基于CT影像组学的机器学习模型在术前鉴别磨玻璃结节良恶性的研究

A Study on Machine Learning Models Based on CT Radiomics for Preoperative Differentiation of Benign and Malignant Ground-Glass Nodules

摘要: 目的：探究基于CT平扫影像组学的机器学习模型在术前鉴别磨玻璃结节良恶性中的价值。方法：本研究采用回顾性分析，分析2020年6月~2024年6月在大理大学第一附属医院行胸部CT平扫的121例明确GGN良恶性患者的CT影像。使用ITK-SNAP软件勾画感兴趣区(region of interest, ROI)，采用PyRadiomics软件提取特征，通过Lasso回归算法降维。采用logistic回归算法建模，对建立的模型分别用受试者工作特征曲线(receiver operating characteristic curve, ROC)和曲线下面积(area under the curve, AUC)评估诊断效能。结果：AUC在训练组中影像组学模型、临床模型和联合诊断模型分别为0.784 (95%CI: 0.692~0.872)、0.770 (95%CI: 0.660~0.856)和0.836 (95%CI: 0.755~0.908)；在验证组中分别为0.770 (95%CI: 0.410~1.000)、0.725 (95%CI: 0.455~0.946)和0.830 (95%CI: 0.658~0.968)。结论：基于CT影像组学特征联合临床和常规CT影像特征构建的联合模型具有最佳预测肺GGN的良恶性的能力，基于影像组学特征构建的机器学习模型具有良好术前区分良性和恶性GGN的诊断能力。

Abstract: Objective: To explore the value of a machine learning model based on non-contrast CT radiomics in the preoperative differentiation of benign and malignant ground-glass nodules (GGNs). Methods: A retrospective analysis was performed in this study, involving the non-contrast chest CT images of 121 patients with pathologically confirmed benign or malignant ground-glass nodules (GGNs). These patients underwent non-contrast chest CT examinations at the First Affiliated Hospital of Dali University from June 2020 to June 2024. The region of interest (ROI) was segmented using ITK-SNAP software, and features were extracted with PyRadiomics software, followed by dimensionality reduction via the Lasso regression algorithm. A logistic regression algorithm was used for model construction, and the diagnostic performance of the established models was evaluated using the receiver operating characteristic (ROC) curve and the area under the curve (AUC).Results: In the training group, the AUC values of the radiomics model, the clinical model, and the combined diagnostic model were 0.784 (95%CI: 0.692~0.872), 0.770 (95%CI: 0.660~0.856), and 0.836 (95%CI: 0.755~0.908) respectively. In the validation group, the corresponding AUC values were 0.770 (95%CI: 0.410~1.000), 0.725 (95%CI: 0.455~0.946), and 0.830 (95%CI: 0.658~0.968). Conclusion: The combined model constructed by integrating CT radiomic features with clinical and conventional CT imaging features exhibits the optimal ability to predict the benignancy and malignancy of pulmonary GGNs. The machine learning model based solely on radiomic features demonstrates good diagnostic performance in the preoperative differentiation of benign and malignant GGNs.