Objective  To investigate the clinical value of an ¹⁸F-FDG PET/CT radiomics model in differentiating benign from malignant solitary pulmonary nodules with negative tumor markers and positive FDG uptake.

Methods The data of 130 patients with such pulmonary nodules admitted to the Department of Nuclear Medicine of Baoding No. 1 Central Hospital from January 2019 to December 2024 were collected. The patients were randomly divided into a training set (n = 91) and an internal test set (n = 39) at a ratio of 7:3. Eligible independent cases were additionally collected from the Cancer Imaging Archive (TCIA) public database as an external validation set (n = 45). PET and CT radiomics features were extracted using LIFEx software. Core features were selected through intraclass correlation coefficient (ICC) stability assessment, Spearman correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression. Logistic regression (LR), random forest (RF), and support vector machine (SVM) models were constructed separately. Receiver operating characteristic (ROC) curves and decision curve analysis (DCA) were used to evaluate and compare the performance of the models.

Results A total of 8 key radiomics features were ultimately selected, including 4 PET features and 4 CT features. In the internal test set, the LR model showed the best performance, with an area under the curve (AUC) of 0.819 (95% CI: 0.738-0.892), a sensitivity of 0.792, and a specificity of 0.800. In the external validation set, the LR model also maintained the best discriminative performance, with an AUC of 0.801 (95% CI: 0.713-0.905), which was significantly higher than those of the RF model (AUC = 0.721, P = 0.041) and the SVM model (AUC = 0.717, P = 0.028). DCA demonstrated that the LR model yielded greater clinical net benefit in both the internal test set and the external validation set.

Conclusions For difficult-to-differentiate solitary pulmonary nodules with negative tumor markers and high FDG uptake, PET/CT radiomics models, particularly the LR model, have good discriminative performance and generalizability and can provide an objective quantitative reference for preoperative decision-making.