CT平扫影像组学预测甲状腺乳头状癌中央区隐匿性淋巴结转移的价值

Value of CT plain scan radiomics in predicting occult central lymph node metastasis in papillary thyroid carcinoma

目的:探讨基于CT平扫的影像组学术前预测甲状腺乳头状癌(PTC)中央区隐匿性淋巴结转移(OLNM)的价值。方法:回顾性分析2家医院(中心1和中心2)经病理证实为PTC患者的临床影像资料。中心1纳入394例,以7∶3的比例随机分为训练集276例和内部测试集118例;中心2纳入143例作为外部测试集。从CT平扫图像中提取病灶的影像组学特征,通过降维获得最佳特征,建立5种机器学习分类器。选择内部测试集和外部测试集中平均AUC最高的分类器作为最佳影像组学模型,并将其结果转换为影像组学评分。将单因素分析中P<0.05的临床和常规CT特征纳入多因素logistic回归分析,得到与OLNM相关的危险因素并建立临床模型。基于临床危险因素和影像组学评分构建联合模型,并绘制列线图。采用ROC曲线评价预测模型的性能。结果:从CT平扫图像中获得10个最佳影像组学特征。在内部测试集和外部测试集中,极端梯度提升具有最佳的预测性能(平均AUC为0.782)。进一步将性别、肿瘤最大径与影像组学评分相结合建立联合模型。联合模型在训练集、内部测试集和外部测试集中预测PTC中央区OLNM的AUC分别为0.869、0.823、0.802。结论:基于CT平扫的影像组学特征对PTC中央区OLNM具有较好的预测价值,进一步结合临床特征建立的联合模型能更好地提升性能。

Objective:To explore the value of radiomics based on CT plain scan in preoperative prediction of central occult lymph node metastasis(OLNM)of papillary thyroid carcinoma(PTC).Methods:A retrospective analysis of clinical imaging data of PTC patients confirmed by pathology in Center 1 and Center 2 was made.Center 1 included 394 patients,who were randomly divided into a training set(276 cases)and an internal test set(118 cases)in a ratio of 7∶3,and Center 2 included 143 patients as an external test set.Radiomic features of lesions were extracted from CT plain scan images,the optimal features were obtained through dimensionality reduction,and five machine learning classifiers were established.The classifier with the highest average AUC value in the internal and the external test set was selected as the optimal radiomics model,and its results were converted into radiomics scores(Rad-scores).Clinical and conventional CT features with statistical difference in univariate analysis were included in multivariate logistic regression analysis to identify risk factors associated with OLNM and establish a clinical model.Subsequently,a combined prediction model was constructed based on clinical risk factors and Rad-scores,and a nomogram was drawn.ROC curve was used to evaluate the efficiency of the combined prediction model.Results:Ten optimal radiomics features were obtained from CT plain scan images.In the internal and external test set,extreme gradient boosting had the best predictive performance(average AUC of 0.782).The combined prediction model was further established with gender,maximum tumor diameter and Rad-scores,and AUCs of the combined prediction model in diagnosing PTC central OLNM in the training set,internal test set and external test set were 0.869,0.823 and 0.802,respectively.Conclusions:The radiomics features based on CT plain scan have good predictive values for central OLNM in PTC,and the combined prediction model established by clinical features and Rad-scores can better improve the diagnostic efficiency.