摘要
目的采用计算机深度神经网络技术,构建一款人工智能模型辅助^(99)Tc^(m)O_(4)^(−)甲状腺图像诊断。资料与方法回顾性纳入四川大学华西医院核医学科2013年1月—2020年6月临床已完成甲状腺全切手术、拟进行131I治疗的甲状腺癌患者3515例图像集,按甲状腺残留程度分类标注后,以8∶2随机分成训练集2811例和测试集704例。利用3种深度神经网络模型Resnet34、InceptionV3和Densenet161分别对训练集样本进行特征提取和训练后,对测试集样本进行效能验证,并与3名初级医师独立阅片结果进行对比,记录3名医师和模型的阅片时间,采用Kappa检验分析医师阅片的一致性,采用受试者工作特征曲线分析不同模型的诊断效能。结果在704例甲状腺图像分类测试时,3名医师的判断准确度分别为89.5%、86.5%、86.6%;Resnet34、InceptionV3和Densenet161神经网络模型的判断准确度分别为91.3%、90.4%和91.2%。3名医师两两比较诊断一致性较好(Kappa=0.773、0.746、0.711,P均<0.05),3名医师判断所需总时间分别为170 min、172 min和131 min;Resnet34、InceptionV3和Densenet161神经网络模型诊断总时间分别为4.5 s、2.9 s和17.3 s。结论人工智能辅助诊断技术可快速、准确地完成甲状腺显像的阅片与甲状腺残留分类工作。
Purpose To construct an artificial intelligence model based on a deep neural network to automatically classify ^(99)Tc^(m)O_(4)^(−) thyroid imaging.Materials and Methods This study retrospectively included 3515 thyroid cancer patients who had completed clinical total thyroidectomy and planned to receive 131I treatment in West China Hospital of Sichuan University from January 2013 to June 2020.All 3515 cases of ^(99)Tc^(m)O_(4)^(−) thyroid images were labeled by two senior physicians and randomly divided into training set(n=2811 cases)and testing set(n=704 cases).Three types of deep neural network models(Resnet34,InceptionV3 and Densenet161)were performed on the training set to obtain features,and efficacy performances were conducted on the testing set to identify residual thyroid.Accuracy and time efficiency were assessed and further compared with three junior physicians.The total reading time of imaging by three doctors and models were recorded,the consistency of reading imaging by doctors was analyzed via Kappa test,the diagnostic efficacy of different models was analyzed by receiver operator characteristic curve.Results Three junior physicians classified 704 images with accuracy of 89.5%,86.5%and 86.6%,respectively.The accuracy of neural network models of Resnet34,InceptionV3 and Densenet161 were 91.3%,90.4%and 91.2%,respectively.There were good diagnostic consistency between physicians 1 and 2,physicians 1 and 3,and physicians 2 and 3(Kappa=0.773,0.746,0.711;all P<0.05).The total time of three physicians was 170 minutes,172 minutes and 131 minutes,respectively,and the total diagnostic time of Resnet34,InceptionV3 and Densenet161 of neural network models was 4.5 seconds,2.9 seconds and 17.3 seconds,respectively.Conclusion Artificial intelligence-assisted technology is valuable to quickly and accurately classify residual thyroid based on ^(99)Tc^(m)O_(4)^(−) thyroid images.
作者
向镛兆
黄秋菊
魏建安
皮勇
蔡华伟
蒋丽莎
杨沛
李玉豪
青春
赵祯
XIANG Yongzhao;HUANG Qiuju;WEI Jianan;PI Yong;CAI Huawei;JIANG Lisha;YANG Pei;LI Yuhao;QING Chun;ZHAO Zhen(Department of Nuclear Medicine,West China Hospital,Sichuan University,Chengdu 610041,China;不详)
出处
《中国医学影像学杂志》
CSCD
北大核心
2023年第2期108-113,共6页
Chinese Journal of Medical Imaging
基金
四川省“十四五”生命健康重大科研专项先进前沿技术在重大疾病和罕见病中的应用研究(2022ZDZX0023)
四川大学华西医院学科卓越发展1·3·5工程临床研究孵化项目(2021HXFH033)。
关键词
甲状腺肿瘤
神经网络
人工智能
诊断
Thyroid neoplasms
Neural networks
Artificial intelligence
Diagnosis
