卷积神经网络诊断甲状腺结节的应用  

Application of Convolution Neural Network in Diagnosing Thyroid Nodules

作　　者：尹爱桃 陆永萍 赵易凡 孙月 张容亮 徐飞[1] YIN Aitao;LU Yongping;ZHAO Yifan;SUN Yue;ZHANG Rongliang;XU Fei(Department of Ultrasound,Affiliated Hospital of Yunnan University,Kunming 650000,China)

机构地区：[1]云南大学附属医院超声科,云南昆明650000

出　　处：《中国医学影像学杂志》2022年第12期1212-1217,1223,共7页Chinese Journal of Medical Imaging

基　　金：云南省“万人计划”名医专项。

摘　　要：目的 基于卷积神经网络的人工智能(AI)计算机辅助诊断(CAD)系统应用于超声诊断,评估其诊断甲状腺结节良恶性的效能。资料与方法 回顾性收集2018年4月—2021年2月云南大学附属医院经手术病理证实的甲状腺结节105例共157个结节,比较超声医师、AI-CAD系统诊断及联合诊断甲状腺结节的结果,采用受试者工作特征曲线下面积评估AI诊断甲状腺结节性质和结节特征分类的效能。结果 超声医师、基于卷积神经网络的AI-CAD系统及联合诊断良、恶性结节的敏感度分别为80.7%、84.5%、92.1%,特异度分别为73.7%、81.0%、86.3%,准确度分别为79.0%、84.1%、91.7%、阳性预测值分别为90.6%、92.4%、95.2%,阴性预测值分别为54.9%、66.7%、82.7%;AI-CAD的诊断效能高于超声医师,两者联合诊断效能最佳,差异有统计学意义(χ^(2)=5.524,P<0.05)。结论 超声医师、AI-CAD系统诊断及联合诊断对甲状腺结节均有较好的诊断价值,AI联合超声医师诊断甲状腺结节效能最好,对临床评估是否手术及手术方案有较高的应用价值。Purpose Artificial intelligence(AI) computer aided diagnosis system(CAD) based on convolutional neural network is applied to ultrasonic diagnosis, and to evaluate its performance in diagnosing benign and malignant thyroid nodules. Materials and Methods 105patients with thyroid nodules confirmed by surgery and pathology in the Affiliated Hospital of Yunnan University from April 2018 to February 2021 were retrospectively collected, with a total of 157 nodules. By comparing the results of ultrasonic diagnosis, AI-CAD system diagnosis and joint diagnosis of thyroid nodules, the performance of AI in diagnosing thyroid nodules and classifying thyroid nodules was evaluated by using area under the curve of ROC. Results The sensitivity of ultrasound doctors, convolutional neural network-based AI-CAD system and combined diagnosis of benign and malignant nodules was 80.7%, 84.5%, 92.1%, respectively;and the specificity was 73.7%, 81.0%, 86.3%,respectively. The accuracy was 79.0%, 84.1%, 91.7%, respectively;the positive predictive value was 90.6%, 92.4%, 95.2%, respectively;and the negative predictive value was 54.9%, 66.7%, 82.7%, respectively. The diagnostic efficiency of AI was higher than that of ultrasound doctors, and the combined diagnostic efficiency of AI and ultrasound doctors was the best, with statistically significant(χ^(2)=5.524, P<0.05).Conclusion Ultrasound doctors, AI-CAD system diagnosis and combined diagnosis have good diagnostic value for thyroid nodules. AI combined with ultrasound doctors has the best diagnostic efficiency for thyroid nodules, and has high application value for clinical evaluation of operation and surgical plan.

关 键 词：甲状腺结节 超声检查 AI-CAD系统 卷积神经网络 病理学 外科 诊断 鉴别 

分 类 号：R445.1[医药卫生—影像医学与核医学] R581.3[医药卫生—诊断学]