AI辅助无线手持超声筛查婴儿发育性髋关节发育不良
本研究开发了一套人工智能无线手持超声髋关节诊断系统(AI-HUDs),基于YOLOv8模型自动识别6个关键解剖标志并按Graf法测量α角和β角。研究纳入1192张超声图像和498个动态视频用于训练测试,并由住院医师操作AI-HUDs采集52名婴儿的104个髋关节超声视频进行验证。结果显示AI-HUDs自动测角与专家手动测量一致性良好,住院医师操作该系统可获得与专家相当的诊断性能。
要点
- AI-HUDs在静态模式下α角MAE为1.05°、β角MAE为1.65°,动态模式误差略有增大。
- 住院医师操作AI-HUDs与专家手动测量的α角ICC为0.69、β角ICC为0.61,Bland-Altman分析95%以上数据点在一致性界限内。
- 该系统有望在资源匮乏地区作为便捷的DDH筛查工具推广。
核心结论
- AI-HUDs在自动测量α角和β角方面表现出较高的可靠性和准确性。
- 住院医师使用AI-HUDs获得的诊断性能与超声专家相当。
- AI-HUDs有望成为资源匮乏地区便捷的DDH筛查工具,促进婴儿髋关节超声筛查的广泛应用。
研究设计与样本
- 研究类型为人工智能诊断系统的开发与验证研究,纳入1192张超声图像和498个动态视频用于模型训练与测试。
- 采用YOLOv8模型构建AI-HUDs,自动识别静态和动态髋关节超声图像中的6个关键解剖标志,依据Graf法测量α角和β角。
- 验证部分由一名住院医师使用AI-HUDs采集52名婴儿的104个髋关节超声视频,与超声专家对相应静态图像的手动测量进行一致性评估。
关键影像表现/征象
- 系统自动识别静态和动态髋关节超声图像中的6个关键解剖标志,并依据Graf法完成α角和β角测量。
- 与专家手动测量相比,静态模式下α角差异标准差为1.38°、β角差异标准差为2.14°,平均绝对误差分别为1.05°和1.65°。
- 动态模式下α角MAE为1.18°、β角MAE为1.86°,较静态模式误差略有增大。
- 住院医师操作AI-HUDs时,α角MAE为1.16°、β角MAE为1.91°,与专家手动测量保持良好一致性。
诊断标准与关键数值
- 静态模式下AI-HUDs与专家手动测量的组内相关系数:α角ICC为0.82,β角ICC为0.64。
- 动态模式下AI-HUDs与专家手动测量的组内相关系数:α角ICC降至0.77,β角ICC降至0.51。
- 住院医师操作AI-HUDs时与专家手动测量的ICC:α角为0.69,β角为0.61。
- Bland-Altman分析显示96.15%的α角数据点和95.19%的β角数据点位于一致性界限内。
临床应用与阅片要点
- AI-HUDs可用于自动测量Graf法α角和β角,辅助DDH筛查中髋关节分型判断。
- 静态图像测量一致性优于动态视频测量,临床应用时宜优先采用静态图像进行角度评估。
- 住院医师等非超声专科人员操作该系统可获得接近专家的诊断表现,适合基层和远程场景部署。
局限性
- β角的一致性指标(ICC为0.51至0.64)明显低于α角,提示β角自动测量的可靠性有限。
- 动态模式下α角和β角的ICC较静态模式下降,动态视频分析的一致性有待提升。
- 验证样本量较小,仅52名婴儿、104个髋关节视频用于住院医师操作评估。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVES: Ultrasonography is increasingly the preferred method for infant hip screening to enable timely diagnosis and treatment of developmental dysplasia of the hip (DDH). However, its reliance on experienced specialists and bulky equipment limits its application in routine screening, particularly in resource-limited and remote settings. We aimed to develop an artificial intelligence wireless handheld ultrasound hip diagnostic system (AI-HUDs) for accurate and robust DDH diagnosis.
METHODS: An AI-HUDs was developed to automatically identify and analyze 6 key anatomical landmarks in static and dynamic hip ultrasound images, employing the Graf method to measure α and β angles. The dataset comprised 1192 ultrasound images and 498 dynamic videos. The YOLOv8 model was trained and tested to construct AI-HUDs. Evaluation focused on the consistency between AI-HUDs' automatic angle measurements and manual measurements by experts using traditional ultrasound equipment. Additionally, 104 hip ultrasound videos from 52 infants were acquired by a resident physician using AI-HUDs, while corresponding static images were manually measured by ultrasound experts to assess agreement between novice-operated AI-HUDs and expert manual assessment.
RESULTS: Compared with expert manual measurements, the AI-HUDs demonstrated good performance in static mode: the differences in α and β angles exhibited standard deviations of 1.38° and 2.14°, and mean absolute errors (MAEs) of 1.05° and 1.65°, respectively. The intraclass correlation coefficients (ICCs) were 0.82 (α) and 0.64 (β). In dynamic mode, the MAEs increased slightly to 1.18° (α) and 1.86° (β), while the ICCs decreased to 0.77 for α and 0.51 for β. When operated by the resident physician, AI-HUDs maintained good performance with expert manual measurements: MAEs were 1.16° (α) and 1.91° (β); ICCs were 0.69 (α) and 0.61 (β). Bland-Altman analysis showed 96.15% (α) and 95.19% (β) of data points within the limits of agreement.
CONCLUSION: AI-HUDs demonstrated high reliability and accuracy in automatically measuring α and β angles. Resident physicians using this system achieved diagnostic performance comparable to experts. AI-HUDs show promise as a convenient DDH screening tool in resource-scarce regions, potentially facilitating wider adoption of infant hip ultrasound screening.
引用
- Zhang, Tao, Chen, et al. Artificial Intelligence-Assisted Wireless Handheld Ultrasound for Screening Developmental Dysplasia of the Hip in Infants. J Ultrasound Med. 2026. PMID: 42319217. doi:10.1002/jum.70318
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。