CT影像组学鉴别儿童化脓性与结核性淋巴结炎  

CT radiomics in differentiation between suppurative from tuberculous lymphadenitis in children

作　　者：张蕊 徐晔[1] 李伟[1] ZHANG Rui;XU Ye;LI Wei(Department of Radiology,Children's Hospital of Chongqing Medical University,National Clinical Research Center for Child Health and Disorders,Ministry of Education Key Laboratory of Child Development and Disorders,Chongqing Key Laboratory of Pediatrics,Chongqing 400014,China)

机构地区：[1]重庆医科大学附属儿童医院放射科,国家儿童健康与疾病临床医学研究中心,儿童发育疾病研究教育部重点实验室,儿科学重庆市重点实验室,重庆400014

出　　处：《放射学实践》2023年第12期1599-1604,共6页Radiologic Practice

摘　　要：目的:探讨基于CT增强静脉期图像采用机器学习方法构建的影像组学模型结合多期CT图像上的影像学特征对儿童颈部伴有坏死的化脓性淋巴结炎与结核性淋巴结炎的鉴别诊断价值。方法:搜集2014年9月-2022年5月我院CT增强检查发现颈部淋巴结坏死并经病理活检或临床确诊为颈部化脓性淋巴结炎(n=52)或颈部结核性淋巴结炎(n=49)患儿的临床和影像资料。分析两组患儿10个主要CT征象(最大坏死淋巴结短径、最大坏死结的坏死区/淋巴结面积比、坏死区内有无分隔、坏死区边缘是否光滑、平扫坏死区是否可见、有无钙化、淋巴结强化形状、淋巴结强化程度、淋巴结边缘、淋巴结融合)的差异。使用Radcloud平台,在CT增强静脉期图像上于病变淋巴结内勾画ROI并提取1409个影像组学特征。利用方差阈值、SelectKBest、最小绝对收缩和选择算法对提取的特征逐步进行降维和筛选,随后采用6种机器学习方法(k-最邻近算法、支持向量机、极限梯度提升、随机森林、逻辑回归和决策树)构建预测模型,采用受试者工作特征(ROC)曲线及符合率、敏感度和特异度等指标来评估模型对两种不同类型淋巴结炎的预测效能。结果:两组之间最大短径、淋巴结边缘、淋巴结融合、坏死区边缘是否规则、坏死区有无分隔、钙化及强化程度这7个征象的差异有统计学意义(P<0.05)。使用方差阈值、SelectKBest、最小绝对收缩和选择算法逐步筛选出9个最佳影像组学特征,采用机器学习方法构建的6种预测模型中,AUC分别为0.76(95%CI:0.54~0.97)、0.89(95%CI:0.72~1.00)、0.72(95%CI:0.52~0.93)、0.66(95%CI:0.45~0.88)、0.81(95%CI:0.60~1.00)和0.58(95%CI:0.35~0.80),其中以支持向量机模型的预测效能最高,其诊断符合率、敏感度和特异度分别为0.88、0.78和0.90。结论:CT表现结合发病年龄和影像组学模型对鉴别儿童颈部伴有坏死的化脓性淋巴结�Objective:To investigate the differential diagnosis value of the venous-phase CT(computed tomography)radiomics model constructed by machine learning combined with the multi-phase CT features in differentiating suppurative lymphadenitis and tuberculous lymphadenitis with necrosis in the neck of children.Methods:The clinical and CT imaging data of 52 children with necrotic suppurative lymphadenitis and 49 children with necrotic tuberculous lymphadenitis confirmed by pathology or clinic from September 2014 to May 2022 were collected.Ten major CT features(the short diameter of the largest necrotic lymph node,the ratio of the short diameter of the necrotic area to that of the largest necrotic lymph node,monolocular/multilocular necrotic areas,sharp/obscure node border,obvious/inconspicuous necrotic area in the nonenhancement phase,node calcification,annular enhancement of the node,nodal attenuation value in the venous phase,regular/irregular silhouette of the necrotic area and fusion of nodes)were identified for statistical analysis.Based on the venous-phase CT image,the lymph node lesions were delineated on the Radcloud platform,and 1409 quantitative radiomic features were extracted.Dimension reduction and screening of these features were gradually carried out using variance threshold,SelectKBest and the least absolute shrinkage and selection operator(LASSO).And then,six prediction models were established through six machine learning method,including k-Nearest Neighbor(KNN),support vector machine(SVM),extreme gradient boosting(XGBoost),random forest(RF),logistic regression(LR)and decision tree(DT).The predictive efficacy of these models for the two different types of lymphadenitis were evaluated using receiver opera-ting characteristic(ROC)curve,accuracy,sensitivity and specificity.Results:There were statistically significant differences between the two groups in the short diameter of the largest necrotic node,the clear edge of the lymph node,the fusion of the lymph nodes,the neat edge of the necrotic area,the se-par

关 键 词：淋巴结炎 结核病 化脓性感染 儿童 影像组学 体层摄影术 X线计算机 预测模型 机器学习 

分 类 号：R814.42[医药卫生—影像医学与核医学] R551.2[医药卫生—放射医学] R632.7[医药卫生—临床医学]