What is occult pleural dissemination? In non-small cell lung cancer (NSCLC), pleural dissemination (PD) means cancer has spread to the lining surrounding the lungs. When this spread is invisible on standard CT scans before surgery, it is called occult PD - a hidden threat that surgeons only discover once they open the chest.
Why it matters so much. NSCLC with pleural spread is classified as stage M1a, the most advanced disease category, with a median survival of only 4 to 11.5 months. Radical surgery is generally not beneficial for these patients, yet between 0.9% and 6.2% of patients who appear to be good surgical candidates are found to have occult PD only when the chest is already open - an outcome known as an "open-close" surgery.
Why imaging alone falls short. Radiologists struggle to detect small pleural metastases on CT, and cannot reliably distinguish them from benign conditions like reactive pleurisy or intrapulmonary lymph nodes. PET scanning also has limitations, with some benign nodules showing high activity and some malignant ones appearing cold.
The goal of this study. Researchers from three major Chinese medical centers set out to build and compare AI-based predictive models using CT imaging that could flag high-risk patients before surgery, potentially sparing them a futile and harmful operation.
A multicenter retrospective cohort. The study enrolled 326 NSCLC patients from three high-volume Chinese hospitals spanning 2016 to 2023. Of these, 163 had surgically confirmed occult PD and 163 were matched controls with pathology-confirmed pleural invasion but no PD - a deliberate design choice to train the model on the subtle differences between local invasion and true dissemination.
How the cohorts were divided. Two hospitals contributed data for training (n = 216) and internal testing (n = 54), while the third hospital provided an entirely separate external test cohort (n = 56). This external validation is essential for checking whether a model can generalize beyond the data it learned from.
Strict inclusion and exclusion rules. Patients with pleural effusion visible before surgery were excluded, because effusion can itself be a sign of PD and would distort both the clinical question and the imaging features. Only patients whose PD was truly occult - meaning preoperative imaging did not raise suspicion - were included in the PD-positive group.
The gold standard for diagnosis. All PD and pleural invasion diagnoses were confirmed by surgical specimens analyzed with intraoperative frozen sections and definitive histopathology, including elastic staining - the most rigorous possible standard.
Standardized CT acquisition and preprocessing. All patients received chest CT scans within one month before surgery using standardized scanning protocols. Images were resampled to a uniform 1x1x1 mm voxel size and pixel intensities were normalized using z-score methods to ensure consistency across different scanner models and centers.
Manual tumor segmentation. A thoracic radiologist with 10 years of experience manually outlined the tumor region on the largest cross-sectional CT slice of each primary tumor using 3D Slicer software. Only the primary lung tumor was included - pleural surfaces and separate pleural lesions were deliberately excluded from the region of interest to avoid data leakage.
Extracting 1,427 radiomic features. Using RIAS software following international standardization guidelines, the researchers extracted 1,427 quantitative features from each tumor region. These included first-order statistics (basic intensity measurements), shape-based features (describing tumor geometry), and texture features (capturing patterns invisible to the human eye).
Rigorous feature selection. Features were filtered for reproducibility using intraclass correlation coefficients (ICC), reduced for redundancy using Spearman correlation analysis, and then trimmed to the most informative 12 features using LASSO regression with 10-fold cross-validation. This sequential filtering prevents overfitting and ensures only robust signals drive model predictions.
A broad machine learning comparison. Ten different ML algorithms were trained on the radiomics features: logistic regression, support vector machine, gradient boosting machine (GBM), neural network, random forest, XGBoost, k-nearest neighbor, AdaBoost, LightGBM, and CatBoost. Comparing this many algorithms ensures the best possible model is selected rather than defaulting to a single approach.
Deep learning with pre-trained networks. Eight convolutional and transformer-based neural networks were tested: AlexNet, ResNet, DenseNet121, GoogleNet, SqueezeNet, VGG, SwinTransformer, and VisionTransformer. All were pre-trained on the large ImageNet dataset, then fine-tuned for the binary task of predicting PD. CT patches were resized to 224x224 pixels for network input.
Data augmentation to handle limited data. Because the dataset was relatively small for deep learning, the training images were artificially expanded using horizontal flipping, vertical flipping, and rotation to help the networks generalize. Early stopping prevented the models from overfitting by halting training once validation performance plateaued.
Two fusion models combining both approaches. A prefusion model combined radiomics and deep learning features before classification, while a postfusion model combined the output probabilities of the best-performing ML model (GBM) and the best deep learning model (DenseNet121) through a stacking ensemble strategy. A neural network classifier then made the final prediction from these combined signals.
GBM led the traditional ML models. Among all 10 radiomics-based machine learning models, the gradient boosting machine (GBM) achieved the highest area under the curve (AUC) of 0.821 in the external test cohort - meaning it correctly distinguished PD-positive from PD-negative patients about 82% of the time based solely on CT texture and shape features.
DenseNet121 topped the deep learning models. Among all 8 deep learning architectures, DenseNet121 achieved the highest AUC of 0.764 in the external test cohort. While slightly lower than GBM, DL models learn directly from image pixels without manual feature engineering, which represents a fundamentally different and complementary source of information.
The postfusion model outperformed all others. By combining the probability outputs from GBM and DenseNet121, the postfusion model achieved AUC scores of 0.828 to 0.978 across the three cohorts - superior to both the prefusion model (0.817 to 0.877) and any single-modality model. Its most striking advantage was sensitivity, reaching 82.1% to 97.2% across cohorts.
Sensitivity is the most critical metric here. In the context of occult PD, failing to detect a true case (a false negative) means a patient undergoes futile open surgery. High sensitivity - correctly catching as many true PD cases as possible - is therefore more important than specificity. The postfusion model's sensitivity of up to 97.2% in the training cohort and 82.1% in the external test cohort is clinically meaningful.
SHAP analysis reveals the top feature. SHapley Additive exPlanations (SHAP) analysis was used to interpret what the GBM model was actually measuring. The most predictive radiomics feature was the square-root transformed first-order standard deviation - a measure of pixel intensity variability within the tumor. Lower values of this feature were associated with higher predicted probability of PD.
Radiomics and deep learning capture different information. A correlation heatmap between the 12 radiomics features and the 115 compressed DL features showed low cross-modality correlations, confirming that the two types of features are genuinely complementary. This independence is exactly why fusing them improves performance - they see different aspects of the same tumor.
Deep learning features were compressed for fusion. DenseNet121 produced 1,024-dimensional feature vectors from its final convolutional layer. These were compressed to 115 features using principal component analysis (PCA), then further refined to 7 key DL features by LASSO regression before being combined with the 12 radiomics features to train the prefusion model.
Decision curve analysis confirmed clinical utility. Beyond AUC numbers, decision curve analysis (DCA) assessed whether using the model would actually benefit patients in practice. When the model was calibrated to a realistic occult PD prevalence of 5%, both the radiomics and postfusion models provided greater net benefit than treating all patients or treating none within the clinically relevant probability threshold range.
CT has well-known blind spots for pleural disease. For PD without pleural effusion, imaging signs include pleural nodules and uneven thickening, but small nodules near the diaphragm are easily obscured by respiratory motion. A prior review of CT scans in patients with known occult PD found that nearly half showed no visible signs at all - even when radiologists knew what to look for.
Benign conditions mimic malignant ones. A study of 164 adenocarcinoma patients without PD found that 25% had nonmalignant pleural or fissural nodules, some of which even showed positive PET uptake. Simultaneously, in actual PD patients with nodules, only 25% showed positive PET uptake. This overlap severely limits the utility of both CT and PET for the subset of occult cases.
Prior models had important limitations. An earlier clinical model using risk factors like age, CEA levels, and pathological findings achieved an AUC of only 0.756 - and required post-surgical pathology data, making it useless for preoperative decision-making. A radiomics model reached an AUC of 0.93 in a single center, but with only 64 PD-positive cases and no external validation.
The case for fusion over single methods. The Delong test showed the postfusion model was slightly but not significantly better than single-modality models (p values greater than 0.05). The authors note this likely reflects the limited deep learning training set - larger multicenter datasets would be expected to widen the performance gap between fusion and single approaches.
A risk stratification tool before the operating room. The intended clinical use is to compute a PD risk score for every NSCLC patient who appears to have pleural invasion on imaging but is being considered for curative surgery. A low score supports proceeding with surgery. A high score triggers a more cautious approach.
What happens with a high-risk prediction. Patients with a high postfusion model score would be candidates for diagnostic thoracoscopy to confirm PD before any major resection. If PD is confirmed, treatment shifts toward systemic therapy - targeted agents or immunotherapy - rather than surgery, sparing the patient a futile and potentially harmful thoracotomy.
The model must be recalibrated for real-world use. The training dataset used a 1:1 ratio of PD-positive to PD-negative cases to address class imbalance, but in the real surgical population occult PD affects only 0.9% to 6.2% of patients. Probability outputs therefore need to be recalibrated to reflect this low baseline prevalence before deployment, converting raw model scores into meaningful patient-specific risk estimates.
Broader implications for surgical oncology. Beyond lung cancer, the fusion modeling approach demonstrated here - combining radiomics-based ML and deep learning - has been shown to improve prediction of occult lymph node metastasis in other cancers. This multicenter study adds to growing evidence that AI models trained on CT radiomics can augment radiologist judgment for staging decisions that determine surgical eligibility.
Retrospective design and selection bias. The study's retrospective nature means only patients who had preoperative CT scans and met strict inclusion criteria were analyzed. This may not fully represent the range of patients encountered in routine clinical practice, and the findings need prospective validation.
Predominantly Chinese patient population. All three participating hospitals are high-volume Chinese centers. Whether the model generalizes to populations with different genetic backgrounds, smoking patterns, or adenocarcinoma subtypes is unknown. International multicenter collaboration is identified as the key next step.
Potential overfitting despite precautions. The model showed higher accuracy on training data than on test data, a classic sign of overfitting. While data augmentation and cross-validation were used to minimize this, the limited training sample and high feature dimensionality remain constraints that only larger datasets can fully address.
Imaging modality expansion. The study used plain CT scans only. Incorporating contrast-enhanced CT or PET/CT for equivocal cases could further improve specificity by adding metabolic and perfusion information, although at greater cost and radiation dose. Future models may also benefit from incorporating clinical variables alongside imaging features.