Deep learning models semi-automatic training system for quality control of transthoracic echocardiography  

深度学习模型半自动训练系统用于经胸超声心动图质量控制

作　　者：QIAN Sunnan WENG Hexiang CHENG Hanlin SHI Zhongqing WANG Xiaoxian GUO Guanjun FANG Aijuan LUO Shouhua YAO Jing QI Zhanru 钱隼南;翁和祥;成汉林;史中青;王小贤;郭冠军;方爱娟;罗守华;姚静;戚占如(江苏省省级机关医院信息处,江苏南京210009;东南大学生物科学与医学工程学院智能医学系,江苏南京210096;南京大学医学院附属鼓楼医院超声医学科,江苏南京210008;南京大学医学院附属鼓楼医院医学影像中心,江苏南京210008;南京鼓楼医院集团仪征医院超声医学科,江苏扬州211400)

机构地区：[1]Department of Information Office,Jiangsu Province Official Hospital,Nanjing 210009,China [2]Department of Intelligent Medicine,School of Biological Sciences and Medical Engineering,Southeast University,Nanjing 210096,China [3]Department of Ultrasound Medicine,Nanjing Drum Tower Hospital,Affiliated Hospital of Medical School,Nanjing University,Nanjing 210008,China [4]Department of Medical Imaging Center,Nanjing Drum Tower Hospital,Affiliated Hospital of Medical School,Nanjing University,Nanjing 210008,China [5]Department of Ultrasound Medicine,Yizheng Hospital of Nanjing Drum Tower Hospital Group,Yangzhou 211400,China

出　　处：《中国医学影像技术》2024年第8期1140-1145,共6页Chinese Journal of Medical Imaging Technology

基　　金：江苏省重点研发计划(BE2022828);江苏省前沿引领技术基础研究重大项目(BK20222002);江苏省卫生健康委2022年度医学科研项目(281);南京鼓楼医院临床研究专项(2022-YXZX-YX-01)。

摘　　要：Objective To explore the value of deep learning(DL)models semi-automatic training system for automatic optimization of clinical image quality control of transthoracic echocardiography(TTE).Methods Totally 1250 TTE videos from 402 patients were retrospectively collected,including 490 apical four chamber(A4C),310 parasternal long axis view of left ventricle(PLAX)and 450 parasternal short axis view of great vessel(PSAX GV).The videos were divided into development set(245 A4C,155 PLAX,225 PSAX GV),semi-automated training set(98 A4C,62 PLAX,90 PSAX GV)and test set(147 A4C,93 PLAX,135 PSAX GV)at the ratio of 5∶2∶3.Based on development set and semi-automatic training set,DL model of quality control was semi-automatically iteratively optimized,and a semi-automatic training system was constructed,then the efficacy of DL models for recognizing TTE views and assessing imaging quality of TTE were verified in test set.Results After optimization,the overall accuracy,precision,recall,and F1 score of DL models for recognizing TTE views in test set improved from 97.33%,97.26%,97.26%and 97.26%to 99.73%,99.65%,99.77%and 99.71%,respectively,while the overall accuracy for assessing A4C,PLAX and PSAX GV TTE as standard views in test set improved from 89.12%,83.87%and 90.37%to 93.20%,90.32%and 93.33%,respectively.Conclusion The developed DL models semi-automatic training system could improve the efficiency of clinical imaging quality control of TTE and increase iteration speed.目的观察深度学习(DL)模型半自动训练系统用于自动优化临床经胸超声心动图(TTE)图像质量控制的价值。方法回顾性收集402例接受TTE检查患者的1250段TTE视频,包括490段心尖四腔(A4C)、310段胸骨旁左心室长轴(PLAX)及450段胸骨旁短轴大血管水平(PSAXGV)切面;按5∶2∶3比例将其分为开发集(含245段A4C、155段PLAX及225段PSAXGV)、半自动训练集(含98段A4C、62段PLAX及90段PSAXGV)及测试集(含147段A4C、93段PLAX及135段PSAXGV)。基于开发集和半自动训练集构建DL模型半自动训练系统,于测试集验证其识别TTE切面及评估TTE质量的效能。结果优化后DL模型识别测试集各切面TTE的整体准确率、精确率、召回率及F1分数分别由97.33%、97.26%、97.26%及97.26%提至99.73%、99.65%、99.77%及99.71%,其判定测试集A4C、PLAX、PSAXGVTTE为标准切面的整体准确率分别由89.12%、83.87%及90.37%提高至93.20%、90.32%及93.33%。结论所获DL模型半自动训练系统可提升临床TTE质量控制性能并加快迭代速度。

关 键 词：ECHOCARDIOGRAPHY quality control artificial intelligence 

分 类 号：R322.1[医药卫生—人体解剖和组织胚胎学] R540.45[医药卫生—基础医学]