超声人工智能辅助诊断发育性髋关节发育不良

Study on artificial intelligence-based ultrasonic-assisted diagnosis for developmental dysplasia of the hip

目的构建一种发育性髋关节发育不良(developmental dysplasia of the hip,DDH)超声图像的人工智能(artificial intelligence,AI)辅助诊断的深度学习系统并验证。方法回顾性收集2019年1月至2021年1月的2000段儿童髋关节超声影像片段,并选取影像片段中标准截面2000张,所有标准截面均由标注小组通过使用基于Python 3.6环境的自编软件使用图片跨媒体数据标注与人工审核标准化流程用统一的标准进行处理。随机选取其中1753张用于训练深度学习系统,余247张用于测试系统。再从测试集中随机选取200张标准截面,并由8位临床医生独立完成读片标注,将8份独立结果与AI结果进行比较。结果测试集共247例,与临床医生测量相比,判读髋关节是否成熟的诊断受试者工作特征曲线(receiver operating characteristic curve,ROC)下的面积(area under the ROC,AUC)为0.865,灵敏度为76.19%,特异性为96.90%。Graf详细分型下的深度学习系统判读的AUC为0.575,灵敏度为25.90%,特异性为89.10%。根据Bland-Altman方法确定的α角[95%LoA(-4.7051°,6.5948°),Bias-0.94,P<0.001]和β角[95%LoA(-7.7191,6.8777),Bias-0.42,P=0.077],与8位临床医生相比AI系统判读结果均更为稳定,且β角效果更为突出。结论新型人工智能系统可快速且准确的测量标准髋关节超声平面的Graf法相关指数。

To construct an artificial intelligence(AI)deep learning system for rapid and accurate ultrasonic-assisted diagnosis developmental dysplasia of the hip(DDH),and then to verify and compare it.Methods Two thousand standard sections images werre collected from 2000 clinical retrospective pediatric hip ultrasound videos from January 2019 to January 2021.All standard sections were annotated by the annotation team through the self-designed software based on Python 3.6 environment for image cross-media data annotation and manual review standardization process with unified standards.Among them,1753 were randomly selected for training the deep learning system,and the remaining 247 were used for testing the system.Further,200 standard sections were randomly selected from the test set,and 8 clinicians independently completed the film reading annotation.The 8 independent results were then compared with the AI results.Results The testing set consists of 247 patients.Compared with the clinician's measurements,the area under the receiver operating characteristic curve(AUC)of diagnosing hip joint maturity was 0.865,the sensitivity was 76.19%,and the specificity was 96.9%.The AUC of AI system interpretation under Graf detailed typing was 0.575,the sensitivity was 25.90%,the specificity was 89.10%.The 95%LoA ofα-angle determined by Bland-Altman method,of-4.7051°to 6.5948°(Bias-0.94,P<0.001),compared with clinicians'measurements.The 95%LoA ofβ-angle,of-7.7191 to 6.8777(Bias-0.42,P=0.077).Compared with those from 8 clinicians,the results of AI system interpretation were more stable,and theβ-angle effect was more prominent.Conclusion The AI system can quickly and accurately measure the Graf correlation index of standard DDH ultrasonic standard diagnosis plane.