摘要
目的:采用4种不同的图像预处理方法,并结合多种分类器构建CT影像组学模型进行对比分析,旨在探究不同归一化方法的模型对诊断骨质疏松症(OP)效能的影响。方法:回顾性纳入2020年1月至2025年1月接受腹部CT检查的患者,并根据双能X线吸收法(DXA)结果将患者分为非OP组和OP组,测量每位患者L1~L4椎体中部横断面CT值,分析各椎体诊断OP的性能,并建立临床模型。利用最大最小归一化(预处理A)、均值方差归一化(预处理B)、窗宽窗位归一化(预处理C)及无归一化(预处理D)对腹部CT图像进行处理,采用支持向量机(SVM)、逻辑回归(LR)、随机森林(RF)及极端梯度提升(XGBoost)分类器构建影像组学模型,计算曲线下面积(AUC)、灵敏度、特异度和准确度,评估各模型的性能。结果:两组间各椎体CT值均与骨密度值(BMD)存在较强的相关性,其中以L2椎体CT值与L1-BMD及L1~L4BMD值相关性最强(均r=0.695,P<0.001),L1~L4椎体平均CT值诊断OP效能高于其他椎体CT值,截断值为89.3HU(AUC=0.896,P<0.05)。年龄、肌少症及性别是OP的独立预测因素(均P<0.05),以预处理C图像预处理后采用RF分类器构建的组学其诊断效能表现较佳(AUC=0.964),两者的联合模型展现最优的效能(AUC=0.967),高于临床模型(AUC=0.880,P<0.05)及CT值模型而与组学模型AUC无统计学意义(AUC=0.964,P=0.139)。结论:基于窗宽窗位归一化图像预处理方法后采用RF分类器结合临床独立预测因素构建的联合模型可有效诊断OP。
Purpose:To construct and compare CT-based radiomics models using four different image preprocessing methods in combination with various classifiers,and to investigate the impact of different normalization techniques on model performance for diagnosing osteoporosis(OP).Methods:This retrospective study included patients who underwent abdominal CT scans from January 2020 to January 2025.Patients were categorized into non-osteoporosis and osteoporosis groups based on dual-energy X-ray absorptiometry(DXA)results.The CT values at the mid-transverse level of the Ll to L4 vertebrae were measured for each patient.The diagnostic performance of each vertebra for OP was analyzed,and a clinical model was established.Abdominal CT images were processed using four preprocessing approaches:(A)min-max normalization,(B)mean-variance normalization,(C)window width/level normalization,and(D)no normalization.Radiomics models were then constructed using support vector machine(SVM),logistic regression(LR),random forest(RF),and extreme gradient boosting(XGBoost)classifiers.Model performance was evaluated using the area under the receiver operating characteristic curve(AUC),sensitivity,specificity,and accuracy.Results:Vertebral CT values showed strong correlations with bone mineral density(BMD)in both groups.The strongest correlation was observed between L2vertebral CT value and L1 BMD,as well as the average L1-L4 BMD(both r=0.695,P<0.001).The average CT value of L1-L4 vertebrae demonstrated a higher diagnostic efficacy for OP than individual vertebrae CT values,with a cutoff of 89.3 HU(AUC=0.896,P<0.05).Age,sarcopenia,and gender were identified as independent predictors of OP(all P<0.05).Among the radiomics models,the one built using the RF classifier on images preprocessed with method C(window width/level normalization)demonstrated the superior diagnostic efficacy(AUC=0.964).The combined model integrating this radiomics model with the clinical predictors achieved the optimal performance(AUC=0.967),which was significantly higher than that of the clinical model alone(AUC=0.880,P<0.05).There was no statistically significant difference between the AUC of this combined model and that of the optimal radiomics model alone(AUC=0.964,P=0.139).Conclusion:A combined diagnostic model,which integrates a radiomics model based on window width/level normalized images using an RF classifier with independent clinical predictors,demonstrates effective performance for diagnosing OP.
作者
廖雯欣
张保朋
周聪
高攀
李依桐
王道清
LIAO Wenxin;ZHANG Baopeng;ZHOU Cong;GAO Pan;LI Yitong;WANG Daoqing(Department of Radiology,The First Affiliated Hospital of Henan University of Chinese Medicine;First School of Clinical Medicine,Henan University of Chinese Medicine;Thrid School of Clinical Medicine,Henan University of Chinese Medicine)
出处
《中国医学计算机成像杂志》
北大核心
2026年第1期101-108,共8页
Chinese Computed Medical Imaging
