基于胸部CT平扫的机器学习影像组学模型筛查骨质疏松的研究

Study on Screening Osteoporosis Using Machine Learning Radiomics Model Based on Chest CT Plain Scan

摘要: 目的：利用常规胸部CT平扫的T12、L1椎体信息及对应DXA骨密度检查结果，结合机器学习与影像组学方法建立辅助筛查模型，实现胸部CT检查人群的骨质疏松机会性快速筛查。方法：回顾性收集青岛大学附属医院2024年1月至2025年6月行胸部CT平扫及DXA检查(间隔 < 30日)的600例患者，按7:3随机分为训练集(420例)和内部验证集(180例)；另纳入济宁市第一人民医院同期263例患者作为外部验证集。采用ITK SNAP3.8软件勾画T12、L1椎体ROI，通过Pyradiomics模块提取影像组学特征，经Mann-Whitney U检验、Spearman相关性分析、mRMR、LASSO及10倍交叉验证筛选特征，构建LR、KNN、SVM等6种机器学习模型；结合临床独立影像因子构建临床模型及列线图模型，以AUC、准确性等指标评估模型性能。结果：每例患者提取1688个影像组学特征，经筛选最终保留25个最佳特征；6种模型中LR模型AUC最高，确定为最终影像组学算法。影像组学模型在训练集、内部验证集、外部验证集的AUC分别为0.880、0.803、0.848，准确率分别为0.838、0.767、0.791；临床模型对应AUC分别为0.759、0.746、0.765，准确率分别为0.757、0.744、0.787；列线图模型对应AUC分别为0.897、0.849、0.875，准确率分别为0.790、0.756、0.871。Delong检验显示，列线图模型及影像组学模型预测效能均显著高于临床模型(均P < 0.05)。结论：基于胸部CT平扫影像组学特征构建的机器学习模型可有效区分骨质疏松与非骨质疏松患者；结合临床变量与影像组学特征的列线图模型预测效能最佳，能更全面精准评估骨质疏松风险，为临床早期筛查及个性化干预提供可靠量化工具。

Abstract: Objective: To establish an auxiliary screening model for opportunistic rapid screening of osteoporosis in patients undergoing chest CT examination by utilizing T12 and L1 vertebral information from routine non-contrast chest CT scans along with corresponding DXA bone mineral density results, combined with machine learning and radiomics methods. Methods: A retrospective collection of 600 patients who underwent non-contrast chest CT and DXA examination (interval < 30 days) at the Affiliated Hospital of Qingdao University from January 2024 to June 2025 was conducted. Patients were randomly divided into a training set (420 cases) and an internal validation set (180 cases) at a 7:3 ratio. Additionally, 263 patients from Jining First People’s Hospital during the same period were included as an external validation set. Regions of interest (ROIs) of T12 and L1 vertebrae were delineated using ITK-SNAP version 3.8 software, and radiomic features were extracted using the Pyradiomics module. Feature selection was performed using Mann-Whitney U test, Spearman correlation analysis, mRMR, LASSO, and 10-fold cross-validation. Six machine learning models (LR, KNN, SVM, etc.) were constructed. Clinical and nomogram models were developed by incorporating independent clinical imaging factors. Model performance was evaluated using AUC, accuracy, and other metrics. Results: A total of 1688 radiomic features were extracted per patient, with 25 optimal features retained after selection. Among the six models, the LR model achieved the highest AUC and was selected as the final radiomic algorithm. The AUCs of the radiomic model in the training, internal validation, and external validation sets were 0.880, 0.803, and 0.848, with accuracies of 0.838, 0.767, and 0.791, respectively. The clinical model achieved corresponding AUCs of 0.759, 0.746, and 0.765, with accuracies of 0.757, 0.744, and 0.787. The nomogram model achieved corresponding AUCs of 0.897, 0.849, and 0.875, with accuracies of 0.790, 0.756, and 0.871. Delong tests showed that the predictive performance of both the nomogram model and the radiomic model was significantly superior to that of the clinical model (all P < 0.05). Conclusion: Machine learning models based on radiomic features from non-contrast chest CT scans can effectively distinguish between osteoporotic and non-osteoporotic patients. The nomogram model combining clinical variables and radiomic features demonstrates the best predictive performance, providing a more comprehensive and accurate assessment of osteoporosis risk and serving as a reliable quantitative tool for early clinical screening and personalized intervention.