Integrating the radiomics prediction with the five clinical features into a combined model (CM) achieved the highest performance of all, reaching AUC values of 0.834 in the training set and 0.874 and 0.894 in the two external test sets. This indicates that radiomics and clinical information capture complementary aspects of STAS risk and work best together.
Decision curve analysis confirms clinical benefit. Beyond accuracy metrics, decision curve analysis evaluated the practical clinical value of each model across a range of decision thresholds. The combined model provided the highest net clinical benefit across all datasets, confirming that it would lead to better clinical decisions than either the radiomics or clinical model alone.
Making AI interpretable. A key limitation of many machine learning models is their "black box" nature - they make predictions without explaining why. This study used SHAP (SHapley Additive exPlanations) analysis to quantify exactly how much each imaging feature contributed to the model's STAS predictions, making the decision process transparent to clinicians.
Wavelet features dominate. SHAP analysis revealed that wavelet-transformed features - mathematical transformations that capture multi-scale texture information at different frequencies and orientations - contributed most strongly to predictions. This indicates that STAS is not detectable from simple visual features but requires mathematically extracting subtle texture patterns that reflect underlying tumor biology.
The top two features. The most important single feature was "wavelet.LLL_glszm_SmallAreaHighGrayLevelEmphasis" from the tumor, which captures small clusters of high-intensity pixels - reflecting dense, heterogeneous regions within the tumor. The second-ranked feature came from the 6 mm peritumoral zone and described small low-intensity clusters, suggesting that even subtle changes in the tissue just outside the tumor boundary contain information about cancer spread.
Both tumor and surroundings matter equally. SHAP rankings showed that tumor features and peritumoral features from all three zones (3, 6, and 12 mm) contributed nearly equally to the model. This validates the study's key methodological innovation - that STAS risk cannot be fully understood by examining only the tumor itself, but requires analysis of the broader tumor microenvironment captured in CT imaging.
STAS predicts survival. Kaplan-Meier survival analysis in the training set confirmed that patients with pathologically confirmed STAS had significantly shorter progression-free survival than STAS-negative patients (p less than 0.001). STAS-positive patients experienced tumor progression at 38.7% vs. 8.3% in STAS-negative patients over the study period, underscoring its importance as a prognostic marker.
Model predictions also stratify survival. Critically, patients classified as high-STAS-risk by the ENR-Rad model also showed significantly worse progression-free survival than model-predicted low-risk patients (p = 0.011 in training, p less than 0.001 in Test Set A). This demonstrates that the model's risk stratification reflects real biological differences - not just statistical patterns in the training data.
Preoperative risk stratification is meaningful. The fact that model-predicted STAS status tracked actual survival outcomes means the tool could meaningfully identify patients before surgery who face higher recurrence risk. These patients might benefit from more aggressive surgical approaches and closer postoperative surveillance.
Implications for treatment decisions. The authors propose a conceptual decision framework: patients predicted STAS-negative might safely undergo less extensive sublobectomy, while those predicted STAS-positive should receive full lobectomy and consideration of adjuvant therapy including chemotherapy or immunotherapy. This is proposed as a framework to inspire future clinical trials, not as a current clinical recommendation.
Surgery type matters for STAS-positive patients. Emerging evidence suggests that patients with STAS-positive early-stage lung adenocarcinoma may have better outcomes with lobectomy (removing an entire lung lobe) compared to sublobectomy (removing a smaller portion). Since current surgical guidelines do not yet specifically account for STAS status, a preoperative STAS prediction tool could help surgeons choose the appropriate surgical extent before operating.
Adjuvant therapy considerations. The 2023 NCCN guidelines and 2024 Chinese lung cancer treatment guidelines both recommend adjuvant chemotherapy for early-stage NSCLC patients with high-risk features, and STAS positivity is increasingly recognized as one such feature. STAS-positive patients may also potentially benefit from immunotherapy, as early data from the KEYNOTE-091 trial showed immunotherapy benefits in high-risk early-stage NSCLC patients.
A framework for personalized treatment. The researchers propose that for pulmonary nodules 3 cm or smaller, preoperative STAS prediction using the combined model could guide surgical planning. Model-predicted STAS-negative patients may safely receive less extensive surgery, while STAS-positive predictions should trigger consideration of lobectomy and postoperative adjuvant therapy. This framework requires prospective validation before clinical implementation.
Bridging the diagnostic gap. The most immediate clinical value of this model is filling the information gap that currently exists between initial CT imaging and final pathological diagnosis. Currently, surgeons must make consequential decisions about surgical extent without knowing STAS status. A validated preoperative prediction model could transform this into an informed, individualized decision.
A validated multicenter advance. This study successfully developed and validated a radiomics-based machine learning model for preoperative STAS prediction in stage I lung adenocarcinoma. Validated in two independent external cohorts including a truly external multi-institution test set, the model demonstrates robust generalizability beyond the training institution - a crucial requirement for clinical utility.
Radiomics outperforms clinical assessment. A key finding is that quantitative imaging analysis substantially outperforms conventional clinical feature-based assessment for STAS prediction. This challenges the assumption that experienced physician evaluation of standard CT features is sufficient for risk stratification and supports broader adoption of radiomics tools in lung cancer workup.
Key limitations. The study has important limitations. All three centers are in northern China, limiting generalizability to other populations. Radiomic features are sensitive to CT scanner parameters and manual segmentation variability, meaning standardization of imaging protocols would be needed for widespread clinical adoption. The retrospective design precludes definitive conclusions about prospective clinical utility.
Path forward. Future work should include prospective validation in diverse populations, development of automated segmentation to reduce operator-dependent variability, and clinical trials specifically examining outcomes in STAS-stratified patient groups. Optimization of the model with larger, more diverse datasets will be essential before clinical translation.