基于X线影像组学特征的预测模型对膝关节周围骨肿瘤的诊断价值

Value of a predictive model based on X-ray radiomic features for diagnosing bone tumors around the knee

目的探讨基于数字化X线影像组学特征,并联合常规临床信息构建的预测模型鉴别膝关节周围骨肿瘤良恶性的临床应用价值。方法回顾性收集433例经手术病理证实的膝关节周围良恶性骨肿瘤病人的术前X线影像及临床资料。根据WHO骨肿瘤分类将病人分为良性组(303例)和恶性组(130例)。病人按照7∶3的比例随机分为训练集(303例)及测试集(130例)。采用ITK-SNAP软件在术前膝关节正侧位X线影像上的病灶区域手动勾画感兴趣区(ROI)并提取影像组学特征,利用最小绝对收缩和选择算子(LASSO)回归算法进行特征选择,并采用决策树(DT)、随机森林(RF)、极端梯度提升(XGB)、逻辑回归(LR)、支持向量机(SVM)和K最邻近(KNN)分类器构建单纯影像组学模型,以及整合了临床信息的联合预测模型。采用受试者操作特征(ROC)曲线下面积(AUC)评估各模型的预测效能,并采用DeLong检验比较各模型之间预测效能的差异。利用SHAP值评估模型纳入的各个特征对诊断结果的重要性。结果恶性组的红细胞沉降率(ESR)大于良性组,且关节活动受限多于良性组(均P<0.05),将ESR、关节活动度作为临床特征。基于6种分类器,分别构建单纯影像组学模型(18个组学特征)和联合预测模型(16个组学特征和2个临床特征)。6种分类器构建的影像组学模型和联合模型的诊断效能均较高(均AUC>0.8),其中XGB联合模型的AUC值最高(0.905)。DeLong检验结果显示,XGB、LR和SVM联合模型的AUC值均高于相应的组学模型(均P<0.05)。其中,XGB联合模型的效能最优。通过SHAP值发现组学特征中灰度依赖矩阵(GLDM)为模型提供了重要的预测信息。结论基于膝关节X线影像组学特征和临床信息构建的XGB联合模型可以在术前有效鉴别良恶性骨肿瘤。

Objective To investigate the clinical application value of a predictive model constructed based on the radiomic features of digital X-ray images and combined with routine clinical information for distinguishing between benign and malignant nature of peripheral bone tumors of knee joint.Methods Preoperative X-ray images and clinical data of 433 patients with benign and malignant bone tumors around the knee joint confirmed by surgical pathology were retrospectively collected.According to the WHO bone tumor classification,patients were divided into benign group(303 cases)and malignant group(130 cases).Patients were randomly divided into training set(303 cases)and test set(130 cases)in a 7∶3 ratio.ITK-SNAP software was used to manually outline the region of interest(ROI)in the lesion area on the preoperative frontal and lateral knee X-ray images and extract radiomic features.The least absolute shrinkage and selection operator(LASSO)regression algorithm was utilized for feature selection.Decision tree(DT),random forest(RF),extreme gradient boosting(XGB),logistic regression(LR),support vector machine(SVM),and k-nearest neighbor(KNN)classifiers were used to construct radiomic models and combined prediction models integrating clinical information.The predictive efficacy of each model was assessed using the area under the curve(AUC)of the subject operating characteristic(ROC),and the DeLong test was used to compare the differences in predictive efficacy between the models.SHAP values were utilized to assess the importance of each feature included in the models for diagnostic outcomes.Results The erythrocyte sedimentation rate(ESR)of the malignant group was higher than that of the benign group,and the joint mobility was more restricted than that of the benign group(both P<0.05).ESR and joint mobility were used as clinical features.Based on the six classifiers,the radiomic model(18 radiomic features)and the combined prediction model(16 radiomic features and two clinical features)were constructed.The diagnostic efficacy of all the radiomic and combined model built by the six classifiers were high(all AUC>0.8),and the highest AUC value(0.905)observed for the XGB combined model.The results of DeLong test showed that the AUC values of the combined XGB,LR,and SVM models were higher than those of the corresponding radiomic models(all P<0.05),indicating that the XGB combined model had the optimal performance.SHAP values indicated that the gray-level dependence matrix(GLDM)among the radiomic features provided significant predictive information for the model.Conclusion A combined XGB model based on knee X-ray radiomic features and clinical features can effectively distinguish benign and malignant bone tumors preoperatively.