摘要
目的构建人工智能(AI)自动识别髋关节超声标准冠状切面及测量相关参数模型,观察其辅助超声筛查婴儿发育性髋关节发育不良(DDH)的价值。方法回顾性分析2164名婴儿共4328侧髋关节超声视频,由5名超声科主治医师采用SonoKit标注软件以统一标准于每段视频中选取1幅标准、2幅非标准髋关节冠状切面声像图,并于标准冠状切面图中标注关键解剖结构。经2名超声科主任医师审核,共获得11100幅声像图(3665幅标准、7435幅非标准),以其中8100幅为训练集(2665幅标准、5435幅非标准)、3000幅为验证集(1000幅标准、2000幅非标准)。基于训练集数据构建AI模型,自动识别髋关节超声标准冠状切面,并于其中自动测量α角、β角和股骨头覆盖率(FHC);以验证集验证AI模型识别标准切面的效能。另选取110名健康婴儿的220幅髋关节标准冠状切面声像图,分别由超声科医师手动测量、以AI模型自动测量其α角、β角和FHC,分析测量结果的一致性及相关性。结果对于验证集髋关节超声标准冠状切面,AI模型与超声科主任医师识别结果的一致性较好(Cohen’s Kappa=0.925);AI模型自动测量与医师手动测量α角、β角及FHC的一致性均较好,组内相关系数分别为0.814、0.730和0.953,均具有强相关性(r=0.826、0.731、0.967)。结论AI模型能有效自动识别髋关节超声标准冠状切面并自动测量相关参数,可辅助超声筛查婴儿DDH。
Objective To construct an artificial intelligence(AI)model for automatically identifying standard coronal section of hip joint ultrasound and measuring related parameters,and to explore its value for assisting ultrasound screening infant developmental dysplasia of the hip(DDH).Methods Bilateral hip ultrasound video of 2164 infants(4328 hips)were retrospectively analyzed.SonoKit annotation software were used by 5 attending physicians to select 1 standard and 2 non-standard coronal ultrasonograms of hip joint for each video according to uniform standards,and the key anatomical structures on the standard coronal sections were marked.Then 11100 images(3665 standard and 7435 non-standard)were obtained and taken as the data set after reviewed by 2 chief physicians,among which 8100 were classified as training set(2665 standard and 5435 non-standard)and 3000 as validation set(1000 standard and 2000 non-standard).AI model was constructed based on the training set to automatically recognize ultrasonic standard coronal section of hip joint and measure the relevant parameters,i.e.α-angle,β-angle and femoral head coverage(FHC).Then the efficacy of AI model for recognizing standard ultrasonic section in the validation set was verified.Meanwhile,other 220 standard coronal section ultrasound images of hip joint of 110 healthy infants were selected,α-angle,β-angle and FHC were measured automatically by AI model and manually by ultrasound physicians,respectively,and the consistency and correlation of the measurement results were analyzed.Results AI model automatically identified results of standard coronal section of hip joint ultrasonograms had good agreement with those of chief ultrasound physicians(Cohen's Kappa=0.925)in validation set.The consistency of α-angle,β-angle and FHC were all good between AI model automatic measurement results and manual measurement results,the intra-group correlation coefficients was 0.814,0.730 and 0.953,respectively,and strong correlations were found(r=0.826,0.731,0.967).Conclusion AI model
作者
徐英
于红奎
林小影
赵杨
黄子殷
许晓
杨星怡
XU Ying;YU Hongkui;LIN Xiaoying;ZHAO Yang;HUANG Ziyin;XU Xiao;YANG Xingyi(Department of Ultrasound,Shenzhen Bao'an District Maternal and Child Health Hospital,Shenzhen 518133,China;Shenzhen Kaili Biomedical Technology Co.,Shenzhen 518107,China)
出处
《中国医学影像技术》
CSCD
北大核心
2023年第8期1229-1233,共5页
Chinese Journal of Medical Imaging Technology
基金
深圳市科技计划项目(JCYJ20210324134203010)。
关键词
髋关节
婴儿
人工智能
超声检查
发育性髋关节发育不良
hip joint
infant
artificial intelligence
ultrasonography
developmental dysplasia of the hip
