汕头大学医学院学报

目的：探讨超声人工智能(artificial intelligence,AI)辅助诊断系统在甲状腺结节良恶性诊断中的应用价值。方法：选取2021年11月—2022年2月在汕头大学医学院第二附属医院进行甲状腺常规超声检查的患者217例(428个结节)，男性59例，女性158例，年龄19～75岁，平均(47±13)岁。其中经病理证实的患者77例(162个结节)，男性17例，女性60例；年龄19～75岁，平均(47±13)岁。采用美国放射协会的甲状腺影像报告与数据系统评估甲状腺结节的良恶性，以中心阅片专家的结果为标准，比较住院医师联合AI辅助诊断前后评估428个甲状腺结节的准确度。以病理结果为“金标准”，比较住院医师、住院医师+AI、主治医师、中心阅片专家、AI组评估162个结节的敏感度、特异度、准确度和受试者工作特征曲线下面积(area under curve,AUC)。结果：428个甲状腺结节以中心阅片专家的评估结果为标准，住院医师诊断的准确度为88.32%(378/428)，联合超声AI诊断系统后准确度提高至94.86%(406/428)，差异有统计学意义(χ²=11.89,P=0.001)。162个甲状腺结节以病理结果为金标准，住院医师、主治医师、中心阅片专家、AI组诊断的敏感度分别为43.90%、78.05%、75.61%、75.61%，准确度分别为67.28%、84.57%、85.80%、84.57%。住院医师在超声AI诊断系统的辅助下，诊断的敏感度提高到78.05%，准确度提高到82.72%，与主治医师、中心阅片专家、AI组比较，差异均无统计学意义(P>0.05)。住院医师、住院医师+AI、主治医师、中心阅片专家、AI组的AUC分别为0.596、0.812、0.824、0.816、0.816。住院医师组在超声AI诊断系统的辅助下诊断效能明显提高，与主治医师、中心阅片专家组比较，差异均无统计学意义(P>0.05)。结论：超声AI辅助诊断系统在甲状腺结节良恶性诊断中的具有较高的价值，可以提高住院医师的诊断效能。

Objective:To investigate the value of ultrasound artificial intelligence (AI) assisted diagnosis system in the differential diagnosis of benign and malignant thyroid nodules.Methods:217 patients (428 nodules),59 males and 158 females,aged 19-75 years,with a mean (47±13) years,who underwent routine ultrasonography of the thyroid gland at the Second Affiliated Hospital of Shantou University Medical College,from November 2021 to February 2022,were selected.There were 77 patients (162 nodules),17 males and 60 females,aged 19-75 years with a mean of (47±13) years,with pathologic confirmation.The benign and malignant nature of thyroid nodules was assessed using the thyroid imaging reporting and data system of the American College of Radiology,and the accuracy of the 428 thyroid nodules assessed before and after combined AI-assisted diagnosis by resident physician was compared,using the results of the center's film-reading specialists as the standard.The sensitivity specificity,accuracy,and area under the receiver operator characteristic curve (AUC) of 162 nodules assessed by the resident physician,resident physician+AI,attending physician,center's reading specialists,and AI groups were compared using the pathology results as the gold standard.Results:428 thyroid nodules were diagnosed by residents with an accuracy of 88.32%(378/428) using the evaluation results of the center film-reading specialists as the standard,and the accuracy increased to 94.86%(406/428) after the combined ultrasound AI diagnostic system,with a statistically significant difference (χ²=11.89,P=0.001).162 thyroid nodules were diagnosed by the gold standard of pathological findings,with sensitivities of 43.90%,78.05%,75.61%,75.61%,and accuracies o67.28%,84.57%,85.80%,and 84.57% for the resident,attending physician,center film-reading specialist,and AI groups,respectively.The sensitivity of the resident's diagnosis with the assistance of the ultrasound AI diagnostic system increased to 78.05% and the accuracy increased to 82.72%,and none of the differences were statistically significant when compared with the attending physician,the center reading specialist,and the AI group (P>0.05).The AUC of the resident,resident+AI,attending physician,center film-reading expert,and AI groups were 0.596,0.812,0.824,0.816,and 0.816,respectively.The diagnostic efficacy of the resident group was significantly improved with the assistance of the ultrasound AI diagnostic software,and none of the differences were statistically significant (P>0.05) when comparing with the attending physicians,and center film-reading expert group.Conclusion:Ultrasound AI-assisted diagnostic systems have high value in the diagnosis of benign and malignant thyroid nodules and can improve the diagnostic efficacy of residents.