Occult lymph node metastasis is a critical challenge in lung cancer staging. In non-small cell lung cancer (NSCLC), lymph node involvement is a key prognostic factor in the AJCC TNM staging system. Occult lymph node metastasis (OLNM) refers to cancer spread to lymph nodes that appears negative on standard PET/CT imaging. In the SEISMIC study, OLNM was found in 12% of patients who appeared node-negative on PET/CT, with a 37% discrepancy between PET/CT and invasive EBUS staging results.
OLNM leads to worse outcomes after radiotherapy. Patients treated with stereotactic body radiotherapy (SBRT) for early-stage NSCLC do not undergo surgical lymph node dissection, making OLNM a plausible explanation for their higher regional relapse rates. Up to 20% of SBRT patients experience mediastinal relapse, compared to approximately 10% after surgical lobectomy. This difference likely reflects undetected nodal disease at the time of treatment planning.
Current tools have significant limitations. While FDG PET/CT has transformed lung cancer staging and achieves negative predictive values above 90% for mediastinal nodes, it still misses a meaningful fraction of node-positive cases. Invasive sampling via EBUS can diagnose OLNM but is not widely available and is not routinely recommended. CT-based radiomics models have achieved AUC values of 0.84, and PET/CT-based models have reached 0.90, but most lack external validation and are limited to specific histologies like adenocarcinoma.
What this study uniquely contributes. This study developed and externally validated a PET radiomics model for OLNM prediction that spans both surgical and SBRT patient populations. Unlike most prior studies limited to one treatment setting or one institution, this work validated the model across three cohorts from multiple international centers, including an entirely separate SBRT application cohort not used in training.
Three independent cohorts from multiple international centers. The study used three distinct cohorts. Cohort A consisted of 201 surgical patients from the University Hospital of Brest, France (2010 to 2024), used for model training. Cohort B consisted of 112 surgical patients from the publicly available Radiogenomics dataset (Stanford and Veterans Affairs, USA, 1990 to 1996), used for external testing. Cohort C consisted of 488 SBRT patients from the University Hospitals of Brest and Liege (2010 to 2020), used to evaluate prognostic value for regional relapse-free survival.
OLNM defined by pathological confirmation. Occult lymph node metastasis was defined as pathologically proven N+ disease at surgery despite appearing cN0 (node-negative) on PET/CT imaging. The OLNM rates were 13.0% in Cohort A and 10.7% in Cohort B, reflecting rates representative of real-world clinical practice rather than artificially enriched research populations.
PET radiomics extracted following standardized guidelines. Pre-treatment FDG PET/CT scans were retrieved and tumors were semi-automatically segmented using the PET-Edge gradient-based tool. All segmentations were reviewed by two experienced radiation oncologists. Radiomics features were extracted using PyRadiomics v3.0.1 following IBSI (Imaging Biomarker Standardisation Initiative) guidelines, with 1x1x1 mm spatial resampling. CT images were excluded due to heterogeneity in contrast enhancement protocols across cohorts.
Harmonization across scanners and time periods. Given the multicentric design spanning different scanners across three decades, NeuroCombat statistical harmonization was applied to standardize radiomic features across imaging systems. Normalization was performed separately on each cohort before harmonization, and harmonization parameters from the training cohort were transferred to the validation cohorts following established methodology to prevent information leakage.
Feature reduction from 819 to 13 retained features. Starting from 75 base radiomic features and 744 wavelet-filtered features, the researchers applied a two-step dimensionality reduction: first, Mann-Whitney testing retained only features significantly associated with OLNM (p less than 0.05); then, Spearman correlation analysis removed redundant features (those with correlation coefficient of 0.7 or higher with another retained feature). This process ultimately retained 13 radiomics features for the ModelPET.
Ensemble multilayer perceptron approach. The model was built using a decremental multilayer perceptron (MLP) neural network. A 10-fold cross-validation approach used 10 separate models, each trained on the 60% development portion of Cohort A, with the 40% internal validation fold held out. At each iteration, the feature with the lowest importance was removed and models were retrained. An ensemble model averaged the predictions of the 10 sub-models to produce the final OLNM probability score.
Two models compared: radiomics-only and combined. The primary ModelPET used only PET-derived radiomic features. A secondary ModelCombined incorporated both radiomic and clinical features (tumor stage, age, histology, and gender). This comparison was designed to test whether clinical data adds value when combined with radiomics, or whether radiomics alone is more robust for generalization across different patient populations.
Threshold optimization using Youden index. The probability threshold for classifying a patient as high OLNM risk was set at 12.5%, selected to maximize the Youden index (sensitivity plus specificity minus 1) on Cohort A. A separate threshold of 14% was identified for predicting N2 (mediastinal) node involvement specifically. Model performance was evaluated using AUC, balanced accuracy, sensitivity, specificity, NPV, PPV, F1 score, and decision curve analysis.
Strong training performance with meaningful external validation. In the training cohort (Cohort A), ModelPET achieved an AUC of 0.92, balanced accuracy of 80.0%, sensitivity of 88.5%, specificity of 71.4%, and a negative predictive value of 97.7%. The ModelCombined achieved higher training performance (AUC 0.99, balanced accuracy 85.3%) by incorporating clinical variables including patient age.
External validation shows robust transferability for ModelPET. In the external surgical cohort (Cohort B, different country and era), ModelPET maintained meaningful performance with AUC of 0.73, balanced accuracy of 71.2%, and importantly an NPV of 96.7%. This compares favorably to standard PET/CT alone, which has an NPV of only 89.3% in this cohort. The ModelCombined failed on external validation (AUC 0.67, balanced accuracy only 51.7%) likely due to age distribution differences between cohorts.
Radiomics-only outperforms combined model in external validation. The clear divergence between ModelPET and ModelCombined in external validation demonstrates a key lesson: clinical features that appear useful in training can introduce overfitting when those features differ between populations. Age was the dominant feature in ModelCombined (17.5% importance), but patients with OLNM were significantly older in Cohort B than in Cohort A, making age a confounding rather than generalizable predictor.
Decision curve analysis confirms clinical utility. Decision curve analysis showed that ModelPET provided net benefit exceeding both PET-alone and treat-all strategies across threshold probabilities of 5 to 25%. This range encompasses clinically meaningful decision thresholds for deciding whether a cN0 patient requires EBUS staging before treatment, supporting the model's potential utility in guiding invasive staging decisions.
ModelPET predicts regional relapse-free survival. In the 488-patient SBRT cohort (Cohort C), patients classified as high OLNM risk by ModelPET experienced significantly worse regional relapse-free survival. The hazard ratio for regional relapse was 1.60 (95% CI 1.03 to 2.48, p = 0.04), and the mean 5-year RRFS was 3.5 months shorter in the high-risk group. This is a crucial finding because SBRT patients never undergo lymph node dissection, so OLNM status cannot be directly measured.
N2-specific prediction shows stronger survival signal. When the model threshold was refined to 14% to specifically target N2 (mediastinal) nodal involvement, the prognostic association strengthened further. N2-targeted prediction was associated with both median RRFS (HR 1.69, 95% CI 1.09 to 2.63, p = 0.02) and 5-year restricted mean overall survival (p = 0.003) and median OS (HR 1.28, p = 0.03). This confirms the dominant impact of mediastinal node involvement on long-term survival.
ModelCombined shows no prognostic value in SBRT setting. Unlike ModelPET, the ModelCombined was not associated with any survival endpoints in Cohort C (p greater than 0.05 for all). This further supports the superiority of the radiomics-only approach for cross-setting generalization and suggests that clinical variables introduce population-specific biases that reduce transferability.
Results varied within the SBRT cohort by institution. When the two SBRT sub-populations (Liege and Brest) were analyzed separately, ModelPET was significantly associated with RRFS in the Liege population but not in the smaller Brest cohort. The non-significant result in Brest is difficult to interpret due to the small sample size, limited event count, and high censoring rate, and should not be interpreted as evidence of model failure.
A triage tool to guide EBUS staging decisions. The model's high NPV of 96.7% makes it suitable for a clinical triage strategy: patients classified as low OLNM risk by ModelPET (despite cN0 on PET/CT) could proceed directly to surgery or SBRT without invasive EBUS staging. Patients classified as high risk should be offered EBUS for tissue confirmation before treatment planning. This could reduce the number of unnecessary EBUS procedures while ensuring high-risk patients receive appropriate staging.
Potential to guide adaptation of systemic therapy. For surgical patients identified as high OLNM risk, particularly those with predicted N2 disease, the model could inform decisions about neoadjuvant or perioperative immunotherapy. Evidence supports overall survival benefit from perioperative immunotherapy in locally advanced NSCLC. For SBRT patients at high risk, the model could support consideration of concurrent systemic treatments, although current phase III trials (KEYNOTE-867 and PACIFIC-4) have not demonstrated superiority of adding immune checkpoint inhibitors to SBRT.
Better patient stratification for clinical trial design. The lack of effective stratification tools has been a barrier to improving outcomes with SBRT plus immunotherapy combinations. The ModelPET could provide a principled basis for patient selection in future clinical trials testing combined SBRT and systemic therapy, ensuring that patients most likely to harbor occult nodal disease are enrolled in arms that address regional disease control.
Important caveats for clinical use. The model cannot localize which specific lymph node stations contain OLNM, so it cannot guide targeted nodal irradiation or dissection. Its use must be interpreted alongside clinical judgment and multidisciplinary team discussion. The retrospective design and moderate calibration in external cohorts also mean prospective validation is needed before routine implementation.
Performance drop from training to external validation. The drop from AUC 0.92 in training to 0.73 in external validation reflects both the expected generalization gap and specific challenges in this study: the external cohort was from a different era (1990 to 1996 vs. 2010 to 2024) with potentially different lymph node dissection templates and staging guidelines. The moderate calibration in validation cohorts suggests that predicted probabilities should be used cautiously without local recalibration.
CT radiomics deliberately excluded despite known value. Including CT features in addition to PET would likely have improved model performance based on prior literature. However, CT acquisition was not standardized across the three cohorts, particularly regarding contrast enhancement, which is known to substantially influence radiomic features. Including CT would have introduced systematic bias between cohorts that would confound interpretation.
The most important radiomic feature has biological meaning. The first-order mean PET intensity feature was the most important predictor, approximating mean tumor metabolic activity (SUV). This has been repeatedly associated with treatment response in lung cancer and reflects the known relationship between metabolic tumor volume, heterogeneity, and nodal spread. Gray-Level Run Length Matrix features, associated with tumor aggressiveness and heterogeneity, were also important contributors.
Comparison to the largest prior multicenter study. Zhong et al. (2023) published the largest prior multicenter study on OLNM prediction in lung cancer with 1354 patients in the validation cohort, achieving NPV of 96.0%. The current study achieves a comparable NPV of 96.7% in a true external cohort from a separate institution, with an OLNM prevalence closer to real-world rates (10.7% vs. 33% in Zhong et al.), making the current NPV a more clinically meaningful estimate.
The first PET radiomics model validated across both treatment modalities. This study is among the first to develop and validate a PET radiomics model for OLNM prediction in NSCLC that has been tested in both surgical and SBRT patient populations from multiple institutions. The ability to predict regional relapse risk in SBRT patients, who cannot undergo pathological staging, is a particularly valuable clinical contribution.
NPV improvement over PET/CT alone is clinically meaningful. Improving the NPV from 89.3% (PET/CT alone) to 96.7% (ModelPET) in the external validation cohort means fewer high-risk patients are falsely reassured that their lymph nodes are clear. In a disease where undetected nodal spread significantly worsens prognosis and changes treatment strategy, this improvement has direct implications for patient safety and treatment quality.
Radiomics alone outperforms combined clinical-radiomic models for generalization. The consistent finding that ModelPET outperformed ModelCombined in external validation challenges the common assumption that adding clinical variables always improves model performance. When clinical variables differ systematically between training and validation populations, they can actively harm generalizability. This has broader methodological implications for how radiomics models should be developed and tested.
Prospective validation is the necessary next step. While the multicohort retrospective design provides meaningful evidence, prospective studies and multi-center validation using standardized prospective data collection are needed before clinical implementation. A high-priority next step is evaluating whether the ModelPET can guide EBUS decisions in a prospective trial, measuring whether selective EBUS based on ModelPET classification improves staging accuracy while reducing the number of invasive procedures.