人工智能辅助诊断系统鉴别甲状腺滤泡性肿瘤  

Artificial intelligence-assisted system to identify follicular thyroid tumours

作　　者：高璐滢 马丽媛 夏宇 安宇昂 潘傲楠 罗能文 顾炯辉 姬姜 姜玉新 Gao Luying;Ma Liyuan;Xia Yu;An Yuang;Pan Aonan;Luo Nengwen;Gu Jionghui;Ji Jiang;Jiang Yuxin(Department of Ultrasound,State Key Laboratory of Compler Severe and Rare Diseases,Peking Union Medical College Hospital,Chinese Academy of Medical Sciences and Peking Union Medical College,Beijing 100730,China)

机构地区：[1]中国医学科学院北京协和医学院北京协和医院超声医学科、疑难重症及罕见病国家重点实验室,北京100730

出　　处：《中华超声影像学杂志》2025年第3期210-215,共6页Chinese Journal of Ultrasonography

基　　金：中央高校基本科研业务费专项资金(3332023011);中央高水平医院临床科研专项(2022-PUMCH-B-066);国家自然科学基金(82402303)。

摘　　要：目的:评估人工智能(AI)辅助诊断系统在诊断甲状腺滤泡性肿瘤中的应用价值,并与不同年资的超声医师的诊断结果进行比较。方法:回顾性收集2016年5月至2018年1月于中国医学科学院北京协和医学院北京协和医院接受手术治疗的86例甲状腺滤泡性肿瘤患者的101个肿瘤。以术后病理为金标准,分为有风险组(29例患者,34个结节,包括滤泡癌15个,恶性潜能不能确定的滤泡性肿瘤19个)和良性组(57例患者,67个结节,包括滤泡型腺瘤15个,结节性甲状腺肿腺瘤样增生52个)。计算AI系统、不同年资的两位超声医师(低年资医师A、高年资医师B)和甲状腺超声恶性风险分层指南[包括2015年美国甲状腺协会(ATA)指南、2017年美国放射学会甲状腺影像报告和数据系统(ACR TI-RADS)、2020年中华医学会超声分会甲状腺影像报告和数据系统(C-TIRADS)](由一位高年资医师C分类)诊断滤泡性肿瘤(有风险组)及滤泡癌的敏感性、特异性和准确性,并进行比较。结果:AI系统诊断滤泡癌的敏感性46.7%,特异性89.6%,准确性81.7%;诊断滤泡性肿瘤(有风险组)的敏感性32.4%,特异性89.6%,准确性70.3%。与低年资医师A相比,AI系统诊断滤泡癌及滤泡性肿瘤(有风险组)的特异性明显更高(89.6%比83.6%,P=0.020;89.6%比73.1%,P=0.020),敏感性差异无统计学意义(46.7%比13.3%,P=0.181;32.4%比11.8%,P=0.073)。AI系统与高年资医师B的敏感性、特异性差异无统计学意义(均P>0.05)。AI系统与低年资医师A、高年资医师B、C-TIRADS、ATA指南和ACR TI-RADS相比,诊断滤泡癌及滤泡性肿瘤(有风险组)曲线下面积差异无统计学意义(均P>0.05)。结论:基于超声AI辅助诊断系统与高年资超声医师诊断甲状腺滤泡性肿瘤的效能相似,AI系统诊断特异性优于低年资超声医师。Objective:To assess the value of artificial intelligence(AI)assisted system in the diagnosis of malignancy in follicular thyroid tumours,and to compare with the diagnostic results of doctors with different levels of experience.Methods:A total of 101 nodules were retrospectively collected from 86 patients with follicular thyroid tumours who underwent surgical treatment at Peking Union Medical College Hospital,Chinese Academy of Medical Sciences,Peking Union Medical College from May 2016 to January 2018.The nodules were classified into risk group(29 patients,34 nodules,including 15 follicular carcinomas and 19 follicular tumours of indeterminable malignant potential)and benign group(59 atients,67 nodules,including 15 follicular adenomas and 52 nodular goitre adenomatoid hyperplasia).The sensitivities,specificities and accuracies of the AI system,two doctors of different seniorities(one junior A and one senior B),and guidelines of thyroid ultrasound malignancy risk stratification[including the 2015 American Thyroid Guidelines(ATA),the 2017 American College of Radiology Thyroid Imaging Reporting and Data System(ACR TI-RADS),the 2020 Chinese Society of Ultrasound,Thyroid Imaging Reporting and Data System(C-TIRADS)](classified by a senior doctor C)for diagnosing follicular tumours in the risk group and follicular carcinomas were calculated and compared.Results:The AI system showed a sensitivity of 46.7%,specificity of 89.6%and accuracy of 81.7%for diagnosing follicular carcinoma;and a sensitivity of 32.4%,specificity of 89.6%and accuracy of 70.3%for diagnosing follicular neoplasms(risk group).Compared with junior doctor A,the specificity of AI system in diagnosing follicular cancer and follicular neoplasms(risk group)was higher(89.6%vs.83.6%,P=0.020;89.6%vs.73.1%,P=0.020),and the differences of sensitivity were not significant(46.7%vs.32.4%,P=0.181;32.4%vs.11.8%,P=0.073).The difference of sensitivity and specificity were not statistically significant between the AI system and senior doctor B(all P>0.05).The differences

关 键 词：超声检查 人工智能 甲状腺滤泡性肿瘤 甲状腺滤泡癌 

分 类 号：R736.1[医药卫生—肿瘤]