LIU Jin, WANG Yi, WANG Qin, WANG Yuxiang, HE Lan, JI Lijun, WANG Fei, XIONG Minchao.
Objective To construct a low-dose CT lung cancer screening model based on Artificial Intelligence (AI) and radiomics, and to verify the model. Methods A total of 460 patients with pulmonary nodule in Ezhou Central Hospital from January 2020 to June 2025 were collected and divided into training set (322 cases) and validation set (138 cases) by a ratio of 7:3. According to the pathological results, 322 patients in the training set were divided into benign group (217 cases of benign nodules) and malignant group (105 cases of malignant nodules). The clinical data of the two groups were compared. LASSO regression was used to screen the radiomics features, and multivariate Logistic regression was used to analyze the influencing factors of lung cancer. The diagnostic models of Logistic Regression (LR), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), and k-Nearest Neighbors (KNN) were constructed in the training set and validation set, and the Receiver Operating Characteristic (ROC) Curve was used to evaluate the efficacy of the models. Results There were statistically significant differences in pleural indentation sign, vascular convergence sign, bronchial sign, spiculation sign, Cyto-keratin 19 fragment antigen 21-1 (CYFRA21-1) and Carcinoembryonic Antigen (CEA) between malignant group and benign group (P＜0.05). LASSO regression was used to screen 7 radiomic features and construct a radiomic score. Pleural indentation sign (OR=4.843, 95%CI: 2.685~8.737), vascular convergence sign (OR=7.586, 95%CI: 3.658~ 15.732), bronchial sign (OR=4.007, 95%CI: 2.257~7.115), spiculation sign (OR=3.014, 95%CI: 1.659~5.477), CYFRA21-1 (OR=1.480, 95% CI: 1.136~ 1.927), CEA (OR=1.399, 95% CI: 1.237~ 1.582) and radiomic score (OR=2.076, 95% CI: 1.584~2.721) were all independent influencing factors for lung cancer (P＜0.05). The results of training set and validation set showed that the AUC value of XGBoost model was the largest, which was significantly larger than that of LR, SVM and KNN models. Conclusion There are many influencing factors of lung cancer. Based on the above factors, a lung cancer screening model is constructed, and the XGBoost model has good diagnostic efficacy.
