Clinical Problem For patients with peripheral lung adenocarcinoma 3 cm or smaller that appears lymph-node-negative on imaging (clinical N0), the decision of whether to perform a full nodal dissection during surgery or a more limited resection hinges on the true risk of occult lymph node metastasis.
Why It Matters Occult nodal metastasis upstages the patient, changes adjuvant therapy decisions, and worsens prognosis. Accurately identifying which small N0 tumors are actually harboring lymph node spread before surgery guides more aggressive surgical and therapeutic planning.
Large Multicenter Study Researchers enrolled 1,132 patients from two centers: 761 in the training set, 327 in an internal validation set, and 44 in an external validation set - a large cohort for a study of clinically N0 small adenocarcinomas.
Published in Scientific Reports 2025 This habitat radiomics study advances the application of intra-tumor heterogeneity analysis to preoperative nodal staging, combining CT-based imaging features with clinical risk factors in a comprehensive prediction model.
K-Means Clustering Each tumor's CT voxels were clustered into three distinct habitat subregions using k-means clustering based on intensity values. These three subregions represent areas of the tumor with different CT attenuation profiles, reflecting different densities of cells, necrosis, and vascularity.
1,834 Features Per Subregion From each of the three habitat subregions, 1,834 radiomic features were extracted, yielding an initial feature space of over 5,500 features per tumor before selection - a comprehensive characterization of intra-tumor heterogeneity.
Feature Selection After intra-class correlation, variance filtering, and LASSO regularization, 28 habitat features and 20 traditional whole-tumor radiomic features were selected for the final combined model, balancing predictive power with model parsimony.
Habitat vs. Whole-Tumor The key hypothesis tested is whether capturing heterogeneity through habitat subregion features provides additional predictive information beyond whole-tumor radiomic features alone - and the results confirmed this with substantial improvement in AUC.
Independent Clinical Predictors Multivariate analysis identified five clinical and radiological factors independently associated with lymph node metastasis: patient age, CT density (solid vs. subsolid), lobulation sign, pleural indentation, and air bronchogram sign.
Lobulation and Spiculation Lobulation refers to an irregular, bumpy tumor margin on CT, indicating rapid and asymmetric tumor growth. Its presence was significantly associated with lymph node involvement, consistent with its established role as a sign of aggressive biology.
Pleural Indentation Pleural indentation, where the tumor pulls the overlying pleura inward on CT, reflects tumor-stroma adhesion and subpleural lymphatic involvement - both of which are pathways for early regional lymphatic spread.
Air Bronchogram Air bronchogram, where air-filled bronchi are visible within the tumor on CT, is typically associated with lepidic growth pattern adenocarcinomas. Its presence in this model correlates with a higher risk of nodal metastasis in small peripheral tumors.
Combined Model AUC The full combined model (habitat radiomics plus clinical features) achieved AUCs of 0.983 on training, 0.950 on internal validation, and 0.877 on external validation - among the highest reported AUCs for preoperative N staging prediction in the literature.
Habitat vs. Standard Radiomics The habitat-only model achieved AUCs of 0.962/0.865/0.853 compared to whole-tumor radiomics alone, confirming that habitat analysis adds value. The combined clinical plus habitat model further improved performance.
Decision Curve Analysis Decision curve analysis confirmed net clinical benefit across a wide range of risk thresholds for the combined model, indicating that using the model to guide decisions about nodal dissection extent would benefit patients compared to treating all cases identically.
Calibration Calibration curves showed that predicted probability of lymph node metastasis closely matched observed proportions across all three cohorts, an important check confirming that the model produces reliable probability estimates rather than just ranked predictions.
Current Standard of Care For clinical N0 NSCLC, guidelines recommend systematic lymph node dissection during surgery. However, this adds operative time, morbidity, and may be unnecessary for truly node-negative cases - prompting interest in tools that can better predict actual nodal status.
High-Risk vs. Low-Risk Stratification Patients identified as high-risk by the combined model could be prioritized for extended nodal dissection and closer postoperative surveillance, while low-risk patients might be candidates for limited resection with sentinel node sampling.
Avoiding Understaging The high sensitivity of the model means fewer high-risk patients are missed. Understaging (treating a node-positive patient as node-negative) has serious consequences including inadequate adjuvant therapy and early recurrence.
Preoperative Staging Enhancement The model can be integrated with PET-CT staging, which currently has limited sensitivity for micrometastatic nodal disease. Combining PET findings with the habitat radiomics score could further improve staging accuracy.
Small External Validation Cohort While the internal validation with 327 patients is robust, the external validation cohort of 44 patients is relatively small. Larger multicenter external validation is needed before recommending clinical implementation.
CT Protocol Dependence Radiomic features, particularly texture features used in habitat analysis, are sensitive to CT acquisition parameters. The study should validate feature robustness across different scanners and reconstruction protocols.
Prospective Validation Study A prospective study in which the model prediction is made before surgery and then compared to actual pathological nodal findings would most rigorously validate the model's clinical accuracy.
Integration with Liquid Biopsy Combining CT habitat radiomics with circulating tumor DNA markers of nodal disease could further improve prediction accuracy, particularly for borderline-risk cases where either tool alone provides insufficient certainty.