Development and validation of a nomogram for preoperative prediction of the invasiveness of stage T1 lung adenocarcinoma utilizing AI-driven radiomics

OBJECTIVE: To develop and validate an AI-radiomics nomogram that preoperately predicts T1 lung adenocarcinoma invasiveness, providing an objective reference when frozen-section and clinical judgment disagree to optimize surgical extent. METHODS: Clinical data and thin-section computed tomography (CT) images from six centers were analyzed using the Shukun AI workstation, a total of 108 features were included. Patients were classified into invasive (invasive adenocarcinoma, IAC) and non-invasive groups. Data from five centers were randomly allocated to a training set (n = 1066) and an internal validation set (n = 457), with the ratio of pre-invasive to invasive lesions kept consistent between the two sets. External validation was conducted using data from the remaining center (n = 562). Statistical analyses were performed using R software (version 4.2.1). Least absolute shrinkage and selection operator (LASSO) regression was used to select significant features, which were then incorporated into logistic regression for model construction. The performance of the model was evaluated using receiver operating characteristic (ROC) curves, decision curve analysis (DCA), and calibration curves. RESULTS: The regression equation was derived as P = ex/(1 + ex), x = 0.0113×(age) + 0.0059× (median tumor size) − 8.5780× (sphericity) + 1.0377× (GLCM Imc1) + 0.8724× (GLSZM small area emphasis) + 2.0429× (GLDM dependence entropy) − 8.8928. The training set’s ROC AUC was 0.941, sensitivity 0.8602, and specificity 0.8879. The internal validation set showed an AUC of 0.934, and the external set an AUC of 0.905. The model demonstrated high accuracy and clinical applicability. CONCLUSION: In the present study, artificial intelligence (AI) techniques were integrated with radiomics analysis to develop a predictive model for assessing the invasiveness of pulmonary nodules. This model provides valuable support for surgeons to plan surgical procedures and guide intraoperative decision-making. Its utility is particularly prominent in scenarios where there is a discrepancy between the results of intraoperative frozen section (IFS) analysis and the preoperative expectations of clinicians.