The study adhered to the CLEAR checklist for conducting and reporting experimental research, detailed in Supplementary Data Sheet S1 . The flowchart of this study is shown in Figure 1 , including clinical data collection, image acquisition, region of interest (ROI) segmentation, feature extraction and selection, model construction and performance evaluation, as well as correlation analysis.

A retrospective analysis was conducted on patients with lung adenocarcinoma who underwent a pretreatment ¹⁸F-FDG PET/CT examination at our hospital between May 2022 and April 2024. Inclusion criteria: (1) All patients used the identical ¹⁸F-FDG PET/CT equipment under uniform scanning conditions; (2) Lung adenocarcinoma was pathologically diagnosed for the first time; (3) Complete clinical and imaging records. Exclusion criteria: (1) Pure ground-glass nodule (pGGN) without FDG metabolism; (2) Indistinct tumor boundaries on ¹⁸F-FDG PET/CT images hindering sketch completion; (3) Previous tumor history or concurrent malignant neoplasms; (4) Chemotherapy, radiotherapy, targeted therapy, immunotherapy, and other anti-tumor treatments before ¹⁸F-FDG PET/CT examination. All selected patients were randomly divided into training and testing sets at a ratio of 7:3.

Clinical information, comprising gender, age, smoking history, semantic features (tumor location, lobulation sign, spiculation sign, pleural indentation sign, and bronchial cut-off sign), serum carcinoembryonic antigen (CEA) levels, and Ki-67 expression level was collected from medical records. The CEA levels were measured by electrochemiluminescence, with a reference range of 0 to 5 ng/ml. Hematoxylin/eosin (H&E) stained whole slide images (WSI) were acquired from some of these patients for further analysis. Criteria for metastatic lymph nodes: The gold standard for diagnosis is the pathological results. For suspicious lesions for which surgical/puncture pathology results could not be obtained, the final diagnosis relied on multiple examinations and subsequent follow-up images over a period of more than 3 months.

This study received approval from the Medical Ethics Committee of our hospital, thereby exempting patients from informed consent.

All ¹⁸F-FDG PET/CT scans were performed on the same device (GE Discovery MI PET/CT, GE HealthCare, Waukesha, WI). All patients fasted for more than 6 hours before the imaging and had blood glucose levels below 11.1 mmol/L. After receiving an injection of 0.11-0.14 mCi/kg of ¹⁸F-FDG, patients rested quietly for approximately 60 minutes. A breath-holding CT scan was performed from the vertex of the skull to the mid thighs, and used for attenuation correction purposes as well as anatomic location of ¹⁸F-FDG uptake. The CT images were acquired using 64-slice helical CT with the following settings: 120 or 140 kV, automatic tube current technique, reconstructed layer thickness of 1.25 mm, a rotation time of 0.6 s, a pitch of 1.375, a matrix of 512×512, and lung window (window width, 1500 HU; window position, -700 HU). Subsequently, the PET images were acquired with a matrix of 128×128 and a layer thickness of 2.78 mm. Following data acquisition, attenuation correction and reconstruction procedures were conducted to generate PET, CT, and PET/CT fusion images for each three-dimensional scanning level (transverse, coronal, and sagittal planes), as well as the whole-body MIP maps of PET images.

In this study, the lesion’s ROI delineation and feature extraction of PET/CT images were completed on LIFEx 7.4.0. The ROI was delineated layer by layer on the CT and PET images by two senior attending nuclear medicine physicians. The ROIs of PET images were semi-automatically delineated using a threshold of 42% of the SUVmax as the optimization criterion, and then resampled to 1mm×1mm×1mm (x, y, z) to standardize the voxel spacing. The feature extraction parameters were set to the default values. In the CT and PET ROIs, 179 radiomic features were extracted from the original images respectively, including 50 morphological features, 73 first-order statistics features, and 56 second-order feature parameters.

For missing data, the extracted PET and CT radiomic features were imputed with the median value and standardized by Z-score normalization. Subsequently, features exhibiting statistical differences were identified using the t-test or Mann-Whitney U test. Spearman correlation analysis was employed to remove features with high correlation, specifically those with a Spearman’s correlation coefficient greater than 0.9. The Max-Relevance Min-Redundancy (mRMR) method and Gradient Boosting Decision Tree (GBDT) algorithm were utilized to further diminish data dimensionality and isolate the most informative radiomic features.

To ensure an accurate assessment, tumors were re-evaluated by two pathologists with over five years of experience in diagnosing lung cancer. Each tumor was categorized in accordance with the WHO classification system for lung cancer (5th version), and the percentage of each histological component was recorded in 5% increments, determining the presence or absence of micropapillary/solid components in the lesions. Decisions were made through collaborative consultation and discussion in case of disagreement. A representative section of each patient was selected and digitized using Pathology Medical Image Analysis System IBL500 at a magnification of 40x, subsequently exporting the images in.svs format.

A mouse anti-human Ki-67 monoclonal antibody was used for the immunohistochemical detection. Positive and negative controls were set up separately, and cells with brownish-yellow nuclei were considered as positive cells. The number of Ki-67 positive tumor cells was counted under 400x microscope in five fields. The percentage of Ki-67 expression level positive staining of tumor cells in each field = the number of positive tumor cells in each field/total tumor cells in each field × 100%. The Ki-67 indices of five visual fields were calculated and averaged.

Tumor regions within WSIs were manually delineated utilizing Qupath 0.5.1 and the feature extraction were completed on CellProfiler 4.2.7 (29). To reduce the computational time, the delineated WSIs were segmented into patches with a field of view of 1024×1024 pixels. For each patient, 20 patches were randomly selected and were clear and unobstructed. All patches underwent color normalization utilizing the Vahadane method (28) for subsequent processing. An automated image processing workflow was developed utilizing CellProfiler to extract quantitative features based on images, tumor nuclei, and tumor cytoplasm with 225, 279, and 271, respectively. The Image-based features comprehensively evaluated each patch, including overall image quality, intensity, granularity, texture, and correlation between stained images. The Nuclei- and Cytoplasm-based features encompass a variety of characteristics, including the number of measured objects, their location, shape, intensity, granularity, texture, and spatial relationships. The mean value of each feature was calculated and aggregated to the WSI level for further analysis.

Initially, the extracted pathomic features were standardized by Z-score normalization for preprocessing. Spearman correlation analysis was employed to eliminate redundant information; if the correlation coefficient was greater than 0.9, one of them was retained. The final feature selection was performed using the least absolute shrinkage and selection operator (LASSO) algorithm with five-fold cross-validation. Features with non-zero coefficients were retained for correlation analysis.

Clinical risk factors predictive of LNM in lung adenocarcinoma were identified by univariate and multivariate logistic regression analysis. Based on the stacking ensemble learning algorithm, a clinical model, a PET/CT radiomics model, and a combined model were developed. The stacking ensemble learning algorithm employed logistic regression (LR), random forest (RF), and naive Bayes (NB) as the base learners, with logistic regression (LR) serving as the meta-learner. Optimal model parameters were automatically determined using five-fold cross-validation, which shuffles data into 5 subsets, trains on 4, and validates on 1, repeating to minimize bias. Furthermore, three conventional machine learning algorithms—LR, RF, and NB—were utilized to develop individual combined models, which were compared to the Stacking-combined model.

The study was statistically analyzed using SPSS 26.0, R 4.4.1, and Python 3.9.1. Continuous variables were compared using the t-test or the Mann-Whitney U test. Categorical variables were compared using the Chi-square test or Fisher’s exact test.

Radiomic feature selection employed mRMR (scikit-learn 1.0.2, Python) and GBDT-based importance ranking (LightGBM 3.3.2, Python), whereas pathomic feature selection utilized LASSO regression (glmnet 4.1-8, R). Predictive models were constructed using scikit-learn 1.0.2 (Python) and evaluated with pROC 1.18.0 (R), with performance quantified by the area under the receiver operating characteristic curve (AUC). AUC differences were evaluated using the DeLong test, and decision curve analysis (DCA) was employed to evaluate the clinical utility of each model. Visualizations were generated with matplotlib 3.5.1 and seaborn 0.11.2 (Python). Feature importance was interpreted using SHapley Additive exPlanations (SHAP) values (shap 0.41.0, Python) with stability validated by 3-fold cross-validation. Spearman’s correlations between radiomic and pathomic features were adjusted for multiple comparisons using the Benjamini-Hochberg false discovery rate (FDR) procedure, with statistical significance defined as FDR-adjusted P (q-value) < 0.05.

Ninety patients diagnosed with lung adenocarcinoma were enrolled in this study, comprising 52 with LNM and 38 without LNM. The flowchart for screening patients is shown in Supplementary Figure S1 . The cohort included 48 males and 42 females. Participants were randomly assigned to the training (n = 62) and testing (n = 28) sets in a 7:3 ratio. Statistically significant differences in CEA levels were observed between the LNM and non-LNM groups in both sets (training set: P = 0.004, testing set: P = 0.001). No statistically significant differences were noted in gender, age, smoking history, tumor location, lobulation sign, and spiculation sign between the two groups (P > 0.05) ( Table 1 ).

Pathological data were accessible for 25 patients in this study, and Table 2 describes the pathological characteristics of the two groups. Notably, LNM in lung adenocarcinoma was significantly correlated with the presence of micropapillary component (P = 0.046). However, there was no statistical difference between the two groups with respect to the presence of solid component or the presence of micropapillary/solid components (P > 0.05). Seventy-five patients underwent immunohistochemical Ki-67 proliferation index assay, and the difference in Ki-67 expression levels between 42 LNM and 33 Non-LNM patients was statistically significant [37.50 (20.00, 50.00)% vs. 10.00 (10.00, 40.00)%, P = 0.002] ( Figure 2 ).

Univariate and multivariate logistic regression analyses were conducted on the clinical risk factors of patients as detailed in Table 3 . The bronchial cut-off sign (OR = 4.55, 95%CI: 1.67-12.43, P = 0.003) and CEA (OR = 1.02, 95%CI: 1.00-1.04, P = 0.024) emerged as clinical risk factors for predicting LNM in lung adenocarcinoma.

A total of 358 PET/CT radiomic features were extracted. Firstly, 156 features were eliminated by statistical methods. Then, Spearman correlation analysis was used to exclude 162 highly correlated features. Finally, 30 and 8 features were further excluded by using mRMR and GBDT algorithms, respectively. After feature selection, PET_GLRLM_LongRunHighGreyLevelEmphasis and CT_INTENSITY-BASED_IntensityBasedEnergy were retained, demonstrating a significant difference between the LNM and Non-LNM groups (P < 0.05) ( Table 4 ). Consequently, these features were integrated to develop the combined models.

The above results indicated that LNM in lung adenocarcinoma was correlated with the presence of micropapillary component. However, differences in tumor cell morphology observed in histopathological images are not easily detected through manual inspection; instead, they could be distinguished using quantitative image features (30). Firstly, according to Spearman correlation analysis, 574 highly correlated redundant features are excluded. Then, the LASSO algorithm was used for further screening. When Lambda was 0.085 ( Figure 3 ), the eight most valuable pathomic features associated with micropapillary component were retained: Nuclei_Neighbors_AngleBetweenNeighbors_Expanded, Cytoplasm_Location_Center_X, ImageQuality_Correlation_Eosin_20, Nuclei_RadialDistribution_RadialCV_Hematoxylin_3of4, Nuclei_Texture_Variance_Hematoxylin_3_03_256, Nuclei_AreaShape_Zernike_4_2, Cytoplasm_Granularity_9_Eosin, Cytoplasm_Granularity_13_Eosin ( Figure 4 ).

The Receiver Operating Characteristic​ (ROC) curve analysis evaluated the clinical model, PET/CT radiomics model, and combined model utilizing the stacking ensemble learning algorithm for diagnosing LNM in lung adenocarcinoma. In the training set, the AUC values for the Stacking-clinical model, Stacking-PET/CT radiomics model, and Stacking-combined model stood at 0.749 (95% CI: 0.638-0.858), 0.893 (95% CI: 0.808-0.964), and 0.971 (95% CI: 0.917-1.000), respectively. In the testing set, the AUC values for the three models were 0.771 (95% CI: 0.615-0.914), 0.854 (95% CI: 0.719-0.959), and 0.901 (95% CI: 0.770-1.000), respectively. The Stacking-combined model demonstrated superior diagnostic efficiency, with the accuracy, sensitivity, and specificity of 0.968, 0.972, and 0.962 in the training set, and 0.857, 0.875, and 0.833 in the testing set ( Figures 5A, B ). In addition, the DeLong test for the training set indicated that the Stacking-combined model exhibited a significantly higher AUC compared to both the Stacking-PET/CT radiomics model (P = 0.015) and the Stacking-clinical model (P = 0.002). Conversely, in the testing set, the Stacking-combined model did not demonstrate a significant difference when compared to the Stacking-PET/CT radiomics model (P = 0.330) or the Stacking-clinical model (P = 0.140).

The conventional combined models utilizing LR, RF, and NB algorithms were compared to the Stacking-combined model, as depicted in Figures 5C, D . Analysis of the ROC curves indicated that the RF-combined model outperformed the other conventional models, with an AUC value of 0.919 (95% CI: 0.854-0.970) in the training set and an AUC value of 0.846 (95% CI: 0.706-0.971) in the testing set. The diagnostic efficacy of the Stacking-combined model for identifying LNM exceeded that of the conventional models. The DeLong test for the training set showed that the Stacking-combined model was statistically different from the LR-combined model (P = 0.033) and the NB-combined model (P = 0.003), but not significantly different from the RF-combined model (P = 0.210). The DeLong test for the testing set demonstrated that the Stacking-combined model was statistically different from the NB-combined model (P = 0.045), but not significantly different from the LR-combined model (P = 0.260) and the RF-combined model (P = 0.210). The diagnostic performance parameters in each predictive model are presented in Table 5 .

DCA demonstrated that the Stacking-combined model exhibits superior performance in distinguishing LNM when compared to the combined models based on LR, RF, and NB algorithms, as illustrated in Figure 6 .

Two radiomic features maintained identical ranking across all 3-fold cross-validation ( Table 6 ). PET_GLRLM_LongRunHighGreyLevelEmphasis consistently ranked as the most impactful feature, with a mean SHAP value ranging from 0.30 to 0.32 (± 0.07-0.09). The second most important feature, CT_INTENSITY-BASED_IntensityBasedEnergy, showed stable performance across folds, with a mean SHAP value of 0.23 to 0.25 (± 0.05-0.07).

Spearman correlation analysis was utilized to evaluate the potential relationships between radiomic features and pathomic features. According to the Table 7 , radiomic-pathomic correlations did not reach statistical significance after FDR adjustment (q > 0.05). However, nominal associations (uncorrected P < 0.05) were observed for partial feature pairs. PET_GLRLM_LongRunHighGreyLevelEmphasis demonstrated a moderate negative correlation with ImageQuality_Correlation_Eosin_20 (r = -0.422, uncorrected P = 0.035), as well as a moderate positive correlation with Nuclei_Texture_Variance_Hematoxylin_3_03_256 (r = 0.408, uncorrected P = 0.043); CT_INTENSITY-BASED_IntensityBasedEnergy demonstrated a moderate negative correlation with Cytoplasm_Location_Center_X (r = -0.407, uncorrected P = 0.044).

In terms of the association between radiomic features and Ki-67 expression levels, PET_GLRLM_LongRunHighGreyLevelEmphasis showed a significantly stronger positive correlation with Ki-67 expression level (r = 0.610, q < 0.001); CT_INTENSITY-BASED_ IntensityBasedEnergy showed a significant moderate positive correlation with Ki-67 expression level (r = 0.332, q = 0.004) ( Table 8 ).