Development and validation of CT radiomics diagnostic models: differentiating benign from malignant pulmonary nodules and evaluating malignancy degree

Background: Lung cancer (LC) is the most prevalent malignancy in China. Early diagnosis is crucial as the 5-year survival rate varies greatly by stage. Radiomics, distinct from invasive pathological diagnosis, can extract features from medical images, offering a new approach for pulmonary nodule (PN) diagnosis. This study aimed to use radiomics to develop models for differentiating <3 cm PNs and assessing malignancy levels to guide early-stage LC treatment and surgical decisions. Methods: A total of 202 eligible patients with PNs who had surgical resection at First Affiliated Hospital of Soochow University (Sep 2022-Sep 2023) were included. They were divided into three groups based on pathology: benign (Group A, n=33), low-grade malignant (Group B, n=77), and high-grade malignant (Group C, n=92). Stratified random sampling created training and validation groups. Univariate and multivariate logistic regression identified risk factors for constructing clinical-radiological models [CM(I) & CM(II)]. Radiomics features were extracted from computed tomography (CT) images, screened by intraclass correlation coefficient (ICC) and least absolute shrinkage and selection operator (LASSO) regression. Radiomics score (Rad score) was calculated for radiomics models [RM(I) & RM(II)]. Composite models [COM(I) & COM(II)] integrated Rad score and risk factors. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). Results: Within the training and validation groups for the analysis of benign versus malignant nodules, RM(I) and COM(I) outperformed CM(I), with RM(I) having areas under the ROC curve (AUCs) of 0.895 (training) and 0.808 (validation), COM(I) 0.927 and 0.854, and CM(I) 0.763 and 0.823. Within the training and validation groups for the analysis of malignancy levels, RM(II) and COM(II) were superior to CM(II), with RM(II) AUCs of 0.966 (training) and 0.959 (validation), COM(II) 0.972 and 0.967, and CM(II) 0.924 and 0.950. Specific sensitivity, specificity, and balanced accuracy were calculated, demonstrating that radiomics could significantly enhance the prediction performance for malignant nodules. Conclusions: Radiomics-based RMs showed good diagnostic performance in differentiating <3 cm lung nodules and assessing malignancy. COMs, which combined independent predictors and RMs, had better diagnostic performance than CMs, indicating potential for clinical use. These models can guide treatment decisions, such as conservative management for benign-predicted nodules, sublobar resection for low-grade malignancies, and radical lobectomy with lymph node dissection for high-grade malignancies.