超声心动图图像人工智能质控系统构建研究  

Preliminary study on the construction of an echocardiogram image quality control system based on artificial intelligence

作　　者：戚占如 成汉林 单淳劼 陈瑞阳 翁和祥 杜悦 郭冠军[1] 王小贤 姚静 罗守华[2] 方爱娟[1] 陈慧[1] 史中青 Qi Zhanru;Cheng Hanlin;Shan Chunjie;Chen Ruiyang;Weng Heriang;Du Yue;Guo Guanjun;Wang Xiaorian;Yao Jing;Luo Shouhud;Fang Aijuan;Chen Hui;Shi Zhongqing(Department of Ultrasound Medicine,Nanjing Drum Tower Hospital,Affiliated Hospital of Medical School,Nanjing University,Medical Imaging Center,Nanjing Drum Tower Hospital,Affiliated Hospital of Medical School,Nanjing University,Nanjing 210008,China;School of Biological Sciences and Medical Engineering,Southeast University,Nanjing 210096,China;School of Biomedical Engineering and Informatics,Nanjing Medical University,Nanjing 211166,China)

机构地区：[1]南京大学医学院附属鼓楼医院超声医学科、南京大学医学院附属鼓楼医院医学影像中心,南京210008 [2]东南大学生物科学与医学工程学院,南京210096 [3]南京医科大学生物医学与工程学院,南京211166

出　　处：《中华超声影像学杂志》2025年第2期107-113,共7页Chinese Journal of Ultrasonography

基　　金：南京鼓楼医院临床研究专项资金(2022-YXZX-YX-01);江苏省卫生健康委2022年度医学科研立项项目(218);江苏省重点研发计划(BE20222828);江苏省前沿引领技术基础研究专项(BK20222002)。

摘　　要：目的探索基于人工智能进行超声心动图图像质量控制的可行性。方法回顾性随机抽取南京大学医学院附属鼓楼医院超声心动图数据库中2021-2023年的超声心动图二维视频图像5000个,其中心尖切面1559个。由医师团队制定评分细则,具体包括增益、缩放比例、心轴角度和结构显示四类评分标准,并对数据进行切面分类及图像质量评分的标记,标记后的数据将其分为训练集(n=643个)、验证集(n=276个)和测试集(n=640个)。训练集和验证集用于构建切面分类和质量评估模型,测试集用于验证模型效果。其中,切面分类模块通过SlowFast模型实现,质量评估模块涉及ResNet、Video Swin Transformer、SSD和U-Net等多个算法。结果分类模型在识别各心尖切面的平均准确度、精确度、召回率和F1得分分别为0.9871、0.9830、0.9871、0.9849,推理时间为(333.4±105.4)ms。质量评估模型包括增益、缩放比例、心轴角度、主要结构显示的平均准确度分别为0.9151、0.9282、0.9387、0.9656,整体评分准确度为0.9127。结论该研究搭建的超声心动图质量控制系统可以有效地对超声心动图心尖切面二维图像进行分类和质量评估,并保证了质控的客观性、及时性和高效性,对超声心动图质控体系的构建具有参考意义。Object:To explore the feasibility of using artificial intelligence for quality control of echocardiographic images.Methods:Retrospectively,5000 two-dimensional echocardiographic video images within the period from 2021 to 2023 were randomly retrieved from the echocardiography database of Nanjing Drum Tower Hospital,Affiliated Hospital of Medical School,Nanjing University.Among these selected images,1559 of them were apical views.The physician team formulated the scoring rules,which specifically included four scoring criteria:gain,scaling ratio,cardiac axis angle,and structure.Subsequently,the data were labeled with view classification and image quality scores.The labeled data were further partitioned into the training set(n=643),the validation set(n=276),and the test set(n=640).The training and validation sets were utilized for constructing the models for view classification and quality assessment,while the test set was employed to verify the models'effectiveness.The view classification module was implemented using the SlowFast model,and the quality assessment module involved algorithms such as ResNet,Video Swin Transformer,SSD,and U-Net.Results:The average accuracy,precision,recall rate and F1 score of the classification model in identifying each apical view were 0.9871,0.9830,0.9871 and 0.9849 respectively,and the inference time was(333.4±105.4)ms.The average accuracies of the quality assessment module in terms of gain,scaling ratio,cardiac axis angle and display of main structures were 0.9151,0.9282,0.9387 and 0.9656 respectively,and the overall scoring accuracy was 0.9127.Conclusions:The echocardiogram quality control system developed in this research can effectively classify and evaluate the quality of two-dimensional images of the apical views in echocardiograms.Moreover,it guarantees the objectivity,timeliness and high-efficiency of quality control,which has reference value for the establishment of the echocardiogram quality control system.

关 键 词：超声心动描记术 人工智能 图像质控 心尖切面 

分 类 号：R540.45[医药卫生—心血管疾病] TP18[医药卫生—内科学]