| 本文已被：浏览 592次 下载 572次 |
码上扫一扫！ |
| 基于AI-SONICTM Thyroid 5.3.3.0的超声图像分析对甲状腺结节恶性风险的预测价值 |
| 郭芳琪1,2,3,刘晟4,5,徐磊6,李勇刚1*,赵佳琦3* |
|
|
(1. 苏州大学附属第一人民医院放射科, 苏州 215000;
2. 海军军医大学(第二军医大学)第二附属医院超声诊疗科, 上海 200003;
3. 同济大学附属上海市第四人民医院超声医学科, 上海 200434;
4. 海军军医大学(第二军医大学)第二附属医院甲乳疝外科, 上海 200003;
5. 同济大学附属上海市第四人民医院甲乳血管外科, 上海 200434;
6. 浙江求是数理医学研究院, 杭州 315032
*通信作者) |
|
| 摘要: |
| 目的 探讨基于超声人工智能（AI）系统AI-SONICTM Thyroid 5.3.3.0的图像分析在甲状腺结节恶性风险评估中的应用价值。方法 选取2019年4月至2021年1月海军军医大学（第二军医大学）第二附属医院收治的453例甲状腺结节患者，共573枚甲状腺结节。以术后病理结果为金标准，通过χ2检验和ROC曲线评估术前AI系统检查对不同性别分组、不同年龄分组及不同结节大小分组的甲状腺结节良恶性的鉴别诊断效能，并通过DeLong检验比较术前AI系统检查与不同年资超声医师术前应用常规超声检查鉴别诊断甲状腺结节良恶性的效能。结果 在术前检查的573枚甲状腺结节中，术后病理证实为恶性411枚（76.5%）、良性162枚（23.5%）。低年资超声医师应用常规超声检查鉴别诊断甲状腺结节良恶性的灵敏度、特异度、准确度分别为85.2%（350/411）、55.6%（90/162）、76.8%（440/573），AUC为0.721（95%CI 0.672～0.771）；高年资超声医师鉴别诊断甲状腺结节良恶性的灵敏度、特异度、准确度分别为93.9%（386/411）、74.1%（120/162）、88.3%（506/573），AUC为0.865（95%CI 0.825～0.904）；AI系统鉴别诊断甲状腺结节良恶性的灵敏度、特异度、准确度分别为92.5%（380/411）、69.1%（112/162）、85.9%（492/573），AUC为0.809（95%CI 0.764～0.854）。DeLong检验结果显示，AI系统鉴别诊断甲状腺结节良恶性的AUC高于低年资超声医师（P=0.032），与高年资超声医师之间差异无统计学意义（P>0.05）。按不同性别、不同年龄分组，AI系统鉴别诊断甲状腺结节良恶性的准确度差异无统计学意义（P>0.05）；按不同结节大小分组，结节最大直径为10～<15 mm时AI系统鉴别诊断甲状腺结节良恶性的AUC最大，为0.882（95%CI 0.723～0.916）。结论 AI-SONICTM Thyroid 5.3.3.0可识别甲状腺结节的良性和恶性声像特征，其诊断效能接近高年资超声医师，有望成为术前预测甲状腺结节恶性风险的实用工具。 |
| 关键词: 甲状腺结节 超声检查 人工智能 计算机辅助诊断 |
| DOI：10.16781/j.CN31-2187/R.20230027 |
| 投稿时间：2023-02-03修订日期：2023-09-26 |
| 基金项目:海军军医大学（第二军医大学）第二附属医院人才建设三年行动计划——“金字塔人才工程”军事医学人才项目（1009），同济大学附属上海市第四人民医院科研启动专项（SYKYQD06101），上海市虹口区卫生健康委员会医学科研课题（虹卫2302-26），上海市虹口区卫生健康委员会临床重点扶持专科项目（HKLCFC202404）. |
|
| Ultrasound image analysis based on AI-SONICTM Thyroid 5.3.3.0 to predict malignant risk of thyroid nodules |
| GUO Fangqi1,2,3,LIU Sheng4,5,XU Lei6,LI Yonggang1*,ZHAO Jiaqi3* |
(1. Department of Radiology, The First Affiliated Hospital of Soochow University, Suzhou 215000, Jiangsu, China;
2. Department of Ultrasound, The Second Affiliated Hospital of Naval Medical University (Second Military Medical University), Shanghai 200003, China;
3. Department of Ultrasound, Shanghai Fourth People's Hospital Affiliated to Tongji University, Shanghai 200434, China;
4. Department of Thyroid, Breast and Hernia Surgery, The Second Affiliated Hospital of Naval Medical University (Second Military Medical University), Shanghai 200003, China;
5. Department of Thyroid, Breast and Vascular Surgery, Shanghai Fourth People's Hospital Affiliated to Tongji University, Shanghai 200434, China;
6. Zhejiang Qiushi Institute for Mathematical Medicine, Hangzhou 315032, Zhejiang, China
*Corresponding authors) |
| Abstract: |
| Objective To investigate the application value of artificial intelligence (AI) system AI-SONICTM Thyroid 5.3.3.0 based on ultrasound image analysis in the risk assessment of malignant thyroid nodules. Methods A total of 573 thyroid nodules were selected from 453 patients who were admitted to The Second Affiliated Hospital of Naval Medical University (Second Military Medical University) from Apr. 2019 to Jan. 2021. With the postoperative pathology as the gold standard, the differential diagnostic efficacies of preoperative AI system examination for benign and malignant thyroid nodules in different gender groups, different age groups, and different nodule size groups were evaluated by χ2 test and receiver operating characteristic (ROC) curve. The efficacies of preoperative AI system examination and conventional ultrasonography by different seniority ultrasound physicians before operation in diagnosing benign and malignant thyroid nodules were compared by DeLong test. Results Of the 573 thyroid nodules examined before operation, 411 (76.5%) were malignant and 162 (23.5%) were benign as confirmed by pathology after operation. The sensitivity, specificity, and accuracy of conventional ultrasonography in diagnosing benign and malignant thyroid nodules were 85.2% (350/411), 55.6% (90/162), and 76.8% (440/573), respectively, with an area under curve (AUC) of 0.721 (95% confidence interval [CI] 0.672-0.771) in the junior ultrasound physicians; the sensitivity, specificity, and accuracy were 93.9% (386/411), 74.1% (120/162), and 88.3% (506/573), respectively, with an AUC of 0.865 (95% CI 0.825-0.904) in the senior ultrasound physicians. The sensitivity, specificity, and accuracy of the AI system were 92.5% (380/411), 69.1% (112/162), and 85.9% (492/573), respectively, with an AUC of 0.809 (95% CI 0.764-0.854). DeLong test results showed that the AUC of the AI system in diagnosing benign and malignant thyroid nodules was significantly higher than that of the junior ultrasound physicians (P=0.032), and there was no significant difference between the AI system and senior ultrasound physicians (P>0.05). There was no significant difference in the accuracy of the AI system in diagnosing benign or malignant thyroid nodules among patients with different genders or different ages (P>0.05). For nodules of different sizes, when the maximum diameter of nodules was 10-<15 mm, the AUC of the AI system was the highest, being 0.882 (95% CI 0.723-0.916). Conclusion AI-SONICTM Thyroid 5.3.3.0 can identify benign and malignant features of thyroid nodules, and its diagnostic efficiency is close to that of senior ultrasound physicians. It is expected to be a practical tool to predict the risk of malignant thyroid nodules before clinical operation. |
| Key words: thyroid nodule ultrasonography artificial intelligence computer aided diagnosis |
