To explore the clinical application value of morphological and radiomics feature models based on computed tomography (CT) in the differentiation of benign and malignant sub-centimeter solid pulmonary nodules (SSPNs).

A retrospective analysis was conducted on the CT image data of 203 SSPNs patients from January 2021 to March 2023. They were randomly divided into the training set 143 cases and the test set 60 cases in a ratio of 7︰3. Baseline low-dose chest CT evaluated the clinical and morphological features of SSPNs, while extracting radiological features of the entire lung. The Least Absolute Shrinkage and Selection Operator (LASSO) is used for radiological feature selection and calculating radiological feature scores, and constructing machine learning models. In the test set, the performance of the model is evaluated through the area under the curve (AUC) of receiver operating characteristic(ROC), Hosmer-Lemeshow test, and decision curve.

The detection rate of malignant nodules was 48.28% (98/203). In both the training set and the test set, there were significant differences between the benign group and the malignant group in terms of nodule diameter, vascular convergence sign, nodule-lung interface, spiculation sign, pleural indentation, vacuole sign, and air bronchogram sign (P<0.05). Multivariate Logistic regression analysis showed that nodule diameter, nodule-lung interface, spiculation sign, pleural indentation, and air bronchogram sign were significantly associated with the malignant risk of SSPNs (P<0.05). Morphological feature models were established using these five CT features. A total of 1, 783 high-throughput radiomics features were extracted. After correlation analysis, Mann-Whitney U test, and LASSO contraction, four of the most significant radiomics features were finally screened to establish the model(wavelet-HLL_gldm_DependenceNonUniformity, squareroot_ngtdm_Strength, gradient_glcm_Imc1, original_glrlm_LongRunLowGrayLeve). Based on the Delong test, the AUC of the combined model in both the training set and the test set was significantly higher than that of the morphological feature model (Z=4.618, 2.534, P<0.05) and the radiomics model (Z=2.449, 1.982, P<0.05). After the H-L goodness-of-fit test, the combined model in the training set and the test set (χ²=3.899, 2.815, P<0.05) fitted well. Decision curve analysis shows that when the high-risk threshold is between 0.02 and 0.98, the joint model offers greater net benefits than the single model.

The combined model of CT morphology and radiomics features has good diagnostic performance in differentiating benign and malignant SSPNs and is expected to become an important auxiliary tool for the clinical diagnosis of malignant SSPNs in high-risk surgical cohorts.