基于CT影像组学特征的机器学习模型鉴别甲状腺乳头状癌与结节性甲状腺肿的价值  

Value of machine learning models based on CT radiomics features in differentiating papillary thyroid carcinoma from nodular goiter

作　　者：吕培锋 张若仙 李成威 王霞[1] 彭莉雯 谢汉民 LU Pei-feng;ZHANG Ruo-xian;LI Cheng-wei;WANG Xia;PENG Li-wen;XIE Han-min(Department of Radiology,Guangdong Maternal and Child Health Hospital,Guangzhou 511400,Guangdong,China;Department of Obstetrical,Guangdong Maternal and Child Health Hospital,Guangzhou 511400,Guangdong,China)

机构地区：[1]广东省妇幼保健院放射科,广东广州511400 [2]广东省妇幼保健院产科,广东广州511400

出　　处：《生物医学工程与临床》2025年第2期172-177,共6页Biomedical Engineering and Clinical Medicine

摘　　要：目的探讨基于CT影像组学特征的机器学习模型对甲状腺乳头状癌(PTC)与结节性甲状腺肿(NG)的鉴别诊断价值。方法选择甲状腺结节患者158例,其中男性19例,女性139例;年龄19~72岁,平均年龄43.34岁;结节大小3.0~88.0 mm,平均大小22.16 mm;病理类型,PTC 78例,NG 80例。基于CT增强扫描动脉期,采取半自动逐层勾画感兴趣区(ROI),所有患者按8∶2的比例随机抽样分为训练集126例,验证集32例。采用Python 3.7提取影像组学特征,用t检验、Pearson相关系数筛选及最小绝对收缩和选择算子(LASSO)对特征进行降维筛选,使用逻辑回归(LR)、支持向量机(SVM)、随机森林(RandomForest)、极端梯度提升(XGBoost)、梯度提升决策树(LightGBM)分类器对提取的特征进行机器学习。通过绘制受试者工作特性(ROC)曲线下面积(AUC)、准确度、灵敏度、特异度、F1分数、召回率6个指标评价各模型的差异。结果训练集与验证集中性别、年龄差异均无统计学意义(P>0.05)。从CT增强动脉期提取出1834个影像组学特征,经过t检验、Pearson相关系数筛选和LASSO回归降维筛选出8个最优特征;构建的5个预测模型中,XGBoost模型诊断效能最好,优于其他模型。该模型在训练集中的AUC、准确度、灵敏度、特异度、F1分数和召回率分别为0.993、0.944、0.912、0.983、0.947、0.912。在验证集中,XGBoost模型的AUC、准确度、灵敏度、特异度、F1分数和召回率分别为0.957、0.906、0.700、1.000、0.824、0.700。结论基于CT增强的影像组学特征的机器学习模型可用于鉴别PTC和NG,对甲状腺结节患者的个体化治疗和随访策略提供重要参考价值。Objective To explore the value of machine learning model based on CT radiomics features in differential diagnosis of papillary thyroid carcinoma(PTC)and nodular goiter(NG).Methods A total of 158 patients with thyroid nodules were enrolled,which included 19 males and 139 females,aged 19-72 years old with mean age of 43.34 years old;nodule size was 3.0-88.0 mm with mean size of 22.16 mm;78 cases of PTC and 80 of NG.Based on arterial phase of CT enhanced scan,the region of interest(ROI)was semi-automatically delineated layer by layer,and all patients were randomly divided into training set(n=126)and validation set(n=32)according to ratio of 8∶2.The Python 3.7 was used to extract radiomics features,the t test,Pearson correlation coefficient screening,least absolute shrinkage and selection operator(LASSO)were used to reduce dimension features.Logistic regression(LR),support vector machine(SVM),random forest(RandomForest),extreme gradient boosting(XGBoost),light gradient boosting machine(LightGBM)classifier were used for machine learning of extracted features.The differences of each models were evaluated by 6 indexes including receiver operating characteristic(ROC)area under curve(AUC),accuracy,sensitivity,specificity,F1 score and recall rate.Results There was no significant difference in gender and age between training set and validation set(P>0.05).A total of 1834 image features were extracted from arterion-enhanced CT,and 8 optimal features were selected by t test,Pearson correlation coefficient screening and LASSO regression dimensionality reduction.Among 5 constructed prediction models,the XGBoost model showed the best diagnostic efficiency and was superior to other models.The AUC,accuracy,sensitivity,specificity,F1 score and recall rate of the XGBoost model in training set were 0.993,0.944,0.912,0.983,0.947 and 0.912,respectively.In validation set,the AUC,accuracy,sensitivity,specificity,F1 score and recall rate of XGBoost model were 0.957,0.906,0.700,1.000,0.824 and 0.700,respectively.Conclusion It is demonstr

关 键 词：甲状腺肿瘤 CT 影像组学 影像特征 机器学习 LOGISTIC模型 预测 

分 类 号：R736.1[医药卫生—肿瘤] R730.44[医药卫生—临床医学]