基于超声影像构建机器学习模型预测甲状腺良恶性结节  

Predict benign and malignant thyroid nodules using machine learning model based on ultrasound images

作　　者：孙芳[1] 石岩 刘菲菲[2] 邹颖 崔广和[1] 夏爽[4] SUN Fang;SHI Yan;LIU Feifei;ZOU Ying;CUI Guanghe;XIA Shuang(Department of Ultrasound,Binzhou Medical University Hospital,Binzhou 256600,China;Department of Ultrasound,Peking University People’s Hospital;Department of Radiology,First Teaching Hospital of Tianjin University of Traditional Chinese Medicine;Department of Radiology,Tianjin First Central Hospital)

机构地区：[1]滨州医学院附属医院超声医学科,滨州256600 [2]北京大学人民医院超声医学科 [3]天津中医药大学第一附属医院放射科 [4]天津市第一中心医院放射科

出　　处：《国际医学放射学杂志》2021年第4期392-397,共6页International Journal of Medical Radiology

摘　　要：目的构建基于超声影像特征的机器学习模型预测甲状腺结节的良恶性,选择最佳模型以准确预测甲状腺结节的良恶性。方法回顾性分析有明确病理结果的甲状腺结节病人2410例共2516个结节的超声影像特征。使用SPSS Modeler18.0统计软件,将结节随机分为训练队列和验证队列,训练队列包括1992个结节(80%),验证队列包括524个结节(20%)。在训练队列和验证队列中,分别使用支持向量机(SVM)、Logistc回归分析、分类回归树(C&R)、决策树(C5.0)、贝叶斯网络和类神经网络6个分类器构建机器学习模型。采用受试者操作特征(ROC)曲线下面积(AUC)分析模型的原始倾向评分,以评估6种模型的预测能力;并使用DeLong检验比较6种模型的预测能力。选择预测能力最好的机器学习模型,筛选预测重要变量。使用R软件,基于训练队列数据绘制列线图,并基于训练队列及验证队列数据绘制校准曲线对列线图进行验证。结果在训练队列和验证队列中,SVM相比其他模型预测甲状腺结节良恶性的能力最好,AUC分别为0.983和0.973(均P<0.05)。选取SVM筛选的6个预测重要变量绘制的列线图显示纵横比>1、微钙化、包膜外侵犯评分最高,其次为边缘、桥本氏甲状腺炎及回声水平。训练队列和验证队列的校准曲线均显示,列线图的预测结果与实际结果有良好的一致性。结论基于超声影像特征构建的机器学习模型可以准确预测甲状腺结节的性质,其中SVM的预测能力最高。Objective Based on the features obtained from ultrasound images,we built machine learning models to predict benign and malignant thyroid nodules,and selected the best model to accurately predict the benign and malignant thyroid nodules.Methods The ultrasound imaging features of 2516 thyroid nodules with clear pathological results from 2410 patients were analyzed retrospectively.Using SPSS Modeler 18.0 statistical software,the nodules were randomly divided into a training cohort and a verification cohort.The training cohort included 1992 nodules(80%),and the verification cohort included 524 nodules(20%).In the two cohorts,six classifiers were used to build the machine learning models,including support vector machine(SVM),logistic regression analysis,classification regression tree(C&R),decision tree(C5.0),Bayesian network,and neural network.The area under the receiver operating characteristic curve(AUC)of the model’s original propensity score was used to evaluate the predictive abilities of the six models,and DeLong test was used to compare the predictive abilities of the six models.We chose the machine learning model with the best predictive ability to screen important predictive variables.R software was used to draw a nomogram based on the training cohort data,calibration curves draw based on the training cohort and verification cohort were used to verify the nomogram.Results In the training cohort and validation cohort,SVM had the best ability to predict benign and malignant thyroid nodules,with AUC of 0.983 and 0.973,respectively.Compared with the other models,the differences were statistically significant(all P values<0.05).Using the six important predictive variables selected by SVM,the nomogram showed that aspect ratio>1,microcalcification,and extra-thyroidal extension had the highest scores,followed by margins,Hashimoto’s thyroiditis,and echogenicity.The calibration curves based on the training cohort and verification cohort displayed that the nomogram prediction was in good agreement with the actual re

关 键 词：甲状腺结节 机器学习 支持向量机 预测模型 

分 类 号：R736.1[医药卫生—肿瘤] R445.1[医药卫生—临床医学]