文献速递骨肌儿科影像技术与安全US

人工智能辅助超声评估发育性髋关节发育不良:涵盖分割、角度测量与诊断任务的系统综述与Meta分析

Artificial Intelligence-assisted Ultrasound for Developmental Dysplasia of the Hip: Systematic Review and Meta-analysis Across Segmentation, Angle Measurement, and Diagnostic Tasks.
PubMed 文献 · Acad Radiol · 2025-11-15

该系统综述与Meta分析纳入34项研究,评估AI辅助婴儿髋关节超声在分割、角度测量和诊断分类中的表现。汇总DSC为0.876,α角ICC为0.883、β角ICC为0.864,诊断分类AUC为0.937、敏感度0.888、特异度0.893,但各领域异质性I²>75%,外部验证仅20.6%。

要点

  • AI辅助髋超声诊断分类汇总AUC 0.937,敏感度0.888、特异度0.893,支持潜在临床应用。
  • 分割汇总DSC 0.876,α角测量ICC 0.883、β角ICC 0.864,角度测量可靠性较高。
  • 各领域异质性I²>75%,外部验证仅20.6%,临床部署前需标准化。

核心结论

  • AI辅助髋超声在分割、角度测量和诊断分类中表现良好,支持潜在临床应用,但高异质性、外部验证有限和报告缺陷要求在临床部署前进行标准化。

研究设计与样本

  • 系统综述与Meta分析,PROSPERO注册号CRD420251133940,检索Medline/PubMed、Web of Science、Scopus和Cochrane Library,纳入34项报告AI评估DDH超声性能的研究,质量评估采用QUADAS-2和CLAIM清单,随机效应模型汇总DSC、ICC和诊断准确性指标,异质性用I²统计量和预测区间评估。

关键影像表现/征象

  • 分割任务7项研究(n=3011)汇总DSC为0.876(95%CI 0.827-0.924)。
  • 角度测量可靠性4项研究(n=2012)α角汇总ICC为0.883(95%CI 0.763-0.945),β角为0.864(95%CI 0.653-0.951)。
  • 诊断分类5项研究(n=3035)汇总AUC为0.937(95%CI 0.885-0.966),10项研究(n=5341)汇总敏感度0.888(95%CI 0.840-0.923)、特异度0.893(95%CI 0.735-0.961)。

诊断标准与关键数值

  • 分割性能以Dice相似系数(DSC)汇总,7项研究(n=3011)汇总DSC为0.876(95%CI 0.827-0.924)。
  • 角度测量可靠性以组内相关系数(ICC)汇总,4项研究(n=2012)α角ICC为0.883(95%CI 0.763-0.945)、β角ICC为0.864(95%CI 0.653-0.951)。
  • 诊断分类以AUC、敏感度和特异度汇总,5项研究(n=3035)AUC为0.937(95%CI 0.885-0.966),10项研究(n=5341)敏感度0.888(95%CI 0.840-0.923)、特异度0.893(95%CI 0.735-0.961)。

临床应用与阅片要点

  • AI辅助髋超声在分割、角度测量和诊断分类中表现良好,支持潜在临床应用。
  • 各领域异质性I²>75%,外部验证仅20.6%,临床部署前需标准化。

局限性

  • 各领域均观察到显著异质性(I²>75%)。
  • 外部验证有限,CLAIM依从性平均67.9%,外部验证关键缺口为20.6%。
  • 报告缺陷需在临床部署前通过标准化解决。

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

RATIONALE AND OBJECTIVES: Artificial intelligence (AI) applications for developmental dysplasia of the hip (DDH) assessment have emerged across multiple analytical tasks, yet comprehensive evidence synthesis remains limited. This systematic review and meta-analysis evaluated AI-assisted hip ultrasonography performance across segmentation, angle measurement, and diagnostic classification tasks in infants.

MATERIALS AND METHODS: Following PROSPERO registration (CRD420251133940), we searched Medline/PubMed, Web of Science, Scopus, and Cochrane Library databases. Studies reporting AI performance for DDH ultrasonographic evaluation were included. Quality assessment employed QUADAS-2 and CLAIM checklists. Random-effects meta-analyses synthesized Dice Similarity Coefficients (DSC) for segmentation, Intraclass Correlation Coefficients (ICC) for angle measurements, and diagnostic accuracy metrics. Heterogeneity was assessed using I² statistics and prediction intervals.

RESULTS: Thirty-four studies met inclusion criteria. Segmentation performance (7 studies, n=3011) yielded pooled DSC of 0.876 (95% CI: 0.827-0.924). Angle measurement reliability (4 studies, n=2012) demonstrated pooled ICC of 0.883 (95% CI: 0.763-0.945) for alpha angles and 0.864 (95% CI: 0.653-0.951) for beta angles. Diagnostic classification showed pooled AUC of 0.937 (95% CI: 0.885-0.966) across 5 studies (n=3035), with pooled sensitivity 0.888 (95% CI: 0.840-0.923) and specificity 0.893 (95% CI: 0.735-0.961) from 10 studies (n=5341). Substantial heterogeneity (I²>75%) was observed across all domains. CLAIM adherence averaged 67.9%, with critical gaps in external validation (20.6%).

CONCLUSION: AI-assisted hip ultrasonography demonstrates strong performance across multiple tasks, supporting potential clinical application. However, high heterogeneity, limited external validation, and reporting deficiencies necessitate standardization before clinical deployment.

来源:Acad Radiol,查看原文。版权归原作者所有。

引用

  1. Sakar R, Onay A, Mercan ED, et al. Artificial Intelligence-assisted Ultrasound for Developmental Dysplasia of the Hip: Systematic Review and Meta-analysis Across Segmentation, Angle Measurement, and Diagnostic Tasks. Acad Radiol. 2025. PMID: 41242894. doi:10.1016/j.acra.2025.10.047

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。