ACR TI-RADS与基于人工智能的报告系统对甲状腺结节的诊断效能及减少不必要穿刺能力的比较

A comparison between ACR TI-RADS and artificial intelligence TI-RADS regarding to diagnostic efficacy and ability to reduce unnecessary fine-needle aspiration cytology

目的比较美国放射学会(American College of Radiology,ACR)2017年提出的甲状腺影像报告和数据系统(Thyroid Reporting And Data System,TI-RADS)与最新基于人工智能(artificial intelligence,AI)的报告系统在甲状腺结节良恶性鉴别中的准确性及临床价值。方法回顾性分析2019年1-12月于东部战区总医院超声诊断科行超声检查并行超声引导下甲状腺细针穿刺活检(ultrasound-guided fine-needle aspiration cytology,US-FNAC)的266例患者资料(共穿刺结节276枚),左叶119枚,右叶157枚。通过绘制ACR TI-RADS与AI TI-RADS两种甲状腺影像报告与评分系统的ROC曲线,计算并比较曲线下面积(AUC)。结果AI TI-RADS的AUC为0.853(95%CI=0.806~0.899),大于ACR TI-RADS的0.754(95%CI=0.689~0.804),差异有统计学意义(Z=2.816,P=0.002)。AI TI-RADS的敏感性、特异性、阳性预测值、阴性预测值(96.62%、62.50%、74.87%、94.12%)均高于ACR TI-RADS(95.27%、44.53%、66.51%、89.06%)。AI TI-RADS能够减少71.74%不必要的FNAC,ACR TI-RADS可减少67.03%不必要的FNAC。结论两种超声影像TI-RADS分类系统对甲状腺结节良恶性均有较好的诊断效能;AI TI-RADS相较于ACR TI-RADS具有更好的诊断效能和减少不必要的FNAC的能力。

Objective To compare the diagnostic efficacy and ability of Thyroid Imaging Reporting Data System version(TI-RADS)of American College of Radiology(ACR)and artificial intelligence(AI)TI-RADS in diagnosis of thyroid nodules.Methods A retrospective analysis was done on 266 patients(276 nodules)proved by ultrasound-guided fine-needle aspiration cytology(US-FNAC)in General Hospital of Eastern Theater Command from January to December 2019.The ROC curve of the two TI-RADS versions was drawn and the area under the curve(AUC)was calculated and compared.Results AUCs of ACR TI-RADS and AI TI-RADS were 0.747 and 0.853.The sensitivity,specificity,positive predictive value and negative predictive value(96.62%,62.50%,74.87%,94.12%)of AI TI-RADS were higher than ACR TI-RADS(95.27%,44.53%,66.51%,89.06%).AI TI-RADS was able to avoid more unnecessary FNAC(71.74%)than ACR TI-RADS(67.03%).Conclusions Both ACR TI-RADS and AI TI-RADS have good performances for differential diagnosis of thyroid nodules.AI TI-RADS is a more simple scoring system with better overall diagnostic performance and ability to exclude unnecessary FNAC with high negative predictive value than ACR TI-RADS.