The occult pleural metastasis problem Pleural metastasis (PM) in lung cancer is classified as stage M1a - advanced, incurable disease with median survival of only 11.5 months and 5-year overall survival of 10%. A subset of patients have 'occult' pleural metastasis (OPM) - pleural tumor deposits found only during thoracoscopy, without obvious pre-operative CT signs. These patients undergo unnecessary surgery.
Why OPM matters clinically Discovering unexpected OPM during planned curative surgery forces surgeons to abandon the resection and shifts the patient to palliative treatment. Identifying OPM before surgery allows appropriate triaging: patients with predicted OPM can undergo diagnostic thoracoscopy first or proceed directly to systemic treatment, avoiding unnecessary thoracotomy.
Study design Researchers at Peking University People's Hospital built a predictive model for OPM using 50 OPM-positive and 50 matched OPM-negative lung cancer patients as the training cohort. CT radiomic features from the tumor and adjacent pleural area were combined with clinical variables (CEA, NLR, cT stage, tumor-pleural relationship) into a nomogram. The model was validated in 545 patients across three medical centers.
Key result The combined radiomic-clinical nomogram achieved AUC 0.890 in the training cohort and 0.855 in the multicenter validation cohort - substantially better than clinical features alone (AUC 0.761). The model included CEA, NLR, and tumor-area radiomic scores as the optimal combination.
Pleural metastasis biology The pleura is a common site of lung cancer spread due to its proximity to the primary tumor. Malignant pleural involvement is staged as M1a in the IASLC TNM 8th edition staging system, upstaging the patient to stage IVA regardless of primary tumor size or lymph node status. This dramatically changes both treatment strategy and prognosis.
The challenge of occult disease While obvious pleural metastasis presents as malignant effusion, multiple pleural nodules, or irregular pleural thickening on CT, occult pleural metastasis lacks these features preoperatively. It is found incidentally during exploratory thoracoscopy at rates of 0.9-5.3% in surgical candidates. Patients with OPM have better prognosis than those with clinically evident PM, but still should not undergo curative resection.
Current diagnostic gaps Preoperative staging CT, PET-CT, and even pleural fluid cytology can miss OPM. There are no established CT imaging criteria to reliably identify OPM before surgery. The tumor-pleural spatial relationship (how close the tumor is to the pleura) provides some information, but individual visual assessment is subjective and lacks standardized quantification.
Radiomic opportunity Radiomics can extract subtle quantitative texture features from both the tumor and adjacent pleural tissue on CT that are invisible to the human eye but correlate with biological characteristics like pleural invasion. Combining radiomic features with clinical biomarkers (CEA, inflammatory ratios) into a nomogram provides a comprehensive preoperative risk assessment.
Training and validation cohorts The training cohort used 50 OPM-positive cases (matched 1:1 by age, sex, BMI with 50 OPM-negative cases). OPM was confirmed by thoracoscopy and histopathology from biopsies of pleural lesions noted intraoperatively. The 545-patient multicenter validation cohort came from three hospitals (January 2023 to June 2024) and had a real-world OPM rate of 4.59% (25 positive).
Dual segmentation approach Two region-of-interest (ROI) types were created for each patient. The Volume of Interest (VOI) encompassed the entire 3D primary lung tumor, semi-automatically delineated using ITK-SNAP. The pleural ROI was manually drawn as a 2cm x 2cm area of the closest pleural surface to the tumor, starting from the tumor edge. Both ROIs were used to extract separate radiomic feature sets.
LASSO feature selection 121 radiomic features were extracted from each ROI using PyRadiomics (first-order, shape, GLCM, GLSZM, GLRLM, NGTDM, GLDM). Univariate analysis, correlation filtering (removing features with pairwise correlation greater than 0.7), and LASSO logistic regression with 10-fold cross-validation reduced features to the most informative predictors. Separate radiomic scores were calculated for the tumor VOI and pleural ROI.
Nomogram construction Multivariate logistic regression identified four significant clinical predictors: NLR, CEA, clinical T stage, and tumor-pleural relationship. Multiple model combinations were tested: clinical model alone, tumor radiomic model, pleural radiomic model, and combined models. The best combined model (CEA + NLR + VOI radiomic score) was built into a visual nomogram for clinical use, with performance assessed by DeLong test, Hosmer-Lemeshow calibration, and decision curve analysis.
Clinical predictors identified Four clinical variables were independently significant in multivariate analysis: NLR (p=0.043), CEA (p=0.010), clinical T stage (p=0.010), and tumor-pleural relationship type (p=0.037). Elevated CEA was present in 54% of OPM-positive vs. 24% of OPM-negative patients. NLR at or above 2 was present in 66% of OPM-positive vs. 42% of OPM-negative patients.
Clinical model performance A nomogram using only the four clinical predictors achieved AUC 0.761 in the training cohort - better than chance but insufficient for reliable clinical decision-making. This establishes the baseline improvement that radiomic features need to surpass.
Combined model outperforms all others The optimal model combining CEA, NLR, and tumor VOI radiomic score achieved AUC 0.890 (training) and 0.855 (multicenter validation) - a statistically significant improvement over the clinical-only model (DeLong p less than 0.05). Adding the pleural ROI radiomic score did not further improve performance beyond the VOI radiomic score alone.
Decision curve analysis Decision curve analysis confirmed that the combined nomogram provided net clinical benefit at threshold probabilities between approximately 10% and 80%, indicating that using the model to guide clinical decisions would result in net benefit over either treating all patients as OPM-positive or treating all as OPM-negative.
Reducing unnecessary surgery By identifying high-risk patients preoperatively, the nomogram allows surgeons to order diagnostic thoracoscopy (with frozen section biopsy) before committing to curative resection. Patients with nomogram-predicted high OPM risk who are confirmed positive at thoracoscopy can be redirected to systemic therapy without the morbidity of a full thoracotomy.
Integration with existing clinical workflow The model uses only variables already collected in routine lung cancer staging: blood CEA level, complete blood count (for NLR calculation), and chest CT scans. No additional imaging, biopsies, or laboratory tests are required. The nomogram can be applied at the time of multidisciplinary tumor board review before surgical planning.
CEA and NLR as accessible biomarkers The identification of CEA and NLR as key predictors has immediate clinical relevance. NLR - the ratio of neutrophils to lymphocytes - reflects systemic inflammatory status and is an emerging pan-cancer prognostic marker. Elevated CEA reflects active tumor biology. Both are cheap, widely available laboratory tests.
Multicenter validation significance The model was validated in 545 patients from three institutions with different CT scanners (Siemens and Philips) and real-world OPM prevalence (4.59%), closely matching published epidemiological rates. This multicenter validation is a critical differentiator from most radiomic studies and substantially increases confidence in the model's generalizability.
Small training cohort with enriched OPM cases The training cohort of 100 patients (50 OPM-positive, 50 OPM-negative) is small, and the 50% OPM prevalence in training is artificially enriched compared to real-world rates of under 5%. While the validation cohort had a more realistic 4.59% OPM rate, the enriched training design can introduce performance optimism in the training AUC.
Retrospective single-institution training The training cohort came from a single center (Peking University People's Hospital) from 2015-2022. Surgical criteria, CT scanning protocols, and OPM detection thoroughness may differ at other institutions. The validation included three centers but was limited to a single geographic region (China).
Segmentation dependency The radiomic model requires semi-automatic tumor VOI segmentation and manual pleural ROI delineation. These steps are time-consuming, operator-dependent, and not yet standardized across institutions. Automated segmentation tools and standardized pleural ROI protocols are needed for scalable clinical implementation.
Future directions Prospective multicenter studies with standardized CT acquisition protocols and pre-specified OPM endpoints would establish the model's clinical validity more rigorously. Development of automated segmentation pipelines and integration into surgical planning software represents the translational pathway. Extension to predict OPM in other cancer histologies (squamous cell carcinoma, SCLC) warrants investigation.