人工智能自动检测系统在甲状腺结节术前超声诊断中的应用

目的 探讨人工智能（AI）自动检测系统在甲状腺结节术前超声诊断中的应用价值。方法 选择2019年4月至2019年7月海军军医大学（第二军医大学）长征医院普通外科收治的98例甲状腺结节患者（共137个甲状腺结节），回顾性分析其病理学资料和超声检查结果。所有患者术前均进行常规超声检查和AI自动检测系统检测，以术后病理学结果为金标准，分析比较常规超声检查与AI自动检测系统检测两种方法对甲状腺结节良恶性的诊断效果，计算其灵敏度、特异度、准确度，采用Kappa检验评估两种方法检查结果与术后病理的一致性。结果 常规超声检查诊断甲状腺结节良恶性的灵敏度、特异度、准确度分别为93.75%（90/96）、80.49%（33/41）、89.78%（123/137），AI自动检测系统诊断甲状腺结节良恶性的灵敏度、特异度、准确度分别为89.58%（86/96）、68.29%（28/41）、83.21%（114/137）。Kappa一致性检验结果显示，常规超声检查与病理诊断结果一致性较高（Kappa=0.75，P<0.001）；AI自动检测系统与病理诊断结果一致性一般（Kappa=0.59，P<0.001）。结论 AI自动检测系统诊断甲状腺结节良恶性的灵敏度、准确度稍逊于常规超声检查，但较相近，可作为术前评估甲状腺结节良恶性的有效补充。

Application of artificial intelligence automatic detection system in preoperative ultrasonic diagnosis of thyroid nodules

Objective To explore the application value of artificial intelligence (AI) automatic detection system in preoperative ultrasonic diagnosis of thyroid nodules. Methods Totally 98 patients with 137 thyroid nodules admitted to the General Surgery Department of Changzheng Hospital of Naval Medical University (Second Military Medical University) from April 2019 to July 2019 were enrolled in this study. Pathological data and ultrasonic diagnosis results were retrospectively analyzed. All patients underwent conventional ultrasonography and AI automatic detection before surgery. The diagnoses for benign and malignant thyroid nodules were compared between conventional ultrasonography and AI automatic detection system, which were based on the postoperative pathology. The sensitivity, specificity and accuracy of the two examination methods were calculated, and Kappa coefficient was performed to measure the consistency between the two methods and postoperative pathological diagnosis. Results The sensitivity, specificity and accuracy of conventional ultrasonography in diagnosis of benign and malignant thyroid nodules were respectively 93.75% (90/96), 80.49% (33/41) and 89.78% (123/137), and those of AI automatic detection were 89.58% (86/96), 68.29% (28/41) and 83.21% (114/137). There was substantial coefficience between conventional ultrasonography and pathological diagnosis results (Kappa=0.75, P<0.001), and that was fair coefficience between AI automatic detection system and pathological diagnosis results (Kappa=0.59, P<0.001). Conclusion The sensitivity and accuracy of AI automatic detection system are slightly lower than but close to those of conventional ultrasonography in differentiating benign from malignant thyroid nodules. AI automatic detection system can be used as an effective supplement to assist conventional ultrasonography for preoperative assessment of thyroid nodules.