文献速递儿科影像骨肌技术与安全DR

基于深度学习的儿童胸片骨密度预测:多中心可行性研究

Deep Learning-based Bone Mineral Density Prediction Using Pediatric Chest Radiographs: A Multicenter Feasibility Study.
PubMed 文献 · Radiology · 2026-01-01

本研究回顾性纳入两家三级医院1464对胸片-DXA配对数据,开发结合胸片与临床变量的人工智能模型预测儿童腰椎面积骨密度。模型预测的Z值与DXA强相关(内部r=0.85,外部r=0.76),检测低骨密度(Z≤-2.0)的AUC内部为0.92、外部为0.90。

要点

  • 胸片+年龄性别身高体重可预测儿童腰椎BMD Z值,与DXA强相关。
  • 检测低骨密度Z≤-2.0的AUC达0.90-0.92,但内部敏感度仅60%。
  • 外部验证敏感度82%、特异度85%,阈值下漏诊率仍不可忽视。

核心结论

  • 基于胸片的人工智能模型能准确预测儿童BMD Z值并识别低骨密度。

研究设计与样本

  • 回顾性研究,纳入18岁以下、3个月内完成DXA与胸片检查的患者,来自两家三级医院(内部测试2014-2023,外部测试2022-2023)。
  • 共1464对胸片-DXA配对(中位年龄13岁,IQR 11-16岁;824例男孩),开发集774、内部测试集376、外部测试集314。
  • 内部数据按时间拆分,开发集采用五折交叉验证;模型结合胸片与临床变量(年龄、性别、身高、体重)预测腰椎L1-L4面积BMD,Z值基于韩国儿童参考计算。
  • 性能评估:回归用Pearson相关系数r,检测低骨密度(Z≤-2.0)用ROC曲线下面积AUC。

诊断标准与关键数值

  • 预测BMD Z值与DXA强相关:内部测试集r=0.85(95%CI 0.82-0.88,P<.001),外部测试集r=0.76(95%CI 0.71-0.81,P<.001)。
  • 检测低骨密度(Z≤-2.0):内部测试集AUC=0.92(95%CI 0.89-0.95),敏感度60%(50/84,95%CI 48-70),特异度95%(276/292,95%CI 91-97)。
  • 外部测试集AUC=0.90(95%CI 0.87-0.94),敏感度82%(54/66,95%CI 70-90),特异度85%(210/248,95%CI 80-89)。

临床应用与阅片要点

  • 可将胸片作为机会性筛查工具,提示需进一步DXA确认的低骨密度儿童。
  • 注意内部测试敏感度仅60%,阴性结果不能排除低骨密度,高风险患儿仍应转诊DXA。
  • 模型依赖年龄、性别、身高、体重等临床变量,报告时需确保输入信息完整准确。

局限性

  • 回顾性设计,仅两家中心,外部验证样本量相对有限。
  • Z值参考基于韩国儿童人群,结论外推至其他人群需谨慎。
  • 内部测试敏感度较低(60%),作为筛查工具存在漏诊可能。

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

Background Measuring bone mineral density (BMD) is essential for pediatric bone health assessment. Dual-energy x-ray absorptiometry (DXA) is the reference standard but has limited accessibility. Purpose To develop and evaluate an artificial intelligence model for predicting BMD from pediatric chest radiography. Materials and Methods This retrospective study included patients aged younger than 18 years who underwent DXA and chest radiography within 3 months at two tertiary hospitals (internal test, 2014-2023; external test, 2022-2023). The internal dataset was temporally split into development (fivefold cross-validation) and test sets. The model combined chest radiographs and clinical variables (age, sex, height, and weight) to predict the lumbar spine (L1 through L4) areal BMD, with Z scores calculated from Korean pediatric reference. Performance was evaluated using the Pearson correlation coefficient (r) for regression and the area under the receiver operating characteristic curve (AUC) for detecting low BMD (Z score ≤ -2.0). Results A total of 1464 radiograph-DXA pairs (median age, 13 years [IQR, 11-16 years]; 824 boys) were included: 774 in the development set, 376 in the internal test set, and 314 in the external test set. The predicted BMD Z scores were strongly correlated with the DXA scores in both the internal (r = 0.85 [95% CI: 0.82, 0.88]; P < .001) and external (r = 0.76 [95% CI: 0.71, 0.81]; P < .001) test sets. For detecting low BMD, the internal test set had an AUC of 0.92 (95% CI: 0.89, 0.95), a sensitivity of 60% (50 of 84 scans; 95% CI: 48, 70), and a specificity of 95% (276 of 292 scans; 95% CI: 91, 97). The external test set achieved an AUC of 0.90 (95% CI: 0.87, 0.94), a sensitivity of 82% (54 of 66 scans; 95% CI: 70, 90), and a specificity of 85% (210 of 248 scans; 95% CI: 80, 89). Conclusion A chest radiograph-based artificial intelligence model accurately predicted pediatric BMD Z scores and identified low BMD. © RSNA, 2026 Supplemental material is available for this article.

来源:Radiology,查看原文。版权归原作者所有。

引用

  1. Choi JW, Ryu YJ, Cheon JE, et al. Deep Learning-based Bone Mineral Density Prediction Using Pediatric Chest Radiographs: A Multicenter Feasibility Study. Radiology. 2026. PMID: 41944722. doi:10.1148/radiol.252761

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。