深度学习联合C-TIRADS在甲状腺4a类结节风险分层管理的应用  

The application of deep learning combined with C-TIRADS in the risk stratification management of thyroid nodules classified as 4a

作　　者：何美情[1] 张均 高燕华[1] 张茜茜 韩磊[1] 李艳川 HE Meiqing;ZHANG Jun;GAO Yanhua;ZHANG Xixi;HAN Lei;LI Yanchuan(Department of Ultrasound,Shaanxi Provincial People's Hospital,Xi'an 710068,China;Department of Obstetrics,Shaanxi Provincial People's Hospital,Xi'an 710068,China)

机构地区：[1]陕西省人民医院超声科,陕西西安710068 [2]陕西省人民医院产科,陕西西安710068

出　　处：《分子影像学杂志》2024年第9期921-927,共7页Journal of Molecular Imaging

基　　金：陕西省重点研发计划项目(2023-YBSF-465);西安市科技局项目(22YXYJ0108);陕西省人民医院2023年科技发展孵化基金项目(2023YJY-75)。

摘　　要：目的探讨深度学习联合中国超声甲状腺影像报告和数据系统(C-TIRADS)在甲状腺4a类结节风险分层管理的应用。方法纳入陕西省人民医院2018年12月~2022年10月收治的179例甲状腺结节患者,依据病理结果分为良性组(n=76)与恶性组(n=103),所有患者均予以超声检查,按照C-TIRADS指南标准、深度学习进行观察诊断。利用多因素Logistic回归分析获取独立预测指标;利用ROC曲线评估预测变量准确性。结果多因素Logistic分析显示,甲状腺结节图像特征的结构、方位、边缘、回声、局灶性强回声及年龄是独立预测甲状腺结节性质的指标(P<0.05)。以病理结果为金标准,深度学习联合C-TIRADS与病理结果完全符合率为96.65%,Kappa值为0.932,一致性好;联合诊断甲状腺4a类结节的符合率、特异度、阳性预测值显著高于深度学习、C-TIRADS(P<0.05)。联合诊断疾病的敏感度、阴性预测值高于深度学习(P<0.001),但与C-TIRADS的差异无统计学意义(P<0.05)。ROC曲线分析显示,C-TIRADS、深度学习及联合诊断的AUC分别为0.873、0.819、0.967;与Az=0.5相比,差异有统计学意义(P<0.001)。结论C-TIRADS在甲状腺4a类结节风险分层管理中敏感性较高,结合深度学习辅助诊断能够准确进行甲状腺结节良恶性鉴别,具备较高诊断效能。Objective To explore the application of deep learning combined with the Chinese thyroid imaging reporting and data system(C-TIRADS)in the risk stratification management of thyroid nodules classified as 4a.Methods A total of 179 patients with thyroid nodules treated at Shaanxi Provincial People's Hospital from December 2018 to October 2022 were included,divided into benign(n=76)and malignant groups(n=103)based on pathological results.All patients underwent ultrasound examination and were diagnosed using C-TIRADS guidelines and deep learning.Multiple factor Logistic regression analysis was used to obtain independent predictive indicators;the accuracy of predictive variables was assessed using the ROC curve.Results Multiple factor Logistic analysis showed that the structural,directional,edge,echo,focal strong echo,and age characteristics of thyroid nodule images are independent indicators for predicting the nature of thyroid nodules(P<0.05).With pathological results as the gold standard,the complete consistency rate of deep learning combined with CTIRADS with pathological results was 96.65%,and the Kappa value was 0.932,indicating good consistency;the consistency rate,specificity,and positive predictive value of the combined diagnosis for thyroid nodules classified as 4a were significantly higher than those of deep learning and C-TIRADS(P<0.05).The sensitivity and negative predictive value of the combined diagnosis for the disease were higher than those of deep learning(P<0.001),but the difference with C-TIRADS was not statistically significant(P>0.05).ROC curve analysis showed that the AUCs for C-TIRADS,deep learning,and combined diagnosis were 0.873,0.819,and 0.967,respectively;compared with Az=0.5,the differences were all statistically significant(P<0.001).Conclusion C-TIRADS has a high sensitivity in the risk stratification management of thyroid nodules classified as 4a,and combined with deep learning for auxiliary diagnosis,it can accurately distinguish between benign and malignant thyroid nodules,with high dia

关 键 词：甲状腺结节 中国超声甲状腺影像报告和数据系统 深度学习 恶性风险 指南 诊断 

分 类 号：R445.1[医药卫生—影像医学与核医学] R581[医药卫生—诊断学]