Why lymph node status is critical. In lung cancer, lymph node (LN) metastasis is one of the most important factors determining a patient's prognosis and treatment plan. Whether cancer has spread to nearby lymph nodes determines TNM staging, guides the extent of surgery, and influences whether less invasive procedures are safe options for a given patient.
Standard CT has a significant blind spot. CT scanning - the primary imaging tool used to assess lymph node involvement before surgery - relies heavily on lymph node size. A short-axis diameter over 10 mm on CT is the traditional cutoff for suspecting malignancy. But this approach misses many cases: cancer can spread to normal-sized lymph nodes, and enlarged nodes can be benign. Studies have found LN metastasis in nearly 37% of NSCLC patients even when all nodes appeared small by CT criteria.
Occult lymph node metastasis: a silent problem. Occult lymph node metastasis (OLNM) refers specifically to microscopic cancer spread in nodes that appear normal by conventional CT evaluation. In this study, 21.8% of patients classified as node-negative by CT (cN0) were found to actually have LN metastasis after surgery - meaning roughly 1 in 5 patients undergoing potentially limited surgery based on CT assessment had hidden nodal disease that standard imaging failed to detect.
The surgical stakes. Accurate preoperative LN assessment matters because limited resection surgery (such as segmentectomy) has been shown to offer equivalent outcomes to larger surgery in node-negative early lung cancer, with better lung function preservation. But if occult LN metastasis is missed, a patient receives less surgery than they need. Conversely, extensive lymph node dissection in node-negative patients may actually worsen outcomes with some immunotherapy regimens. Better preoperative prediction would help surgeons choose the right approach for each patient.
844 patients across two cohorts. This retrospective study enrolled 844 patients with peripheral lung cancer who underwent surgery with systematic lymph node dissection between 2019 and 2021. Patients were split 70/30 into a training cohort (591 patients) and a validation cohort (253 patients). All had been classified as node-negative on preoperative CT, making the prediction of hidden metastasis the clinical challenge.
Going beyond the tumor boundary. The key innovation of this study was extending radiomics analysis into the region surrounding each tumor. For each patient, radiologists manually outlined the tumor (the gross tumor volume, or GTV) and then mathematically expanded this boundary outward by 3 mm, 6 mm, and 9 mm to create three successive peritumoral volumes (PTV3, PTV6, and PTV9). Each zone was analyzed separately to determine which distance captured the most useful predictive information.
1,688 features extracted per region. Using specialized software, the researchers extracted 1,688 radiomics features from each of the four regions (GTV, PTV3, PTV6, PTV9). These features quantified aspects of the tumor and its surroundings including shape, internal texture, and patterns of image density at different spatial scales - capturing information invisible to the human eye.
Rigorous feature selection to build predictive models. From 1,688 initial features, a multi-step selection process was applied: features with poor reproducibility between observers were excluded, statistically non-significant features were removed, redundant correlated features were pruned, and then minimum redundancy maximum relevance (mRMR) filtering and LASSO regularization identified the final small set of most predictive features for each region. This disciplined approach reduced each model to 8-10 key features.
Predictive power increases with distance from the tumor. A key finding was that radiomics models performed progressively better as the peritumoral zone expanded. The model based on the 9 mm peritumoral ring (PTV9) outperformed both the 3 mm (PTV3) and 6 mm (PTV6) models in predicting OLNM. This makes biological sense: microscopic cancer cells spreading toward lymph nodes alter the tissue microenvironment in a gradually expanding radius around the main tumor.
PTV9 model: AUC 0.76 in validation. The PTV9 radiomics model alone achieved an area under the curve (AUC) of 0.761 in the training cohort and 0.733 in the validation cohort - significantly better than the clinical model based on standard CT features (AUC 0.676 training, 0.706 validation).
Combining tumor and peritumoral features: the GPTV model. The researchers then built a combined model integrating the radiomics scores from both the tumor interior (GTV) and the 9 mm peritumoral zone (PTV9) into a single GPTV score. This combination outperformed either region analyzed alone, achieving an AUC of 0.799 in training and 0.772 in validation, confirming that information from inside and outside the tumor is complementary.
Histologic subtype influences optimal peritumoral distance. In a subgroup analysis by cancer type, PTV9 was the best-performing peritumoral zone for adenocarcinoma (the most common subtype), while PTV6 performed best for squamous cell carcinoma. This is consistent with known pathology: adenocarcinoma tends to invade more distantly (mean microscopic extension: 2.69 mm) than squamous cell carcinoma (1.48 mm), meaning its biological signature extends further into the surrounding tissue.
Three independent clinical predictors identified. Before adding radiomics, the researchers identified three standard clinical and CT features independently associated with OLNM through multivariable analysis: elevated serum CEA level (a blood-based tumor marker), lobulation sign (an irregular, bumpy tumor border on CT), and a Type II tumor-pleura relationship (tumor in contact with the lung lining). Each carried an odds ratio of approximately 2.4-2.6, meaning patients with these features were more than twice as likely to have hidden LN metastasis.
Combining clinical features with GPTV radiomics: AUC 0.819. When the three clinical predictors were integrated with the GPTV radiomics score into a single clinical-radiomics nomogram, performance improved substantially. The combined model achieved an AUC of 0.819 in the training cohort and 0.801 in the validation cohort - significantly better than the clinical model alone (p < 0.05) and the best single radiomics model.
Sensitivity of 82% in validation. At the optimal cutoff, the clinical-radiomics model achieved 81.8% sensitivity and 69.2% specificity in the validation cohort. This means the model correctly identified 82 out of 100 patients with hidden lymph node metastasis - a meaningful improvement over standard CT assessment, which missed these same patients by definition.
Clinical utility confirmed by decision curve analysis. Decision curve analysis demonstrated that using the clinical-radiomics nomogram provided a higher net benefit compared to either the clinical model alone or treating all patients the same (either always dissecting extensive lymph nodes or never doing so) across most clinically relevant risk thresholds. This confirms that the model would lead to better patient outcomes in real-world practice.
A practical tool for surgical planning. The clinical-radiomics nomogram was designed to be usable with information available before surgery: a routine preoperative CT scan and a blood CEA level. This means the tool can inform surgical decision-making without requiring any additional invasive procedures. Surgeons could use the predicted OLNM probability to decide between limited resection and more extensive surgery with broader lymph node dissection.
Implications for immunotherapy planning. The study notes a nuanced and counterintuitive finding from other research: in patients with lung cancer who later receive immunotherapy, extensive lymph node dissection is associated with worse immunotherapy outcomes. This suggests that for patients planned to receive chemoimmunotherapy, avoiding unnecessary lymph node dissection in truly node-negative patients may matter more than previously recognized - increasing the value of accurate OLNM prediction.
Biological basis for peritumoral radiomics. The peritumoral microenvironment - the tissue immediately surrounding a tumor - reflects important aspects of cancer biology including local invasion, inflammation, and vascular patterns. The fact that the 9 mm zone captures meaningful predictive information supports the view that cancer-related biological changes extend well beyond the visible tumor boundary, and that these invisible changes are detectable through quantitative CT analysis.
Limitations and future directions. As a retrospective single-institution study, the results need validation in multicenter and prospective settings. The manual tumor segmentation required is time-consuming and could introduce variability in practice. Future work should explore automated segmentation methods, integration with PET/CT data, and prospective validation to establish whether OLNM-guided surgical strategy actually improves patient survival outcomes.