基于数据挖掘方法的甲状腺结节良恶性CT分类模型的建立  

Establishment of CT Classification Model of Benign and Malignant Thyroid Nodule Based on Data Mining Method

作　　者：梁红琴 朱立强[2] 王健 LIANG Hongqin;ZHU Liqiang;WANG Jian(Department of Radiology,Southwest Hospital,Army Medical University,Chongqing 400038,China;不详)

机构地区：[1]陆军军医大学第一附属医院放射科,重庆400038 [2]解放军第三二二医院质控科,山西大同037006

出　　处：《中国医学影像学杂志》2019年第10期740-744,共5页Chinese Journal of Medical Imaging

摘　　要：目的根据甲状腺结节CT特征指标建立分类模型,以病理结果作为“金标准”,预测甲状腺结节的性质。资料与方法回顾性分析143例甲状腺结节患者的临床和影像学资料,按照病理结果分为良性组(n=56)及恶性组(n=87),分别采用Fisher线性判别方法和非条件Logistic回归算法建立数据挖掘模型,采用受试者工作特征(ROC)曲线下面积比较两个模型的诊断效能。另选取55例甲状腺结节患者进行数据验证。结果两组患者的年龄、结节大小、淋巴结显示及增强强化后CT值比较,差异均有统计学意义(P<0.05)。Fisher判别方法的敏感度为96.6%,特异度91.1%,约登指数为94.4%。非条件Logistic回归法的敏感度为95.4%,特异度为92.9%,约登指数为94.4%。两个模型的ROC曲线下面积比较,差异无统计学意义(Z=0.3,P=0.765)。使用55例甲状腺结节患者进行验证,Fisher判别及非条件Logistic回归敏感度均为66.7%,特异度均为92.3%,约登指数为72.7%,但对具体样本的判别结果有所差异。结论建立的两种模型对甲状腺结节性质的判断结果均较为理想,判别模型与非条件Logistic回归的泛化能力一致。Purpose To establish the classification model according to the CT characteristics of thyroid nodule and predict the nature of thyroid nodule by taking pathological results as gold standard.Materials and Methods Retrospective analysis was conducted for the clinical and imaging data of 143 patients with thyroid nodule.According to pathologic findings,the patients were divided into benign group(n=56)and malignant group(n=87).Fisher linear discriminant analysis method and unconditional Logistic regression algorithm were used to establish the data mining model respectively.The area under the receiver operating characteristic(ROC)curve was used to compare the diagnostic efficacy of the two models.Another 55 patients with thyroid nodule were selected for data verification.Results No statistical difference in age,nodule size,lymph node display and enhanced CT values was observed between the patients of the two groups(P<0.05).The sensitivity,specificity and Yoden index of Fisher discrimination method were 96.6%,91.1%and 94.4%,respectively;those of unconditional Logistic regression algorithm were 95.4%,92.9% and 94.4%,respectively.The areas under ROC curves of the two models were compared and showed statistical significance(Z=0.3,P=0.765).A total of 55 patients with thyroid nodule were used for verification.The sensitivity of Fisher discrimination and unconditional Logistic regression was 66.7%,the specificity was 92.3%,and the Yoden index was 72.7%.However,there were differences in the discrimination results of specific samples.Conclusion The two established models have relatively ideal results in determining the nature of thyroid nodules.The generalization ability of discrimination model is consistent with that of unconditional Logistic regression.

关 键 词：甲状腺结节 体层摄影术 X线计算机 病理学 外科 数据挖掘 算法 诊断 鉴别 

分 类 号：R445.3[医药卫生—影像医学与核医学] R581[医药卫生—诊断学]