Development and validation of an interpretable medical deep foundation model for predicting occult lymph node metastasis in early-stage small cell lung cancer: a multicenter retrospective diagnostic study Authors: Xu Jiang, Qin Peng, Jing Zhang, Alex Zwanenburg, Xin Peng, Chao Luo, Jialiang Ren, Lin Yang, Yajia Gu, Steffen Löck, Kunfeng Liu, Meng Li, Jianwei Wang Published: 2026 ABSTRACT AbstractBackgroundOccult lymph node metastasis (OLM) is common in patients with clinical stage T1–2N0M0 small cell lung cancer (SCLC) and is strongly associated with inferior survival. Accurate preoperative identification of OLM remains challenging using conventional imaging, yet is critical for nodal staging, surgical planning, and individualized treatment decision-making. This study aimed to develop and externally validate an interpretable deep foundation model-based approach using preoperative contrast-enhanced computed tomography (CECT) to predict OLM in early-stage SCLC.MethodsIn this multicenter retrospective diagnostic study, 416 surgically treated patients with postoperative pathologically confirmed SCLC and preoperative cT1–2N0M0 status were included. Patients from the largest center were randomly divided into training (n=201) and internal validation (n=86) cohorts, while patients from four additional centers constituted an external test cohort (n=129). Deep imaging features were extracted from preoperative CECT using a cancer-specific pretrained foundation model, and predictive models were constructed using machine-learning classifiers. Model performance was compared with handcrafted radiomics, conventional convolutional neural network (CNN)-based models, and a clinical model. Discrimination, calibration, and clinical utility were assessed using the area under the receiver operating characteristic curve (AUC), calibration analysis, and decision curve analysis.ResultsThe overall incidence of OLM was 33.9% (141/416). Among the evaluated imaging approaches, the foundation model-derived DFM50N achieved an external test AUC of 0.787 [95% confidence interval (CI): 0.703–0.871]. In unadjusted pairwise DeLong comparisons, its AUC was higher than those of RadN and the clinical model (P=0.02 and P=0.02, respectively), whereas no significant difference was observed compared with the combined model (AUC, 0.790; 95% CI: 0.705–0.874; P=0.90). The DFM50N model demonstrated relatively preserved external calibration and provided greater net benefit than strategies of additional staging for all or no patients across a clinically relevant range of threshold probabilities. Model interpretability analysis using SHapley Additive exPlanations (SHAP) revealed similar feature-ranking patterns associated with OLM risk across cohorts.ConclusionsAn interpretable deep foundation model based on preoperative CECT achieved favorable and stable external performance for the noninvasive prediction of OLM in early-stage SCLC. The model showed favorable discrimination compared with handcrafted radiomics, conventional CNN-based approaches, and clinical assessment alone, and its integration with clinical predictors enabled individualized risk estimation. This approach may facilitate the identification of patients requiring more comprehensive mediastinal evaluation and support more informed multidisciplinary surgical planning and lymph node management.Keywords: Small cell lung cancer (SCLC), occult lymph node metastasis (OLM), foundation model, radiomics, contrast-enhanced computed tomography (CECT) Highlight box Key findings • In this multicenter retrospective diagnostic study of 416 patients with clinical T1–2N0M0 small cell lung cancer (SCLC), occult lymph node metastasis (OLM) was observed in 33.9% of patients. • An interpretable deep foundation model based on preoperative contrast-enhanced computed tomography, particularly the DFM50N model, achieved favorable external test performance and higher discrimination than the primary handcrafted radiomics comparator and clinical model. • The DFM50N model showed favorable calibration, clinical net benefit, and similar SHapley Additive exPlanations (SHAP)-based feature-ranking patterns across cohorts. What is known and what is new? • OLM is common in early-stage SCLC and is associated with inferior survival, but conventional imaging remains insensitive for detecting microscopic nodal disease before surgery. Previous radiomics and deep learning approaches have shown promise, but their generalizability may be limited by small datasets, handcrafted feature dependence, and inter-institutional imaging heterogeneity. • This study introduces and externally validates a cancer-specific foundation model–based CT biomarker for individualized preoperative OLM risk estimation in early-stage SCLC. What is the implication, and what should change now? • The proposed model may support preoperative nodal risk stratification, identify patients who may benefit from more comprehensive mediastinal evaluation, and facilitate individualized multidisciplinary surgical planning. • Prospective and international validation is warranted before routine clinical implementation. Introduction Small cell lung cancer (SCLC) is an aggressive neuroendocrine malignancy characterized by rapid progression, early dissemination, and a poor prognosis (1). In patients with early-stage disease who are clinically staged as T1–2N0M0, surgical resection followed by adjuvant therapy may offer curative potential (2,3). Nonetheless, a substantial proportion of these patients harbors occult lymph node metastasis (OLM) (4,5). OLM refers to postoperative pathologically confirmed lymph node metastasis in patients who had been clinically staged as N0 before surgery based on preoperative clinical and imaging assessment, primarily contrast-enhanced computed tomography (CECT). The presence of OLM in SCLC is strongly associated with reduced overall survival, and its detection may influence postoperative staging and adjuvant treatment decisions (3,6,7). Because OLM can markedly alter staging, affect perioperative management, and influence long-term outcomes, accurate preoperative identification remains a critical unmet clinical need in early-stage SCLC. Current preoperative modalities for detecting lymph node metastasis include computed tomography (CT), typically performed as CECT, and positron emission tomography/CT (PET/CT). CECT relies primarily on size-based criteria (e.g., short-axis diameter >10 mm) and may fail to identify micrometastases or morphologically normal nodes with microscopic involvement, leading to false-negative findings (8,9). PET/CT provides important whole-body staging information but may still yield false-negative findings in small-volume or microscopic nodal disease; its accessibility may also be affected by cost and infrastructure requirements (10-12). In addition, invasive procedures such as endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA) or mediastinoscopy provide pathological confirmation but are generally performed selectively according to nodal accessibility, clinical suspicion, and institutional practice (13,14). Radiomics has recently emerged as a promising approach to address this gap by extracting high-dimensional features from medical images to model tumor phenotypes (15-17). However, traditional handcrafted radiomics faces inherent challenges, including reliance on handcrafted features, vulnerability to batch effects caused by heterogeneous imaging protocols, and restricted generalizability across institutions (18,19). Deep learning (DL) models, which automate feature extraction via convolutional neural networks (CNNs), have also shown potential in medical imaging but are constrained by the scarcity of labeled SCLC datasets for supervised learning. Small cohort sizes increase the risk of overfitting, and models trained on limited samples often fail to generalize across centers (20,21). Foundation models, large-scale DL models pretrained on large and diverse datasets, often through self-supervised learning, may help address these limitations by providing transferable representations for downstream tasks (22,23). A cancer imaging foundation model pretrained on 11,467 diverse cancer lesions demonstrated favorable performance in data-limited downstream applications and generated imaging representations associated with tumor microenvironmental and genomic characteristics (24). Such representations may also support interpretable imaging biomarker discovery when combined with appropriate model-explanation methods (25). Foundation model-derived features may therefore be particularly attractive for SCLC, in which limited sample sizes and interinstitutional imaging heterogeneity create substantial challenges for conventional supervised learning. Thus, the purpose of this study was to develop and externally validate an interpretable foundation model-based approach using preoperative CECT to predict OLM in patients with clinical stage T1–2N0M0 (cT1–2N0M0) SCLC. We compared its performance with that of handcrafted radiomics, conventional CNN-derived models, and a clinical model, aiming to provide a noninvasive tool that enhances preoperative nodal risk stratification. Such a model also holds promise for integration into multidisciplinary decision-making, including selection for additional mediastinal evaluation and individualized surgical planning, thereby supporting personalized management of early-stage SCLC. We present this article in accordance with the STARD and TRIPOD reporting checklists (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0714/rc). Methods Study workflow and patient recruitment This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committee of the Cancer Hospital, Chinese Academy of Medical Sciences (No. NCC2021C-213). All participating hospitals were informed of and agreed to the study. The requirement for informed consent was waived by the ethics committee due to the retrospective nature of this study and complete anonymization of patient data. A total of 416 patients with surgically resected, pathologically confirmed SCLC and preoperative cT1–2N0M0 status were included from five Chinese centers: Center 1, the National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College; Center 2, the First Hospital of China Medical University; Center 3, Shanxi Cancer Hospital, Shanxi Medical University; Center 4, Sun Yat-sen University Cancer Center; and Center 5, Fudan University Shanghai Cancer Center. Clinical N0 (cN0) was defined primarily on CECT as the absence of intrathoracic lymph nodes exceeding 10 mm in short-axis diameter or other unequivocal evidence of nodal metastasis; PET/CT and invasive staging were considered when available. OLM was defined as postoperative pathological nodal involvement in a clinically N0 patient. Pathological reports were independently reviewed by a senior pathologist blinded to imaging and clinical data. All patients were retrospectively staged according to the eighth edition of the AJCC tumor-node-metastasis (TNM) classification (26). Inclusion and exclusion criteria are detailed in Appendix 1, and the enrollment process is illustrated in Figure 1. Center 1 contributed the largest eligible cohort (n=287; January 2005–December 2024) and was therefore used for model development; patients were randomly assigned to the training (n=201) and internal validation (n=86) cohorts in a 7:3 ratio. Patients from Centers 2–5 (January 2010–December 2024) were held out and combined to form an institutionally independent external test cohort (n=129), thereby preventing information leakage and enabling a rigorous assessment of model transportability across institutions, imaging protocols, and clinical practices. Flowchart of the study enrollment process. Center 1: the National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College; Center 2: the First Hospital of China Medical University; Center 3: Shanxi Cancer Hospital, Shanxi Medical University; Center 4: Sun Yat-sen University Cancer Center; Center 5: Fudan University Shanghai Cancer Center. AJCC, American Joint Committee on Cancer; CECT, contrast-enhanced computed tomography; CT, computed tomography; NAT, neoadjuvant therapy; SCLC, small cell lung cancer; TNM, tumor-node-metastasis. Before model evaluation, the minimum sample size required for external testing was estimated using ROC-curve power analysis (pROC, R), assuming an OLM prevalence of approximately 33.9% (case-control ratio, 1.95) (27), an alternative area under the receiver operating characteristic curve (AUC) of 0.75 versus a null AUC of 0.50, a two-sided α of 0.05, and 90% power. This yielded a minimum sample size of 80 patients, including approximately 27 OLM-positive and 53 OLM-negative patients. Further details are provided in Appendix 2. The overall workflow is shown in Figure 2. The overall flowchart of this study. 3D ROI, 3D region of interest; CNN, convolutional neural network; CT, computed tomography; DCA, decision curve analysis; ICC, intraclass correlation coefficient; LR, logistic regression; mRMR, minimum redundancy maximum relevance; OLM, occult lymph node metastasis; RF, random forest; ROC, receiver operating characteristic; SA, simulated annealing; SVM, support vector machine. Clinical data, semantic CT features, and nodal assessment Baseline demographic and clinical characteristics, including age, gender, smoking history, and clinical tumor (T) stage, were collected from electronic medical records. The definitions of all semantic CT features followed standardized thoracic imaging criteria and are provided in Appendix 3. Two radiologists with 4 and 5 years of experience in chest imaging independently evaluated the prespecified semantic CT features while blinded to pathological nodal status. Disagreements were resolved by consensus with a senior thoracic radiologist with 25 years of experience. Preoperative diagnostic and nodal staging data, including pathological confirmation of SCLC and use of PET/CT, EBUS-TBNA, or mediastinoscopy, were retrospectively collected. Indeterminate or suspicious PET/CT nodal uptake without definitive evidence of metastasis was not considered positive. Surgical and pathology records were reviewed for nodal assessment strategy, numbers of examined lymph nodes and stations, evaluation of N1, N2, and station 7, and the distribution of metastatic nodes. In the training cohort, univariable logistic regression (LR) was first performed for all baseline clinical and semantic CT features. Variables (P<0.05) were then entered into a multivariable LR to identify independent predictors of OLM. CT acquisition, preprocessing and semi-automatic tumor segmentation For all patients, CECT scans were performed using CT scanners from multiple manufacturers (Siemens Healthineers, Forchheim, Germany; GE Healthcare, Waukesha, WI, USA; Toshiba Medical Systems, Otawara, Japan; United Imaging Healthcare, Shanghai, China; and Philips Healthcare, Best, the Netherlands) across the participating centers. Details of the CT scan protocols are depicted in Appendix 4. For handcrafted radiomics extraction, CECT images were resampled to 1×1×1 mm3 isotropic voxels using linear interpolation, whereas segmentation masks were resampled using nearest-neighbor interpolation. Gray levels were subsequently discretized using a fixed bin width of 25 Hounsfield units (HU) for both original and filtered images. To further standardize voxel intensities across scanners, we optionally applied a scanner-agnostic tracheal air–based centering normalization (“N” variants), in which the mean intensity of the segmented tracheal lumen was shifted to approximately –1,000 HU (tracheal air-based centering normalization, Appendix 5). This preprocessing was evaluated alongside unnormalized counterparts to assess whether such centering provided additional benefit beyond standard HU calibration. The region of interest (ROI) was defined as the primary lung tumor on preoperative CECT and semi-automatic segmentation of the three-dimensional (3D) ROI was performed by two radiologists with over 4 years of experience, who were blinded to clinicopathological information, using ITK-SNAP software (version 3.8.0, www.itksnap.org). When available, PET/CT scans were referenced to more accurately delineate tumor boundaries. All segmented ROIs were subsequently reviewed and manually refined by a senior radiologist with over 25 years of experience to ensure accuracy and consistency, with any discrepancies resolved through consensus. To assess feature robustness, a subset of cases underwent repeated segmentation to estimate intra- and interobserver reproducibility. Feature extraction: handcrafted radiomics and DL-based radiomics (foundation model and CNNs) Handcrafted radiomics features were extracted from each ROI with PyRadiomics (v3.1.0) following Image Biomarker Standardization Initiative (IBSI) recommendations (18,28), with fixed parameters to enhance reproducibility, leading to 1,781 features per case. All settings and additional details are given in Appendix 6. Foundation model-derived features were extracted using a cancer-specific encoder pretrained on a large-scale dataset of 11,467 cancer lesions from CT images (24). This medical foundation model employs a 3D ResNet50-based deep convolutional encoder, trained via self-supervised contrastive learning, to capture robust imaging features of lesions and their microenvironments. The application process in the present study involved several steps: (I) preprocessing: CT datasets were resampled to 1×1×1 mm3 isotropic voxels using tricubic spline interpolation. Voxel intensity values were clipped to the range of –1,024 to 2,048 HU to match the foundation model’s original training specifications. (II) ROI input: three ROI selection strategies were designed based on the tumor centroid: (i) DFM35N, a fixed 35×35×35 voxel cube centered on the tumor centroid; (ii) DFM50N, a fixed 50×50×50 voxel cube centered on the tumor centroid; and (iii) DFMdynN, a dynamically sized ROI defined as the tumor bounding box with a margin, subsequently resampled and zero-padded to match the fixed input size of the ResNet50 encoder. The 50×50×50 voxel cube was specifically evaluated because it matched the standard input window of the pretrained foundation model and, after 1-mm isotropic resampling, corresponded to a 5×5×5 cm3 volume. This size is clinically reasonable for cT1–2N0M0 SCLC, in which the maximal tumor diameter is up to approximately 5 cm, and can therefore encompass the primary tumor while retaining immediate peritumoral context potentially related to occult nodal spread. (III) Feature extraction: for each patient, deep imaging features were extracted from the final average pooling layer of the pretrained foundation model, yielding a 4,096-dimensional feature vector for each ROI configuration. For comparison, DL-based radiomics (CNNs) were derived by training standard CNN backbones, including two-dimensional (2D) ResNet-50, 2D DenseNet-121, 2D VGG-16, and 3D ResNet variants, initialized with ImageNet or self-supervised weights, using ROI crops normalized and resized to a uniform input. Training used Adam (initial learning rate 1×10−4), batch size 32, and standard augmentations (random flips, rotations, scaling). For a fair comparison, we extracted features from each network’s final layer (e.g., 2,048-dimensional for ResNet-50, 1,024 for DenseNet-121, 4,096 for VGG-16). Feature selection Handcrafted radiomic and DL-based (foundation/CNN) radiomic features were screened exclusively within the training cohort following a multi-step process: (I) reproducibility filtering—retained only features with high inter- and intra-observer reproducibility (ICC >0.80). Z-score normalization—all retained features were standardized to zero mean and unit variance within the training cohort to eliminate scale differences and improve the stability of subsequent feature selection and modeling. (II) Univariate filtering—applied the Mann-Whitney U test to identify features significantly associated with OLM status (P<0.05). (III) Redundancy reduction—employed the minimum redundancy maximum relevance (mRMR) algorithm to retain the top 20 features with the highest relevance to OLM while minimizing redundancy. (IV) Optimal subset selection—finally, simulated annealing (SA) (29), a global stochastic optimization algorithm, was employed to determine the optimal feature subset. In this step, the feature set itself was perturbed by randomly adding, removing, or replacing individual features, and each candidate subset was re-evaluated using a random forest (RF) classifier with 10-fold cross-validation. Improvements were explicitly defined as a higher mean cross-validated AUC of the RF classifier in the training cohort. Subsets yielding such improvements were always accepted, while inferior subsets were accepted with a probability determined by the performance gap and the current iteration, allowing occasional escapes from local minima. The final subset obtained through this process constituted the feature signature for model construction. Model construction and performance assessment For each feature set [handcrafted radiomics, DL-based radiomics (foundation model), and DL-based radiomics (CNN)], three machine learning classifiers—LR, support vector machine (SVM), and RF—were independently trained. Model hyperparameters were optimized using five-fold cross-validation within the training cohort: LR: grid search over regularization strengths (C values from 10 −3 to 103) with L2 penalty; SVM: radial basis function kernel, with cost (C) and kernel width (γ) optimized; RF: number of trees (n_estimators =100–1,000) and maximum tree depth (max_depth =2–20) optimized. To address class imbalance between OLM-positive and OLM-negative cases, all classifiers were trained with the parameter class_weight=‘balanced’, which automatically adjusts class weights inversely proportional to class frequencies in the training data (30). To ensure a fair comparison, all handcrafted radiomics, conventional CNN-derived radiomics, and foundation model-derived features were evaluated using the same cohort partitioning, training-only preprocessing and feature selection strategy, classifier candidates, hyperparameter tuning framework, and performance assessment metrics. The external test cohort was not used for model selection or hyperparameter tuning. Because SA optimized the feature subsets using the mean cross-validated AUC of an RF classifier, RF-based signatures were prespecified as the primary models. LR- and SVM-based models using the same selected feature subsets were evaluated as secondary classifier comparisons. Among the RF-based foundation model variants, the model with the most favorable internal validation performance was selected for subsequent analysis and integration with clinical predictors. The external test cohort was reserved exclusively for final evaluation and was not used for model selection. All models without the suffix “N” used the original HU values after clipping, whereas models with the suffix “N” additionally underwent tracheal air–based centering normalization. The selected clinical variables identified in multivariable LR (section “Clinical data, semantic CT features, and nodal assessment”) were combined with the best-performing imaging signature (e.g., the DFM50N score) to build a combined model using LR analysis. Performance metrics were calculated in the training, internal validation, and external test cohorts, including: (I) discrimination: receiver operating characteristic (ROC) curves, AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The Youden-index cutoff was derived in the training cohort and applied unchanged to the internal validation and external test cohorts. Two exploratory operating points were additionally selected in the internal validation cohort: a sensitivity-oriented threshold requiring sensitivity ≥90% while maximizing specificity, and a specificity-oriented threshold requiring specificity ≥90% while maximizing sensitivity. Both thresholds were fixed before application to the external test cohort, and these operating points were considered exploratory rather than validated clinical cutoffs; (II) statistical comparison: AUCs compared using the DeLong test; (III) calibration curves were constructed using 1,000 bootstrap resamples and locally weighted scatterplot smoothing (LOWESS) smoothing. Calibration intercept and calibration slope were estimated using logistic recalibration of the observed outcome on the logit of the predicted probability, following van Calster et al. (31); and (IV) clinical utility: decision curve analysis (DCA) was performed using additional preoperative mediastinal staging as the model-directed clinical action. Model interpretability, nomogram, and research-use calculator To improve transparency and facilitate clinical translation, we applied SHapley Additive exPlanations (SHAP) to interpret the predictions of the best-performing model. SHAP values were computed for each DL feature to quantify its individual contribution to the predicted probability of OLM for every patient. To further contextualize the selected embeddings, their associations with semantic CT and handcrafted radiomic features were evaluated using Pearson correlation analysis. To visualize individualized risk estimates, we constructed a multivariable LR-based nomogram using the best-performing model’s feature signature combined with selected clinical predictors (section “Clinical data, semantic CT features, and nodal assessment”). The combined model was additionally implemented as a research-use web calculator to demonstrate individualized probability estimation. Statistical analysis All statistical analyses were performed using R version 4.3 and Python version 3.8; package details are provided in Appendix 7. Continuous variables are presented as median (interquartile range), and categorical variables as number (percentage). Continuous variables were compared using the Mann-Whitney U test between OLM-positive and OLM-negative patients within each cohort and the Kruskal-Wallis test across the training, internal validation, and external test cohorts. Categorical variables were compared using the Pearson Chi-squared test or Fisher exact test, as appropriate. To quantify cohort imbalance, absolute standardized mean differences (SMDs) were calculated between the training and internal validation cohorts and between the combined development cohort and external test cohort. The development cohort comprised the training and internal validation cohorts. An absolute standardized mean difference (SMD) of ≥0.10 was considered indicative of potential imbalance. All statistical tests were two-sided, with P<0.05 considered statistically significant. Results Patient characteristics and nodal assessment Baseline characteristics are summarized in Table 1, with across-cohort comparisons in Table S1. OLM prevalence was 37.3%, 37.2%, and 26.4% in the training, internal validation, and external test cohorts, respectively (P=0.09). Across the five participating centers, OLM prevalence ranged from 7.7% to 44.8% (Center 1, 37.3%; Center 2, 7.7%; Center 3, 24.2%; Center 4, 22.2%; and Center 5, 44.8%), with an overall prevalence of 33.9% (141/416) (Table S2). Tumor location and short-axis diameter differed significantly across cohorts (P=0.03 and P=0.041), while SMD analyses indicated modest imbalances in selected tumor characteristics. Table 1Clinical and radiological characteristics according to OLM status in the training, internal validation, and external test cohortsCharacteristicTraining cohort (N=201)Internal validation cohort (N=86)External test cohort (N=129)OLM-negative (N=126)OLM-positive (N=75)P valueOLM-negative (N=54)OLM-positive (N=32)P valueOLM-negative (N=95)OLM-positive (N=34)P valueClinical T stage0.100.02*0.28 T167 [53]31 [41]37 [69]14 [44]55 [58]16 [47] T259 [47]44 [59]17 [31]18 [56]40 [42]18 [53]Lobe0.650.810.58 RUL38 [30]16 [21]17 [31]7 [22]24 [25]6 [18] RML7 [5.6]4 [5.3]3 [5.6]3 [9.4]1 [1.1]1 [2.9] RLL25 [20]20 [27]10 [19]5 [16]20 [21]5 [15] LUL31 [25]19 [25]15 [28]10 [31]28 [29]13 [38] LLL25 [20]16 [21]9 [17]7 [22]22 [23]9 [26]Location0.006*0.15<0.001* Central32 [25]33 [44]11 [20]11 [34]10 [11]15 [44] Peripheral94 [75]42 [56]43 [80]21 [66]85 [89]19 [56]Gender0.120.007*0.76 Female22 [17]20 [27]8 [15]13 [41]22 [23]7 [21] Male104 [83]55 [73]46 [85]19 [59]73 [77]27 [79]Age (years)63.5 [54.0,68.0]60.0 [55.0,66.0]0.1163.0 [58.0,69.0]59.0 [55.0,65.5]0.05463.0 [56.0,68.0]62.0 [57.0,71.0]0.74Smoking history0.04*0.002*0.50 No34 [27]31 [41]10 [19]16 [50]25 [26]11 [32] Yes92 [73]44 [59]44 [81]16 [50]70 [74]23 [68]Long-axis diameter (cm)2.9 [1.9,4.0]3.4 [2.5,4.4]0.005*2.3 [1.7,3.2]3.6 [2.5,4.6]<0.001*2.7 [1.9,3.8]3.2 [2.6,4.2]0.02*Short-axis diameter (cm)2.0 [1.4, 2.7]2.2 [1.8, 3.4]0.002*1.6 [1.3,2.2]2.5 [1.8,2.8]<0.001*2.0 [1.5,2.6]2.3 [1.9,2.9]0.03*Shape0.110.260.32 Irregular80 [63]39 [52]40 [74]20 [63]64 [67]26 [76] Round/oval46 [37]36 [48]14 [26]12 [38]31 [33]8 [24]Lobulation0.090.740.13 No8 [6.3]10 [13]7 [13]3 [9.4]14 [15]9 [26] Yes118 [94]65 [87]47 [87]29 [91]81 [85]25 [74]Spiculation0.004*0.670.43 No82 [65]63 [84]40 [74]25 [78]72 [76]28 [82] Yes44 [35]12 [16]14 [26]7 [22]23 [24]6 [18]Calcification0.56>0.990.57 No125 [99]73 [97]53 [98]32 [100]92 [97]34 [100] Yes1 [0.8]2 [2.7]1 [1.9]0 [0]3 [3.2]0 [0]Tumor-lung interface0.580.700.73 Unclear40 [32]21 [28]18 [33]12 [38]25 [26]10 [29] Clear86 [68]54 [72]36 [67]20 [63]70 [74]24 [71]Obstructive change0.01*0.060.29 No101 [80]48 [64]45 [83]21 [66]78 [82]25 [74] Yes25 [20]27 [36]9 [17]11 [34]17 [18]9 [26]Enhancement heterogeneity0.900.350.38 Homogeneous21 [17]13 [17]13 [24]5 [16]11 [12]6 [18] Heterogeneous105 [83]62 [83]41 [76]27 [84]84 [88]28 [82]BVB sign0.760.690.80 No70 [56]40 [53]28 [52]18 [56]61 [64]21 [62] Yes56 [44]35 [47]26 [48]14 [44]34 [36]13 [38]Feeding vessel sign0.370.620.25 No104 [83]58 [77]43 [80]24 [75]72 [76]29 [85] Yes22 [17]17 [23]11 [20]8 [25]23 [24]5 [15]Vascular adjacent sign0.530.04*0.23 No51 [40]27 [36]24 [44]7 [22]36 [38]9 [26] Yes75 [60]48 [64]30 [56]25 [78]59 [62]25 [74]Pleural retraction0.080.220.40 No107 [85]70 [93]48 [89]25 [78]80 [84]31 [91] Yes19 [15]5 [6.7]6 [11]7 [22]15 [16]3 [8.8]Pleural attachment0.100.0510.32 No96 [76]49 [65]44 [81]20 [63]70 [74]22 [65] Yes30 [24]26 [35]10 [19]12 [38]25 [26]12 [35]Peripheral emphysema0.230.120.009* No63 [50]44 [59]26 [48]21 [66]45 [47]25 [74] Yes63 [50]31 [41]28 [52]11 [34]50 [53]9 [26]Interstitial pneumonia>0.990.360.61 No121 [96]73 [97]52 [96]29 [91]92 [97]32 [94] Yes5 [4.0]2 [2.7]2 [3.7]3 [9.4]3 [3.2]2 [5.9]Data are presented as n [%] for categorical variables and median (interquartile range) for continuous variables. Percentages were calculated within each OLM-status group. P values compare OLM-negative and OLM-positive patients within each cohort and were calculated using the Mann-Whitney U test for continuous variables and the Pearson Chi-squared test or Fisher’s exact test for categorical variables, as appropriate. All tests were two-sided. BVB, broncho-vascular bundle thickening; IQR, interquartile range; LLL, left lower lobe; LUL, left upper lobe; OLM, occult lymph node metastasis; RLL, right lower lobe; RML, right middle lobe; RUL, right upper lobe; T, tumor. Clinical and radiological characteristics according to OLM status in the training, internal validation, and external test cohorts Data are presented as n [%] for categorical variables and median (interquartile range) for continuous variables. Percentages were calculated within each OLM-status group. P values compare OLM-negative and OLM-positive patients within each cohort and were calculated using the Mann-Whitney U test for continuous variables and the Pearson Chi-squared test or Fisher’s exact test for categorical variables, as appropriate. All tests were two-sided. BVB, broncho-vascular bundle thickening; IQR, interquartile range; LLL, left lower lobe; LUL, left upper lobe; OLM, occult lymph node metastasis; RLL, right lower lobe; RML, right middle lobe; RUL, right upper lobe; T, tumor. Overall, 392 patients (94.2%) underwent systematic nodal dissection and 24 (5.8%) underwent systematic sampling. The median numbers of examined lymph nodes and nodal stations were 16 [interquartile range (IQR), 11–24] and 6 (IQR, 5–7), respectively. Both N1 and N2 regions were assessed in 94.2% of patients, and station 7 in 90.6%. Corresponding proportions among pathologically node-negative patients were 92.4% and 88.4%, indicating generally comprehensive nodal assessment despite some interinstitutional variation (Table S2). Among 141 patients with OLM, 96 (68.1%) had N1-only involvement, 26 (18.4%) had N2-only involvement, and 19 (13.5%) had both N1 and N2 involvement (Table S3). A definitive pathological diagnosis of SCLC before surgery was available in 87 of 416 patients (20.9%), including 54 of 201 patients (26.9%) in the training cohort, 23 of 86 patients (26.7%) in the internal validation cohort, and 10 of 129 patients (7.8%) in the external test cohort. All preoperative diagnoses were established by bronchoscopic biopsy (Table S4). PET/CT was performed in 58 (13.9%), including 7 with indeterminate or suspicious nodal uptake. EBUS-TBNA and mediastinoscopy were each performed in two patients, with all results negative (Table S5). Feature selection In the training cohort, univariable and multivariable LR analyses identified short-axis diameter and spiculation as significant predictors of OLM. Smoking history showed a nonsignificant association in the multivariable analysis (P=0.11) but was retained in the clinical model because of its established clinical relevance in SCLC (27) and its prespecified inclusion as a candidate clinical predictor (Table S6). The stepwise selection of foundation model features, conventional CNN-derived features, and handcrafted radiomic features is detailed in Appendix 8 and Table S7. Model construction Predictive models were developed using three categories of imaging features: handcrafted radiomics (RadN), foundation model-derived features (DFM35N, DFM50N, and DFMdynN), and features derived from conventional 2D and 3D CNN architectures, including DenseNet, ResNet, and VGG variants. Because feature-subset optimization was based on an RF cross-validation objective, RF-based signatures were treated as the primary models, whereas LR- and SVM-based models were evaluated as secondary classifier comparisons. Among the RF-based foundation model variants, DFM50N achieved the highest internal validation AUC (0.790) and was therefore selected as the primary imaging signature for integration with clinical predictors. The resulting combined model is detailed in Table S8. The selected DFM50N embeddings showed similar correlation patterns without extreme collinearity across cohorts, and their distributions according to OLM status are presented in Figures S1,S2. Model performance assessment Table 2 summarizes the primary performance metrics, and the corresponding ROC curves are shown in Figure 3. DFM50N achieved AUCs of 0.859, 0.790, and 0.787 in the training, internal validation, and external test cohorts, respectively. The corresponding AUCs of the combined model were 0.870, 0.799, and 0.790. Thus, discrimination remained closely similar from internal validation to external testing, with absolute AUC differences of 0.003 for DFM50N and 0.009 for the combined model. In the external test cohort, DFM50N achieved an AUC of 0.787 [95% confidence interval (CI): 0.703–0.871], closely approaching that of the combined model (AUC, 0.790; 95% CI: 0.705–0.874). Table 2Diagnostic performance of the models for predicting OLM in the training, internal validation, and external test cohortsModelAUC (95% CI)Accuracy (95% CI)Sensitivity (95% CI)Specificity (95% CI)PPV (95% CI)NPV (95% CI)Training cohort RadN0.787 (0.722–0.851)0.731 (0.664–0.791)0.680 (0.562–0.783)0.762 (0.678–0.833)0.630 (0.515–0.734)0.800 (0.717–0.867) DFM50N0.859 (0.809–0.909)0.761 (0.696–0.818)0.813 (0.707–0.894)0.730 (0.644–0.805)0.642 (0.537–0.738)0.868 (0.788–0.926) Clinical0.694 (0.618–0.770)0.706 (0.638–0.768)0.507 (0.389–0.624)0.825 (0.748–0.887)0.633 (0.499–0.754)0.738 (0.657–0.808) Combined0.870 (0.822–0.917)0.771 (0.707–0.827)0.893 (0.801–0.953)0.698 (0.610–0.777)0.638 (0.539–0.730)0.917 (0.842–0.963)Internal validation cohort RadN0.684 (0.565–0.803)0.628 (0.517–0.730)0.500 (0.319–0.681)0.704 (0.564–0.820)0.500 (0.319–0.681)0.704 (0.564–0.820) DFM50N0.790 (0.690–0.890)0.756 (0.651–0.842)0.844 (0.672–0.947)0.704 (0.564–0.820)0.628 (0.467–0.770)0.884 (0.749–0.961) Clinical0.724 (0.607–0.841)0.698 (0.589–0.792)0.375 (0.211–0.563)0.889 (0.774–0.958)0.667 (0.410–0.867)0.706 (0.583–0.810) Combined0.799 (0.697–0.901)0.733 (0.626–0.822)0.844 (0.672–0.947)0.667 (0.525–0.789)0.600 (0.443–0.743)0.878 (0.738–0.959)External test cohort RadN0.657 (0.548–0.766)0.636 (0.546–0.719)0.559 (0.379–0.728)0.663 (0.559–0.757)0.373 (0.241–0.519)0.808 (0.703–0.888) DFM50N0.787 (0.703–0.871)0.698 (0.611–0.775)0.735 (0.556–0.871)0.684 (0.581–0.776)0.455 (0.320–0.594)0.878 (0.782–0.943) Clinical0.644 (0.535–0.752)0.674 (0.586–0.754)0.412 (0.246–0.593)0.768 (0.671–0.849)0.389 (0.231–0.565)0.785 (0.688–0.863) Combined0.790 (0.705–0.874)0.690 (0.603–0.768)0.765 (0.588–0.893)0.663 (0.559–0.757)0.448 (0.317–0.585)0.887 (0.790–0.950)Data are presented as point estimates with 95% CIs. *, P<0.05. The Youden-index cutoffs were derived in the training cohort and applied unchanged to the internal validation and external test cohorts: 0.170 for RadN, 0.250 for DFM50N, 0.458 for the clinical model, and 0.254 for the combined model. AUC, area under the receiver operating characteristic curve; CI, confidence interval; NPV, negative predictive value; OLM, occult lymph node metastasis; PPV, positive predictive value. Diagnostic performance of the models for predicting OLM in the training, internal validation, and external test cohorts Data are presented as point estimates with 95% CIs. *, P<0.05. The Youden-index cutoffs were derived in the training cohort and applied unchanged to the internal validation and external test cohorts: 0.170 for RadN, 0.250 for DFM50N, 0.458 for the clinical model, and 0.254 for the combined model. AUC, area under the receiver operating characteristic curve; CI, confidence interval; NPV, negative predictive value; OLM, occult lymph node metastasis; PPV, positive predictive value. Predictive performance of OLM. The ROC curves of the DFM50N, clinical, and combined models in the training cohort (A), internal validation cohort (B), external test cohort (C), respectively. AUC, area under the receiver operating characteristic curve; CI, confidence interval; COMB, combined; OLM, occult lymph node metastasis; ROC, receiver operating characteristic. Within the foundation model family, the tracheal-air-normalized variants generally yielded higher AUCs than their corresponding unnormalized variants, and DFM50N achieved higher discrimination than DFM35N and DFMdynN. Confusion matrices for DFM50N and the combined model are shown in Figures S3,S4, respectively. Foundation model-derived approaches generally achieved higher AUCs than handcrafted radiomics and conventional CNN-derived models, whose external test AUCs were mostly between approximately 0.65 and 0.72. Across classifiers, the RF-based DFM50N model achieved an external test AUC of 0.787, compared with 0.767 for SVM and 0.729 for LR (Table S9). In the external test cohort, DFM50N showed higher AUCs than RadN and the clinical model in unadjusted paired DeLong comparisons (P=0.02 and P=0.02, respectively), whereas no significant difference was observed between DFM50N and the combined model (P=0.90; Table S10). Two exploratory operating points selected in the internal validation cohort were applied unchanged to the external test cohort. At the sensitivity-oriented thresholds, DFM50N (0.200) and the combined model (0.214) achieved sensitivities of 91.2% and 88.2%, with NPVs of 94.2% and 93.1%, respectively. At the specificity-oriented thresholds, DFM50N (0.460) and the combined model (0.526) achieved specificities of 86.3% and 85.3%, with PPVs of 51.9% and 51.7%, respectively. Complete threshold-specific metrics and 95% CIs are provided in Table S11. Calibration curves showed generally reasonable agreement between predicted and observed probabilities for DFM50N across cohorts (Figure S5). In the formal individual-level logistic recalibration analysis, DFM50N yielded intercepts/slopes of 0.854/1.449, 0.715/1.169, and −0.131/0.960 in the training, internal validation, and external test cohorts, respectively. Its external test intercept was close to 0 and its slope was close to 1, indicating relatively preserved external calibration. In contrast, the combined model showed slopes of 0.626 and 0.584 in the internal validation and external test cohorts, respectively, suggesting overly extreme predictions and a potential need for recalibration before clinical implementation (Table S12). Decision-curve analysis was interpreted using additional preoperative mediastinal staging as the model-directed clinical action (Figure S6). In the external test cohort, DFM50N and the combined model provided greater net benefit than strategies of additional staging for all patients or for no patients across approximate threshold-probability ranges of 0.03–0.55 and 0.07–0.49, respectively. The corresponding range for the clinical model was 0.22–0.35. The DFM50N score was significantly higher in OLM-positive than in OLM-negative patients in all three cohorts (all P<0.001; Figure S7). DFM50N interpretability and combined-model clinical interface Figure 4 shows that SHAP analyses of the DFM50N model produced similar feature-ranking patterns across all cohorts. The same embeddings (e.g., FCIB_1105, 2550, 2857, 3248) consistently ranked among the top features across cohorts, with mean absolute SHAP values typically ranging from 0.03 to 0.05. Beeswarm distributions show that higher values of several top embeddings push predictions toward OLM, indicating directional associations between embedding values and model output. The concordant rankings and effect magnitudes across cohorts suggested partially consistent representations, indicating that DFM50N captures lesion/perilesional cues beyond semantic CT descriptors. Furthermore, patient-level waterfall plots of the DFM50N score are shown in Figure S8 and demonstrate separation between OLM-negative and OLM-positive cases across the three cohorts. DFM50N feature importance based on SHAP analysis in the training cohort (A), internal validation cohort (B) and external test cohort (C). SHAP, Shapley Additive Explanations. To further interpret the DFM50N embeddings, we examined their correlations with conventional CT features (Figure 5). Several embeddings were moderately correlated with tumor size and first-order intensity, including FCIB1790 with long diameter (r=0.56), FCIB1790 with volume (r=0.61), FCIB1790 with maximum 3D diameter (r=0.52); FCIB3248 with volume (r=0.56), and FCIB3248 with maximum 3D diameter (r=0.56; all P<0.001). FCIB3248 and FCIB2165 also correlated with mean intensity (r=0.54; r=0.55; all P<0.001). Weaker associations were observed with enhancement heterogeneity and intensity, such as FCIB2550 with entropy/variance (r≈0.20–0.40) and FCIB3792 with mean intensity (r=0.27). In contrast, correlations with morphological descriptors such as lobulation and enhancement heterogeneity were negligible (all |r|<0.2). These findings suggest that DFM50N embeddings partly reflect established size- and texture-related properties while also capturing latent, nonredundant patterns beyond handcrafted features. Correlations between DFM50N embeddings and conventional semantic CT and radiomic features. (A) The heatmap illustrates the Pearson correlation coefficients (r) between the nine most influential DFM50N embeddings (FCIB_F3248, F93, F3580, F1105, F3792, F1790, F2165, F2550, F2857) and conventional semantic CT features, including long diameter, shape, lobulation, spiculation, enhancement heterogeneity, feeding vessel sign, and vascular adjacent sign. (B) The heatmap depicts the correlations between the nine most influential DFM50N embeddings and selected interpretable radiomic features, including tumor volume, maximum 3D diameter, sphericity, mean intensity, maximum intensity, skewness, variance, entropy, GLCM contrast, and uniformity. Correlation coefficients are marked with *, P<0.05; **, P<0.01; ***, P<0.001. 3D, three-dimensional; CT, computed tomography; GLCM, gray-level co-occurrence matrix. For practical demonstration, the combined model was implemented as a research-use web calculator, allowing users to obtain individualized OLM risk estimates. The DFM50N score is generated from preoperative CECT through a reproducible pipeline comprising 1-mm isotropic resampling, a 50×50×50-voxel lesion-centered crop, FCIB version 1.0.0 embedding extraction, and an RF model based on nine selected foundation model-derived features. The resulting score is combined with the selected clinical predictors to estimate individualized OLM probability—detailed in our public repository (https://xujiang-ai.github.io/SCLC_OLM/). The combined-model nomogram is provided in Figure S9. Discussion In this multicenter retrospective diagnostic study, we developed and externally validated an interpretable DL model based on a cancer-specific medical foundation model (DFM50N) for the preoperative prediction of OLM in patients with cT1–2N0M0 SCLC. DFM50N demonstrated favorable discrimination compared with handcrafted radiomics, conventional CNN-derived features, and clinical models, with closely comparable performance in the internal validation and external test cohorts. Integration with clinical predictors enabled the construction of a combined model, which was further implemented as a web-based research prototype to facilitate individualized risk estimation, model interpretation, and potential future clinical translation. Consistent with our previous findings (27), OLM was identified in 33.9% of patients with early-stage SCLC, a proportion higher than that generally reported in early-stage NSCLC. Center-specific prevalence ranged from 7.7% to 44.8%, potentially reflecting differences in patient selection, tumor characteristics, surgical and pathological assessment, and the small sample sizes of some external cohorts. These center-level estimates, particularly those derived from smaller cohorts, should therefore be interpreted with caution. Nevertheless, the overall prevalence underscores the difficulty of detecting microscopic nodal disease and the need for improved preoperative risk-stratification tools. Conventional CT assessment, which relies largely on nodal size and morphology, may fail to identify microscopic metastases (32). PET/CT plays an important role in whole-body staging, especially for detecting otherwise occult distant disease, but its sensitivity for small-volume nodal involvement remains imperfect, and its availability may be limited by cost, scanner capacity, and supporting infrastructure, particularly in resource-constrained settings (33). PET/CT and the proposed model address complementary clinical questions: PET/CT provides whole-body staging information and may identify metabolically suspicious nodal or distant disease, whereas the model estimates the probability of occult nodal involvement among patients already classified as cN0 on conventional assessment. Because PET/CT was performed selectively in only 58 of 416 patients, a reliable head-to-head comparison with the proposed model was not possible. Accordingly, the model should be regarded as an adjunct to, rather than a replacement for, established imaging and pathological staging procedures. Foundation models represent an emerging direction in medical artificial intelligence (AI) (34). The cancer imaging foundation model developed by Pai et al. has demonstrated potential as a transferable feature extractor across multiple downstream tasks (24). To our knowledge, this is the first multicenter study to apply a foundation model–based imaging approach to OLM prediction in SCLC. By leveraging pretrained cancer imaging representations, DFM50N reduces dependence on manually designed radiomic features and on task-specific representation learning from a relatively limited rare-cancer cohort. Its favorable performance in the independent external test cohort supports the potential of foundation model-derived features to provide informative and transferable imaging biomarkers for preoperative OLM prediction. Our findings also compare favorably with prior studies of OLM prediction in early-stage SCLC. A previous multicenter model combining intratumoral and peritumoral radiomic features achieved an external validation AUC of 0.772 (27), supporting the feasibility of CT-based OLM risk stratification. In the present study, DFM50N achieved an external test AUC of 0.787, while the combined model achieved an AUC of 0.790. Although numerical comparisons across independently conducted studies should be interpreted cautiously, the present results suggest that pretrained cancer imaging representations may offer incremental information beyond handcrafted radiomics and conventional CNN-derived features. Importantly, the model maintained similar discrimination in the internal validation and external test cohorts, supporting its potential usefulness across participating institutions. Beyond predictive performance, we evaluated model interpretability to improve transparency and facilitate potential clinical translation. SHAP analysis identified several embeddings that repeatedly contributed to model predictions across cohorts. Correlation analyses further showed that some influential embeddings were moderately associated with established CT and radiomic phenotypes, including tumor size-related, first-order intensity, and texture characteristics, whereas others showed only weak associations with conventional morphological descriptors. These findings suggest that DFM50N embeddings may capture both recognizable imaging phenotypes and latent information not fully represented by handcrafted descriptors. The recurrence of several influential embeddings among the top-ranked features across cohorts suggests that the model was not entirely dependent on cohort-specific signals. Nevertheless, SHAP values describe contributions to prediction rather than biological causality, and the observed correlations should therefore be regarded as exploratory. Future studies integrating imaging with pathological, molecular, and spatial microenvironmental data may help clarify the biological basis of these representations. Although the combined model integrating DFM50N with clinical predictors yielded only a marginal improvement in external test AUC compared with DFM50N alone (0.790 vs. 0.787), this does not imply that clinical variables are uninformative. Rather, the foundation model–derived imaging signature appeared to capture a substantial proportion of the discriminative information relevant to OLM. The principal value of the combined model may therefore lie in embedding the AI-derived score within a clinically interpretable framework using routinely available patient-level variables, thereby facilitating individualized risk communication and multidisciplinary discussion. Some predictor directions differed from conventional clinical expectations, which may reflect multivariable interactions between imaging phenotypes and clinical characteristics rather than simple univariable relationships. These associations should be interpreted as predictive rather than causal. The web calculator provides a practical research interface for demonstrating individualized model output, although prospective validation, standardized preprocessing, and probability recalibration where necessary will be required before routine implementation. In contemporary clinical practice, the proposed model would most plausibly be applied to patients with biopsy-confirmed or strongly suspected early-stage SCLC who are being considered for surgery. However, only 87 of 416 patients (20.9%) in the present cohort had a definitive pathological diagnosis of SCLC before resection. This relatively low proportion reflects the retrospective and surgically selected nature of the cohort, in which many patients underwent resection for an apparently localized primary lung malignancy before the histological subtype was definitively established. The cohort therefore represents a historical real-world practice pattern that may differ from contemporary diagnostic pathways, and the model’s applicability specifically to patients with preoperatively confirmed SCLC requires prospective evaluation. Within this clinical context, the model may serve as an adjunctive preoperative triage tool for patients classified as cN0. At sensitivity-oriented thresholds, the external test NPVs were 94.2% for DFM50N and 93.1% for the combined model, potentially helping identify patients at relatively low predicted risk. At specificity-oriented thresholds, the corresponding high-risk subgroups had observed OLM prevalences of 51.9% and 51.7%, respectively, which may support more comprehensive mediastinal evaluation. Decision-curve analysis further demonstrated greater net benefit than strategies of performing additional mediastinal staging in all or no patients across threshold-probability ranges of 0.03–0.55 for DFM50N and 0.07–0.49 for the combined model. A high model-predicted risk should prompt multidisciplinary review and additional staging rather than cancellation of surgery on the basis of the model output alone. Further evaluation may include PET/CT, repeat review of CECT, and tissue confirmation using EBUS-TBNA, with complementary EUS- or EUS-B-guided sampling according to nodal location and local expertise. If endosonographic sampling is negative or nondiagnostic but clinical suspicion remains high, surgical mediastinal staging should be considered. For suspicious stations 5 or 6, which are not routinely accessible by conventional EBUS-TBNA or standard cervical mediastinoscopy, tissue confirmation may require left video-assisted thoracoscopic surgery, anterior mediastinotomy (the Chamberlain procedure), or extended cervical mediastinoscopy (35). The choice of procedure should be individualized according to nodal location, procedural risk, institutional expertise, and multidisciplinary assessment. Conversely, a low predicted risk should not be used to omit adequate intraoperative and pathological nodal assessment. Thus, treatment decisions should be based on comprehensive staging and pathological confirmation whenever feasible, rather than on the model score alone. The model is therefore best positioned as a complementary tool within established staging pathways, supporting individualized preoperative evaluation and multidisciplinary surgical planning. Several limitations should be acknowledged. First, the retrospective design may have introduced selection bias, and the postoperative pathological reference standard restricted the cohort to surgically treated patients, limiting generalizability to non-surgical populations. Second, although nodal assessment was generally comprehensive, interinstitutional variation and relatively limited nodal evaluation in a small proportion of node-negative patients may have resulted in occasional false-negative classification. Third, modest differences in OLM prevalence and several clinical and radiological characteristics were observed across cohorts, and some external centers had relatively small sample sizes, limiting the precision of center-specific estimates. Fourth, only 20.9% of patients had a definitive pathological diagnosis of SCLC before surgery, including only 10 patients in the external test cohort. This precluded a reliable subgroup analysis and limits direct inference regarding model performance specifically in contemporary populations with biopsy-confirmed SCLC. Fifth, PET/CT was performed selectively in only a minority of patients. Although cost, accessibility, historical staging practice, and clinical selection may have contributed to its limited use, the reasons for non-use were not systematically recorded. This precluded a reliable head-to-head comparison between PET/CT and the proposed models. Finally, although external validation across four independent institutions supported the transportability of the proposed model, development was performed using data from a single center, which may have limited the diversity of imaging protocols, clinical practices, and patient characteristics represented during training. Future studies should investigate multicenter model development, including pooled or federated learning strategies, and conduct prospective international validation in more diverse populations and healthcare settings before widespread clinical implementation. Conclusions In conclusion, the proposed DFM50N model demonstrated favorable and stable discrimination for preoperative OLM prediction in cT1–2N0M0 SCLC, including in an independent multicenter external test cohort. The model-derived score may provide complementary information beyond conventional clinical assessment and handcrafted imaging features, while the combined model offers an interpretable framework for individualized risk estimation. These findings support further prospective evaluation of foundation model-based imaging biomarkers as adjunctive tools for preoperative nodal risk stratification and staging decision support. Supplementary The article’s supplementary files as Acknowledgments The authors express their sincere gratitude to all patients who participated in this multi-institutional study. We also acknowledge the contributions of senior colleagues who provided continuous guidance and advice throughout the conduct of this study. Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committee of the Cancer Hospital, Chinese Academy of Medical Sciences (No. NCC2021C-213). All participating hospitals were informed of and agreed to the study. The requirement for informed consent was waived by the ethics committee due to the retrospective nature of this study and complete anonymization of patient data. Footnotes Reporting Checklist: The authors have completed the STARD and TRIPOD reporting checklists. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0714/rc Funding: This work was supported by the Noncommunicable Chronic Diseases-National Science and Technology Major Project (grant Nos. 2024ZD0520000 and 2024ZD0520002), the Fundamental Research Funds for the Central Universities, Peking Union Medical College (grant No. 3332024207), Beijing Natural Science Foundation (grant No. 7222148), and the National Natural Science Foundation of China (grant No. 81971616). Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0714/coif). The authors have no conflicts of interest to declare. Data Sharing Statement Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0714/dss References Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0714/prf Supplementary Materials The article’s supplementary files as Data Availability Statement Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2026-0714/dss