The Clinical Stakes N2 lymph node metastasis (involvement of mediastinal lymph nodes) is the dividing line between potentially curable NSCLC and disease requiring non-surgical treatment. Current guidelines recommend invasive mediastinoscopy or EBUS to confirm N2 status before surgery - procedures that carry procedural risk and require specialized expertise not available at all centers.
Limitations of Current Imaging PET/CT, the most advanced non-invasive staging tool, has pooled sensitivity of only 62% for mediastinal nodal metastasis. This means PET/CT alone cannot reliably rule out N2 disease, forcing invasive confirmation for most surgical candidates. An AI model combining imaging and clinical data could improve non-invasive accuracy and reduce the need for invasive staging in low-risk patients.
Study Approach 472 NSCLC patients who underwent cervical mediastinoscopy were analyzed. Three types of predictive models were developed and compared: logistic regression (standard statistical), ResNet50 deep learning on CT images, and multiple machine learning algorithms using clinicopathological and radiological features.
Best Result Linear SVM achieved the highest accuracy at 95.7% with AUC 93.5% and F1-score 92%, outperforming the logistic regression baseline (accuracy 90.6%, AUC 85.7%). Multiple other ML algorithms (LDA, QDA, Naive Bayes, ANN) also exceeded 90% accuracy, demonstrating robust and consistent superiority of ML over standard statistics.
Patient Population 1,489 patients who underwent standard cervical mediastinoscopy at Istanbul University between 2006 and 2021 were initially screened. Of these, 472 NSCLC patients with complete pre-operative CT within two weeks of surgery and full mediastinal lymph node dissection at stations 2R, 2L, 4R, 4L, and 7 were included. Histopathological confirmation of N2 status was the gold standard.
Outcome Distribution 356 patients (75.4%) had N0 disease and 116 (24.6%) had confirmed N2 metastasis, reflecting the expected prevalence in a mediastinoscopy-selected surgical population. Male sex and right-sided tumors showed significantly higher N2 rates on univariate analysis.
Input Features Features included age, sex, tumor side and lobe, histological subtype (adenocarcinoma vs. non-adenocarcinoma), radiological tumor size, SUVmax of the primary tumor, highest SUVmax of mediastinal N2 lymph nodes, hilar lymph node SUVmax, tumor-to-N2 SUVmax ratio, and long and short axis of the largest N2 lymph node.
Deep Learning Model The ResNet50 deep learning model was trained on raw DICOM CT images adjusted to a mediastinal window (HU -140 to 260), with regions outside the lymph node segmentation mask zeroed out. The model was trained for 100 epochs on a 70/10/20 train-validation-test split. This CT-only model was compared to the tabular ML models using clinical and radiological features.
Machine Learning Rankings Among all tested algorithms, linear SVM scored highest with accuracy 95.7%, AUC 93.5%, and F1-score 92%. LDA, QDA, Gaussian Naive Bayes, and ANN all exceeded 90% accuracy. Random forest and AdaBoost performed slightly below the top tier but still substantially outperformed the logistic regression baseline.
Deep Learning Performance The ResNet50 CT-based model performed worse than the tabular ML algorithms despite using raw imaging data. This counterintuitive result likely reflects the limitation of training a complex deep learning model on a relatively modest dataset of 472 patients - deep learning typically requires thousands to tens of thousands of images to outperform simpler models on structured medical tasks.
Logistic Regression Baseline Standard stepwise logistic regression achieved accuracy 90.6% and AUC 85.7%, representing the conventional statistical benchmark. While this performance is clinically useful in isolation, the ML improvement to 95.7% accuracy is practically meaningful - at a hypothetical false negative rate, even a few percentage points less missed N2 disease translates to fewer patients receiving inappropriate surgery.
Key Discriminating Features Across ML algorithms, SUVmax of N2 lymph nodes, the tumor-to-N2 SUVmax ratio, and lymph node short axis were consistently among the highest-ranked features. These radiological parameters capture the biological signal of metastatic involvement (hypermetabolism and lymph node enlargement) more quantitatively than simple categorical classifications used in traditional staging.
Mediastinoscopy Risk Cervical mediastinoscopy, while the gold standard for N2 staging, carries procedural risks including recurrent laryngeal nerve injury, tracheal injury, hemorrhage, and pneumothorax. Complication rates of 1-3% are reported even at experienced centers. A reliable non-invasive predictor could spare low-predicted-risk patients these procedural risks.
Decision Framework Proposal Patients predicted to be N0 by the SVM model with high confidence could potentially bypass mediastinoscopy and proceed directly to surgery, while those predicted N2 would still require invasive confirmation before deciding against resection. This risk-stratified approach preserves invasive staging for ambiguous or high-risk cases.
Resource Allocation Mediastinoscopy requires general anesthesia, an operating theater, and thoracic surgery expertise. Reducing the number of mediastinoscopies performed annually would free these resources for therapeutic procedures, reduce waiting times, and lower healthcare costs - particularly relevant in health systems with limited thoracic surgery capacity.
Generalizability Considerations The model was trained at a single high-volume thoracic surgery center with a specific mediastinoscopy indication protocol (ESTS guidelines). External validation at centers with different institutional criteria and patient populations is needed before widespread adoption.
Retrospective Single-Center Limitation All patients were from Istanbul University over a 15-year period, creating temporal and institutional biases. Imaging protocols, PET/CT scanners, and surgical criteria evolved over this period, potentially introducing confounding variation. A prospective multicenter validation study is needed to establish generalizability.
Deep Learning Scalability Opportunity The ResNet50 model underperformed tabular ML with the current dataset size. However, with thousands of CT scans from multiple institutions, deep learning could ultimately surpass tabular ML by extracting spatial features invisible to radiologists. A federated learning approach across multiple centers could build a training dataset large enough to unlock deep learning's potential.
Missing Molecular Data The model uses imaging and clinical features but lacks molecular biomarkers such as EGFR mutation status, ALK rearrangements, or liquid biopsy ctDNA levels, which have independent associations with lymph node involvement patterns in NSCLC. Integrating molecular data could further improve predictive accuracy.
Future Directions Future work should pursue prospective multicenter validation, develop real-time decision support integration into pre-surgical planning workflows, explore EBUS-guided staging as the reference standard (which is less invasive than mediastinoscopy), and test whether the model generalizes to specific patient subgroups such as post-neoadjuvant restaging.