The study was approved by the Ethics Committee of X Hospital (KYLH2022-1). Written informed consent was obtained from all patients before dPET and FDG PET/CT imaging.

A total of 323 patients underwent dPET (chest, 65 min) and static FDG PET/CT (sPET/CT) imaging (whole body, 10–20 min) from May 2021 to December 2024. All patients had lung nodules or masses identified on a chest CT scan, and none of the patients received anti-infective or antitumor therapy prior to undergoing a dPET+sPET/CT scan. Of these patients, 261 had lung cancer confirmed by puncture and/or surgical pathology.

We retrospectively collected 279 FDG-avid LNs from 74 patients with pathologically confirmed lung cancer. The 74 patients were selected from 261 lung cancer patients. On the sPET/CT scan, mediastinal or pulmonary hilar region LNs were considered FDG-avid LNs if their FDG uptake exceeded the mediastinal blood pool. All 279 FDG-avid LNs were confirmed by pathology, and the LNs were included according to their distribution and size on the sPET/CT scan after a one-to-one correspondence with the pathological findings. The locations of the LNs according to the International Association for the Study of Lung Cancer (IASLC) are shown on the LN map (16). The time interval between the dPET+sPET/CT scan and receipt of the pathology results was less than 2 weeks.

We collected the dPET+sPET/CT scans and the clinical features of FDG-avid LNs. The dPET+sPET/CT scan features included the primary focus site, the primary focus SUV_max, the FDG-avid LN zoning, the LN short and long diameters, the LN-SUV_max, and the LN-K _i. The clinical characteristics included gender, age, primary lung cancer pathology, and LN pathology.

Based on the pathological findings, of the 279 FDG-avid LNs, 161 (57.71%) were metastatic and 118 (42.29%) were non-metastatic. The participants were randomly divided into two groups: a training group (n = 196) and a test group (n = 83).

Both the dPET and sPET/CT scans were performed using a Discovery MI PET/CT (GE Healthcare, Milwaukee, WI, USA). Figure 1 illustrates the dPET and sPET/CT examination processes, including data acquisition, image reconstruction, and metabolic parameter acquisition.

Dynamic K _i images and quantitative metabolic values were obtained based on a two-tissue irreversible compartment model. In this model, it was assumed that ¹⁸F-FDG was taken up unidirectionally (i.e., k ₄ = 0) and was irreversibly trapped in tissue as ¹⁸F-FDG-6-PO (17). The image-derived input function (IDIF) was extracted from the ascending aorta by drawing a region of interest (ROI) with a diameter of 10 mm on six consecutive slices in an image obtained by combining early time frames (0–60 s), in which the effects of motion and partial volume are less pronounced than that in the left ventricle. Two experienced nuclear medicine physicians used the ITK-snap software (version 4.9) to display the 3D volume of interest (VOI) for each LN in the K _i images and to calculate the quantitative values.

Two experienced nuclear medicine physicians independently reviewed the static images. Based on the distribution of the LNs in the puncture and/or pathological findings, the LN long and short diameters were measured on 5-mm CT scans according to the one-to-one correspondence principle, and the LN site and LN-SUV_max were recorded on the sPET/CT scan.

The diagnosis was based on two factors: the appearance under the microscope and the immunohistochemical results. Two experienced pathologists made the diagnosis independently.

Figure 1 illustrates the radiomics feature extraction process. All of the patients’ 2.79-mm PET, 3.75-mm chest CT, and 2.79-mm K _i images were exported to DICOME from the PET/CT workstation. The DICOM format files were then imported into the radiomics version of the uAI Research Portal (version 3.0.1; https://pyradiomics.readthedocs.io/en/) to create outlines and to extract the radiomics features. A junior physician performed manual delineation of the VOI layer-by-layer on the PET (SUV threshold of 40%), CT, and K _i images in a blinded fashion using a software annotation tool. The results outlined by the VOI were then reviewed by another senior doctor.

Prior to the radiomics feature extraction, the distribution of the image voxels in all segmented VOIs was standardized using mean normalization. A total of 4,362 radiomics features were extracted based on the CT, PET, and K _i images, 1,454 of which were CT features, 1,454 were PET features, and 1,454 were K _i features. These radiomics features included: first-order statistics and shape features, gray-level co-occurrence matrix (GLCM) features, gray-level run length matrix (GLRLM) features, gray-level size zone matrix (GLSZM) features, neighboring gray tone difference matrix (NGTDM) features, and gray-level dependence matrix (GLDM) features. Advanced features were achieved using five filters: original, Laplacian of Gaussian (LoG), mean, box mean, and additive Gaussian noise. The parameters were as follows: for original, native image intensities were used without any spatial filtering; for LoG, edge enhancement was performed using 3D LoG filtering with a Gaussian kernel (σ = 3.0 mm); for the mean, uniform mean filtering was applied with a 3 × 3 × 3 voxel smoothing kernel; for the box mean, cubic mean filtering was implemented using a 5 × 5 × 5 voxel kernel; and for additive Gaussian noise, a zero-mean Gaussian noise (10% of the VOI standard deviation) was introduced to simulate acquisition noise.

Figure 1 shows the radiomics feature screening and modeling processes. The extracted radiomics features were then put through a process of Z-score normalization. This was performed so that any differences in the dimensions of the index could be managed. Subsequently, Student’s t-test was used on the training set to compare the features that conformed to a normal distribution in order to distinguish between FDG-avid LN metastasis and non-metastasis. For features that did not follow a normal distribution, the Mann–Whitney U test was used for the initial feature selection. Among these features, Pearson’s correlation coefficient was calculated between each feature–label pair that follows a normal distribution, and features with |r| > 0.6 were selected. LASSO (least absolute shrinkage and selection operator) logistic regression was used to select the radiomics features and to calculate the radiomics score (Rad-score), which was then iteratively validated using 10-fold cross-validation. Three machine learning models were constructed according to the radiomics features of the images: support vector machine (SVM), random forest (RF) classifier, and logistic regression (LR) models.

For the construction of the quantitative model, one-way logistic regression analyses were first performed for SUV_max and K _i. The correlated features were then further incorporated into the multifactor logistic regression to determine the risk predictors. For the construction of the PET/CT+K _i machine learning model, after comparing the efficacy of three machine learning models (i.e., RF, SVM, and LR), the machine learning model with the best overall prediction efficacy was selected to obtain the PET/CT+K _i machine learning model. For the construction of the composite model, a PET/CT+K _i+quantitative composite model was obtained by applying logistic regression analysis to the PET/CT+K_i+quantitative composite model after averaging the weights of the predicted values of the PET/CT+K_i optimal machine learning model and the quantitative model. Figure 1 shows the machine learning modeling, quantitative modeling, and composite model development and evaluation processes.

Statistical analyses and model construction were carried out using the R statistical software package (version 4.1.1) and Python programming language (version 3.7). The groups were compared using the Wilcoxon rank-sum test or the independent-samples t-test. Single-factor and LASSO logistic regression were used to determine the significant risk predictors and radiomics features, as well as the calibration curves. The predictive ability of the model was assessed using receiver operating characteristic (ROC) curves, the area under the curve (AUC), and calibration curves. DeLong’s test was used to determine whether the difference in the efficacy between the models was statistically significant.

The patient and LN features are shown in Table 1 . This study included a total of 279 LNs from 74 patients with lung cancer, of whom 51 (70.83%) were men and 23 (31.93%) were women, with an average age of 61.8 ± 10.0 years. Of the 279 LNs that were pathologically confirmed, 118 (42.29%) were non-metastatic and 161 (57.71%) were metastatic. The enrolled LNs were randomly divided into a training group and a test group in a 7:3 ratio. There were 196 LNs in the training group and 83 LNs in the test group.

Figure 2 shows the process and the results of the screening for radiomics features. A total of 4,362 radiomics features were extracted from the CT, PET, and K _i images: 1,454 from CT, 1,454 from PET, and 1,454 from K _i. A total of 319 radiomics features were selected using Student’s t-test or the Mann–Whitney U test and Pearson’s correlation analysis. The six most significant radiomics features were subsequently selected using LASSO logistic regression. These included one feature for CT (GLCM), three features for PET (GLRLM, GLCM, and GLSZM), and one feature for K _i (GLDM). The final PET/CT+K _i radiomics feature score formula was calculated by summing the coefficients for the retained radiomics features, which were weighted according to their importance.

Rad_score = −0.1979*CTlog_glcm_log.sigma.2.0.mm.3D.InverseVariance+-0.0843*pet_wavelet_glrlm_wavelet.HLH.ShortRunLowGrayLevelEmphasis+-0.08257*pet_wavelet_glrlm_wavelet.HHL.ShortRunLowGrayLevelEmphasis+-0.07745*pet_wavelet_glcm_wavelet.LLL.Correlation+0.035966*pet_wavelet_glszm_wavelet.HLL.SmallAreaLowGrayLevelEmphasis+0.057188*KI_log_gldm_log.sigma.4.0.mm.3D.DependenceNonUniformityNormalized.

Three machine learning models (i.e., SVM, RF, and LR) were constructed using six radiomics features. Figure 3 shows the predictive performance of the three machine learning models for the differential diagnosis of FDG-avid LNs in lung cancer. The ROC curve analysis showed that the RF model had better predictive efficacy in both the training (AUC = 0.823) and test groups (AUC = 0.819). The DeLong’s test showed that, in the training group, there was a statistical difference between LR and RF (p < 0.01), but no statistical difference between LR and SVM or RF and SVM (p = 0.158 and p = 0.058) ( Figure 4 ). There were no statistical differences between LR and RF, LR and SVM, or RF and SVM in the test group (p = 0.671, 0.554, and 0.447, respectively). Table 2 shows the predictive efficacy of the three models. For the RF model, the respective values for AUC, sensitivity, specificity, and accuracy were 0.823 (0.766–0.877), 0.69, 0.819, and 0.745 in the training group and 0.818 (0.727–0.898), 0.667, 0.800, and 0.723 in the test group. Therefore, the RF model was included as a predictive model for the PET/CT+K _i radiomics features.

The quantitative model included SUV_max and K _i. In the training group, the results of the single-factor logistic regression showed that both the SUV_max and K _i were statistically different in the metastatic and non-metastatic groups (p < 0.01, respectively). Multifactor logistic regression analysis showed that K _i differed significantly between the metastatic and non-metastatic groups (p = 0.001), whereas the SUV_max did not (p = 0.917), as shown in Table 3 . Therefore, K _i was included in the quantitative prediction model. Using ROC curve analysis ( Figure 5 ), the AUCs of the quantitative model were 0.772 (0.701–0.831) and 0.805 (0.711–0.893) in the training and test groups, respectively.

A composite prediction model was created based on the RF model and the quantitative model, which was named the PET/CT+K _i+quantitative composite model. According to the ROC curve analysis ( Figure 5 ), the AUC, the sensitivity, the specificity, and the accuracy were respectively 0.844 (0.787–0.894), 0.611, 0.928, and 0.745 in the training group and 0.835 (0.745–0.911), 0.604, 0.943, and 0.747 in the test group. The DeLong’s test showed that, in the training group, the quantitative and composite models had a statistical difference (p = 0.002), while the quantitative and RF models, as well as the RF and composite models, did not (p = 0.120 and p = 0.101) ( Figure 4 ). There were no statistical differences between the quantitative model and the RF model, the quantitative model and the composite model, or the RF model and the composite model (p = 0.750, p = 0.278, and p = 0.382, respectively). Table 4 shows the predictive efficacy of the three models.

Figure 6 shows a nomogram of the composite model, which was constructed based on the machine learning model and the quantitative K _i prediction scores. Agreement between the predicted and the actual values on the nomogram was evaluated. The results of the Hosmer–Lemeshow goodness-of-fit test for both the training and test groups showed no statistical significance (p = 0.978 for the training group and p = 0.227 for the test group), indicating that the predictions of the nomogram constructed in this study were unbiased and a perfect fit, as shown in Figure 7 . The curves demonstrated that the values predicted by the composite model are in close alignment with the actual values.

Figure 8 shows the decision curve analysis of the composite model in predicting the differential diagnosis of FDG-avid LNs in lung cancer. According to the decision curve analysis, the composite model has a better net benefit and clinical value in the differential diagnosis of FDG-avid LNs in lung cancer.