Noise2Noise自监督人工智能在乳腺超声存图质控的应用研究  

Study on the Application of Self-Supervised Artificial Intelligence Based on Noise2Noise in the Quality Control of Breast Ultrasonography

作　　者：蒋南 张元恒 金壮 温馨 张祎 滕月阳[2] 曹军英 Jiang Nan;Zhang Yuanheng;Jin Zhuang;Wen Xin;Zhang Yi;Teng Yueyang;Cao Junying(Department of Ultrasound Diagnosis,The General Hospital of Northern Theater Command,Shenyang 110016,China;College of Medicine and Biological Information Engineering,Northeastern University,Shenyang 11o016,China)

机构地区：[1]中国人民解放军北部战区总医院超声诊断科,沈阳市110016 [2]东北大学医学与生物信息工程学院,沈阳市110016

出　　处：《中国超声医学杂志》2023年第12期1345-1348,共4页Chinese Journal of Ultrasound in Medicine

基　　金：辽宁省重点研发项目(No.2020JH2/103300122);沈阳市科学技术计划(No.20-205-4-058)。

摘　　要：目的 利用Noise2Noise自监督人工智能(AI)提高乳腺超声存图质控的准确度。方法 对手动提取“体标”、“大小”和“血流”3种标识创建大型数据集,利用自监督训练和Noise2Noise的耦合方式集进行人工智能模型训练。分割准确度和重建相似性两类定量指标对模型的性能进行分析,筛选出最优AI模型。随机选取行乳腺检查的1 000例患者图像,分别记录AI和人工质控完成3种标识所用的时间和准确度,进行AI与人工质控比较。结果 Noise2Noise模式下训练的U-Net模型可作为后续研究使用的最优AI模型;人工在完成质控标识的时间差异无统计学意义(P>0.05);“体标”、“大小”和“血流”3种标识的AI质控准确度为95.44%、78.50%和73.50%,人工质控准确度为99.99%、100%、100%,AI质控3种标识的每张图像耗时较人工质控分别节省0.45 s、0.84 s和0.50 s。结论 Noise2Noise自监督AI进行超声图像标识质控较人工质控可缩短耗时,提高质控效率,但准确度有待进一步提高。Objective To improve the accuracy in quality control of breast ultrasonographic image using the selfsupervised artificial intelligence(AI)based on Noise2Noise.Methods A data set was created for manual extraction of body marker,radial line and vascular flow.The AI model was trained by the self-supervised training coupled with the Noise2Noise method.Two quantitative indicators,segmentation accuracy and reconstruction similarity were used to analyze the performance of the model,and select the optimal Al model.Images of 1 ooo patients who underwent breast examination at were randomly selected,and the time and accuracy of Al and manual quality control to complete the three marks were recorded respectively,and the results of quality control between Al and manual was compared.Results The U-Net model trained in the Noise2Noise was considered the optimal AI model for subsequent research.There was no significant difference in time spent manually completing quality control identification(P>0.05).The Al quality control accuracy of body mark,radial line and vascular flow was 95.44%,78.50%and 73.50%,while the manual quality control accuracy was 99.99%,100%and 100%.Each image of Al quality control of each image of the three identifiers saved 0.45 s,0.84 s and 0.50 s respectively,compared with manual control.Conclusions Compared with manual quality control,self-supervised AI based on Noise2Noise can shorten time and improve quality control efficiency,and the accuracy needs to be further improved.

关 键 词：人工智能 超声存图 质控 乳腺 

分 类 号：R445.1[医药卫生—影像医学与核医学]