Lymph node metastasis shapes lung cancer outcomes. In lung cancer, whether cancer has spread to regional lymph nodes is a critical factor in the TNM staging system used by the American Joint Committee on Cancer. The N stage directly determines whether a patient is eligible for surgical resection or requires chemotherapy and radiation therapy instead.
Dramatic survival differences. Patients without lymph node metastasis (LNM) who undergo surgical resection have a 5-year survival rate exceeding 70%. In contrast, patients with confirmed LNM who receive neoadjuvant or adjuvant chemotherapy and radiotherapy have 5-year survival rates of only 26 to 53%. Accurate preoperative staging is therefore essential for appropriate treatment selection.
Limitations of current staging tools. CT scanning typically uses a lymph node short-axis diameter greater than 10 mm as the criterion for metastasis, but enlarged nodes can result from inflammation, reducing specificity. PET-CT offers better metabolic information but has a false-positive rate of 5 to 21%, is expensive, limited in availability, and adds radiation exposure. Invasive procedures like mediastinoscopy are not routinely recommended.
Radiomics as a solution. Radiomics extracts large numbers of quantitative features from standard CT images, capturing tumor characteristics beyond what the human eye can detect. Meta-analyses show CT-based radiomic models achieve sensitivity of approximately 0.84 and specificity of 0.82, with AUC of 0.90 for predicting LNM in non-small cell lung cancer (NSCLC), making it a promising non-invasive staging tool.
Retrospective cohort of 403 lung cancer patients. The study enrolled 403 patients with pathologically confirmed lung cancer treated between March 2020 and March 2025 at Panzhihua Traditional Chinese Medicine and Western Medicine Hospital. Of these, 143 patients (35.5%) had mediastinal lymph node metastases and 260 had no metastases. The vast majority of tumors were adenocarcinoma (369 cases), with smaller numbers of squamous cell carcinoma (23) and small cell lung cancer (7).
Training and test set division. Patients were randomly divided into a training set of 282 patients and an internal test set of 121 patients at a 7:3 ratio. An additional 30 patients from Panzhihua Central Hospital were collected as an independent external test set to evaluate model generalizability across institutions.
Comprehensive clinical feature collection. Clinical variables collected included age, sex, smoking history, and family history. Laboratory values included carcinoembryonic antigen (CEA), CA125, CA199, alpha-fetoprotein, neuron-specific enolase, and squamous cell carcinoma antigen. Pathological data included tumor type and immunohistochemical markers. TNM staging was based on the 9th edition AJCC classification.
Key clinical associations with LNM. Multivariate analysis identified smoking history and T stage as the two clinical variables most significantly associated with lymph node metastasis status. Other variables including sex, age, family history, M stage, clinical stage, tumor location, and CEA levels did not show statistically significant differences between LNM and non-LNM groups.
851 radiomic features extracted per patient. Non-contrast chest CT images were acquired using standardized scanning parameters and imported into 3D Slicer software. A radiologist with 8 years of experience manually delineated the tumor region of interest layer by layer, reviewed by a 10-year-experienced attending physician to ensure an intraclass correlation coefficient of 0.85 or higher. The SlicerRadiomics module extracted 851 features including 14 shape features, 93 texture features, and 744 frequency-domain features.
Three-step dimensionality reduction. Because 851 features would cause severe overfitting in models trained on hundreds of patients, three sequential selection steps were applied: first, variance thresholding removed low-variance features; second, univariate statistical testing filtered out features not significantly associated with LNM; third, LASSO regression was applied to identify the most predictive features while minimizing multicollinearity.
16 radiomic features retained. After selection, 16 features were retained for modeling, spanning multiple feature families including wavelet-transformed gray-level co-occurrence matrix (GLCM) features, gray-level dependence matrix (GLDM) features, gray-level size zone matrix (GLSZM) features, first-order statistical features, and shape features such as sphericity. These were combined into a single radiomic score (RadScore) using LASSO-derived coefficients.
Biological meaning of selected features. The selected features reflect tumor heterogeneity in biologically meaningful ways: first-order intensity features reflect tissue density and cellularity; sphericity captures macroscopic growth patterns related to invasion; texture features quantify regional non-uniformity linked to genomic instability and elevated tumor mutation burden; wavelet features provide multi-scale assessment of microstructural patterns related to proliferation and immune-stromal interactions.
Nine models compared across three algorithms. Three machine learning algorithms were used to build prediction models: logistic regression (LR), support vector machine (SVM), and random forest (RF). For each algorithm, models were built using clinical features alone, radiomic features alone, and a combined clinical-radiomic feature set, resulting in nine total models.
Handling class imbalance. With only 35.5% positive cases, the dataset was class-imbalanced. Two strategies were tested: SMOTE oversampling of the minority class, and setting the class_weight parameter to balanced (which adjusts attention to minority vs. majority classes). The balanced weighting approach outperformed SMOTE, yielding AUC gains of 0.23 in the internal test set and 0.15 in the external test set for the combined SVM model.
Systematic hyperparameter tuning. All three algorithms underwent systematic hyperparameter optimization using GridSearch with internal 5-fold cross-validation. For SVM, the search covered kernel types (rbf, linear, poly), gamma values, and C parameters. For random forest, optimization covered number of estimators, max depth, minimum samples split, and max features settings. This thoroughness ensured fair performance estimates.
Clinical utility evaluation with decision curve analysis. Beyond AUC, decision curve analysis (DCA) was used to measure net clinical benefit across a range of threshold probabilities. A nomogram was also constructed based on the combined LR model to provide a visual, interpretable tool for individual risk prediction mapping T stage, smoking status, and RadScore to estimated metastasis probability.
SVM with combined features achieves highest AUC. Among all nine models, the combined clinical-radiomic SVM model demonstrated the best performance in the training set (AUC = 0.927) and internal test set (AUC = 0.852). The model using radiomics alone also performed well (AUC = 0.921 training, 0.824 internal test), while clinical features alone produced substantially lower AUC values (0.725 training, 0.707 internal test), highlighting the dominant contribution of imaging features.
External validation confirms generalizability. In the external test set from a separate hospital, the combined RF model performed best (AUC = 0.833), with the combined SVM close behind (AUC = 0.812). The radiomics-only SVM also showed strong external performance (AUC = 0.840), suggesting that radiomic features generalize well across institutions even when clinical variables vary.
Superiority over conventional imaging methods. Compared to conventional CT criteria (sensitivity 0.81, specificity 0.55) and 18F-FDG PET (sensitivity 0.77, specificity 0.86), the combined clinical-radiomic model achieved significantly higher specificity (0.93) with comparable sensitivity (0.81). This superior specificity is particularly important for reducing false positives from reactive lymph node enlargement due to inflammation.
Calibration and clinical utility confirmed. The Hosmer-Lemeshow test (p = 0.398) confirmed that the nomogram's predicted probabilities are well-calibrated to actual outcomes. Decision curve analysis showed that the combined model provided higher net clinical benefit than clinical-only or radiomics-only models across threshold probabilities of 0.2 to 0.8, supporting its utility for real-world treatment planning.
SHAP explains individual predictions. SHapley Additive exPlanations (SHAP) analysis was applied to identify which features drove individual model predictions. SHAP assigns each feature a value indicating how much it contributed to pushing a prediction toward or away from a lymph node metastasis diagnosis, making the model's reasoning transparent and interpretable.
Tumor stage and radiomics lead importance rankings. Tumor T stage was identified as the most critical variable for predicting LNM, consistent with the clinical understanding that larger tumors have a higher likelihood of nodal spread. Radiomic features reflecting tumor gray-level distribution and shape ranked second, followed by smoking history, which was the only clinical variable beyond staging to show significant impact.
All features show positive SHAP values. Notably, all key features exhibited positive SHAP values, indicating a synergistic tendency to increase the predicted probability of LNM rather than canceling each other out. This demonstrates a cooperative effect between clinical and radiomic features, reinforcing the rationale for combining both feature types in a single model.
Nomogram for clinical communication. A visual nomogram was constructed based on the combined LR model to facilitate understanding by clinicians unfamiliar with machine learning. The nomogram maps T stage, smoking status, and RadScore to risk points, allowing individual patients to be given estimated metastasis probabilities that can guide shared decision-making about whether to proceed to PET-CT or invasive staging procedures.
Retrospective design introduces potential bias. Because the study retrospectively collected data from existing medical records, it is subject to selection bias in terms of which patients received CT scans and surgical lymph node assessment. This type of bias is inherent to retrospective radiomics studies and means that patient populations may not perfectly represent the broader lung cancer population.
Small external validation cohort. The external test set contained only 30 patients, which, while providing meaningful cross-institution validation, limits the precision of generalizability estimates. A larger and more diverse external cohort from multiple hospitals would be needed to fully establish the model's robustness before clinical adoption.
Feature reproducibility not formally tested. Radiomic feature stability was not formally assessed through test-retest analysis or phantom-based reproducibility testing. Although standardized imaging protocols were used, factors such as respiratory motion, patient positioning, and technical variations could influence feature values. Manual tumor segmentation also introduces potential interobserver variability despite standardization procedures.
Recommended next steps. The authors call for prospective multicenter studies incorporating repeat imaging acquisitions and phantom-based analysis following the Image Biomarker Standardisation Initiative (IBSI) guidelines. Automated segmentation methods would also help address interobserver variability. Future work should also explore deep learning radiomic models and investigate whether adding contrast-enhanced or PET-CT features would further improve the model.
A practical advance in preoperative staging. This study demonstrates that combining non-contrast CT radiomics with two readily available clinical variables (T stage and smoking history) can predict lymph node metastasis with AUC values of up to 0.927, outperforming conventional CT and PET-CT staging criteria. The non-contrast approach is particularly valuable because it avoids contrast agents, eliminates the cost and radiation of PET-CT, and is accessible at most hospitals.
Radiomics adds substantial value beyond clinical data alone. The consistent superiority of models that included radiomic features over clinical-only models demonstrates that CT images contain prognostically relevant information that cannot be captured by visible lymph node size or standard clinical variables. Radiomics effectively mines this hidden imaging data to improve staging accuracy.
Clinical decision support potential. The nomogram and SHAP visualizations make the model interpretable for clinicians, enabling practical use in preoperative planning. High-risk patients identified by the model could be directed to PET-CT or invasive staging, while low-risk patients could potentially avoid unnecessary procedures, improving both clinical efficiency and patient experience.
Foundation for future development. With further validation in larger prospective cohorts, the combined radiomic-clinical model offers a pathway toward a reliable, affordable, and non-invasive routine tool for individualized lung cancer staging and treatment planning, potentially reducing overtreatment and undertreatment that results from inaccurate preoperative lymph node assessment.