不同机器学习模型在颈部淋巴瘤诊断中的应用与比较  

Application and Comparison of Different Machine Learning Models in Diagnosis of Cervical Lymphoma

作　　者：杨秋婷 王月桂 杨彩云 钟嵘 王康健 沈浩霖 Yang Qiuting;Wang Yuegui;Yang Caiyun;Zhong Rong;Wang Kangjian;Shen Haolin(School of Clinical Medicine,Fujian Medical University,Fuzhou 350122,China;Department of Ultrasound,Zhangzhou Affiliated Hospital of Fujian Medical University,Zhangzhou,Fujian 363005,China)

机构地区：[1]福建医科大学临床医学部,福州市350122 [2]福建医科大学附属漳州市医院超声科,福建省漳州市363005

出　　处：《中国超声医学杂志》2023年第9期974-978,共5页Chinese Journal of Ultrasound in Medicine

基　　金：福建省自然科学基金(No.2022J011478)。

摘　　要：目的基于颈部淋巴结的超声特征,构建并对比不同机器学习模型诊断淋巴瘤的性能。方法纳入颈部淋巴结肿大患者714例,按7∶3将患者随机分为建模队列和验证队列。结合临床数据(年龄、性别和肿瘤史)和淋巴结的超声特征,构建K近邻算法(KNN)、支持向量机(SVM)、决策树(DT)、反向传播算法神经网络(BP)及随机森林(RF)模型。通过计算受试者工作特征(ROC)曲线下面积(AUC)、准确度、精准度、召回率、F1-score及Brier-score等指标比较模型诊断效能。结果KNN模型和RF模型的AUC在训练集中比较高,分别为0.8531和0.8674;在验证集中下降为0.6981和0.6629。在验证集中,DT模型和SVM模型的AUC(分别为0.7586和0.7476)与精准度(分别为0.9130和0.8696)均比其他3个模型高。同时,也相应高于训练集中的指标。DT模型最终纳入了网格状回声、淋巴门、血供模式、年龄和淋巴结短轴径5个变量。结论网格状回声是淋巴瘤最重要的超声特征。与其他4种模型相比,DT模型诊断能力高,且具有更好的泛化能力和可解释性,可以在临床诊断颈部淋巴瘤中发挥重要作用。Objective To construct and analyze the performance of different machine learning models for the diagnosis of lymphoma,based on ultrasonic features of cervical lymph nodes.Methods We enrolled 714 patients with enlarged cervical lymph node,who were randomly split into the modeling cohort(70%)and the validation cohort(30%).The K-nearest neighbor(KNN),support vector machine(SVM),decision tree(DT),back propagation(BP)neural network,and random forest(RF)were constructed combining clinical data(age,gender,and history of tumor)and ultrasound features.The diagnostic performance of the models was analyzed by area under the receiver operating characteristic curve(AUC),accuracy,precision,recall rate,F1-score,and Brier-score.Results In the training set,the AUC values of the KNN and RF model were 0.8531 and 0.8674,respectively;while in the validation set,these decreased to 0.6981 and 0.6629.In the validation set,the AUC values(0.7586 and 0.7476,respectively)and precision values(0.9130 and 0.8696,respectively)of the DT and SVM model were higher than those of the other three models;and the values were also higher than those in the training set.The DT model finally included five variables:grid-like echo,echogenic-hilum,vascular-pattern,age,and diameter on the short axis view of the lymph node.Conclusions Grid-like echo is the most crucial ultrasound feature of lymphoma.Compared with the other four models,DT model has the best diagnostic ability and better generalization ability and interpretability,which can play an important role in clinical diagnosis of cervical lymphoma.

关 键 词：颈部淋巴瘤 机器学习 超声 决策树 

分 类 号：R445.1[医药卫生—影像医学与核医学] R739.91[医药卫生—诊断学]