Abstract
BACKGROUND: Preoperative identification of early-stage lung adenocarcinoma, particularly minimal lesions, holds critical implications for clinical decision-making. However, accurate diagnosis of pulmonary nodules' malignancy remains challenging when relying solely on conventional clinical parameters or radiological features, due to the lack of specificity in certain instances of early lesions. This study aims to develop a nodule-level composite model using radiomics methods and compare its performance with the Lung CT Screening Reporting and Data System (Lung-RADS) in predicting the invasiveness using preoperative low-dose computed tomography (LDCT) scan with 1 mm slice thickness. METHODS: From June 2019 to October 2023, the patients who were pathologically diagnosed with the initial primary lung nodules were enrolled in the retrospective cohort. To predict invasiveness of nodule, nodule-level models, including radiomics, radiological, and combined models, were developed to categorize the nodules into the invasiveness and non-invasiveness groups. Radiomics features were extracted from the tumor region to develop a radiomics model after a three-stage feature reduction procedure, resulting in a Radscore for each nodule. The nodules were randomly divided into two groups, with a ratio of 7:3. The area under the receiver operating characteristic curve (AUC), accuracy, and other relevant measures were employed to assess the models' performance and discriminative ability. A nomogram that combines optimal radiomics and clinical-radiological features was compared to Lung-RADS using various evaluation methods, including calibration curve analysis, decision curve analysis (DCA), net reclassification improvement (NRI), and integrated discrimination improvement (IDI). RESULTS: A total of 393 patients were enrolled, including 67 with multiple nodules. Altogether, 487 nodules were analysed, comprising 374 in the invasiveness and 113 in the non-invasiveness group. The radiological model identified two risk factors: nodule type and a vessel convergence sign. The nomogram achieved a training-set AUC of 0.91, surpassing the Lung-RADS system (AUC: 0.669) in nodule assessment. Notably, significant improvements in NRI and IDI, as well as higher net benefits, were shown by the DCA. CONCLUSIONS: The nomogram based on radiomics and radiological features offers supplementary information for preoperative prediction of the invasiveness of lung nodules, surpassing both the commonly used Lung-RADS and single models in guiding diagnosis and treatment strategies for lung nodules.