AI工具在四肢骨骼骨折X线检测中的独立骨水平诊断准确性研究
本研究在七家丹麦医院回顾性纳入2783例创伤X线患者，以临床放射报告为参考标准，评估商用AI骨折检测工具的诊断准确性。AI总体敏感度89%、特异度88%，但在陈旧骨折特异度降至57%，短小不规则骨敏感度差异大。
要点
- AI骨折检测总体敏感度89%、特异度88%，但陈旧骨折特异度仅57%。
- 骨水平敏感度差异大：腕骨25%-75%，跗骨0%-53%，短小不规则骨准确性差。
- 阅片时对陈旧骨折及短小骨骨折需谨慎，AI结果不能替代人工判读。
核心结论
- AI工具总体诊断准确性高，在多数特定情境下表现稳健，但陈旧骨折存在时特异度显著降低，骨水平分析显示短小不规则骨准确性差。
研究设计与样本
- 回顾性研究，纳入七家丹麦医院连续创伤X线患者2783例（中位年龄38岁，IQR 21-64，女性1443例），以临床放射报告（由急诊医师、资深骨科外科医生和放射学专家评估）为参考标准。
诊断标准与关键数值
- AI工具总体敏感度89%（95% CI 87%-91%），特异度88%（95% CI 86%-89%）。
- 陈旧骨折检查中特异度为57%（95% CI 49%-65%）。
- 腕骨骨折敏感度从其他腕骨25%（95% CI 1%-81%）到三角骨75%（95% CI 43%-95%）。
- 跗骨骨折敏感度从内侧楔骨0%（95% CI 0%-60%）到距骨53%（95% CI 27%-79%）。
临床应用与阅片要点
- AI可作为骨折检测的辅助工具，但需注意陈旧骨折假阳性增加及短小不规则骨假阴性风险。
- 报告时应结合临床病史和人工阅片，对AI提示阴性但临床可疑的短小骨骨折保持警惕。
局限性
- 回顾性研究设计，参考标准为临床放射报告而非独立专家复核，可能引入偏倚。
- 仅纳入七家丹麦医院，外推性受限。
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: To perform an in-depth evaluation of the diagnostic test accuracy of a commercially available AI tool for assistance in fracture detection on radiographs.
MATERIALS AND METHODS: This retrospective study included consecutive patients with trauma radiographs at seven Danish hospitals. The AI output was evaluated using the clinical radiologic report as a reference standard for a binary fracture outcome. The report is based on assessments by an emergency physician, a senior orthopedic surgeon, and a radiology expert. Sensitivity, specificity, positive- and negative predictive values were calculated. Sensitivity and specificity were additionally stratified for children, degenerative disease, metal, old fractures, casting, obvious fractures, and inter-hospital differences. Bone-wise sensitivity and specificity were assessed for multiple fracture cases and individual bones.
RESULTS: The study sample consisted of 2783 patients (median age 38 years, IQR, 21, 64, 1443 female), and 948 (34%) had the target finding. The AI tool demonstrated an overall sensitivity of 89% (95% CI: 87%-91%) and specificity of 88% (95% CI: 86%-89%). The specificity was 57% (95% CI: 49%-65%) in examinations with old fractures. Bone-wise sensitivity for carpal fractures ranged from other carpals 25% (95% CI: 1%-81%] to triquetrum 75% (95% CI: 43%-95%). Tarsal fractures ranged from medial cuneiform 0% (95% CI: 0%-60%) to talus 53% (95% CI: 27%-79%).
CONCLUSION: The AI tool demonstrated high overall diagnostic accuracy and performed robustly across most specific situations. However, specificity was substantially reduced in the presence of old fractures. The bone-wise analysis showed great variability, with a pattern of poor accuracy for short, irregular bones.
KEY POINTS: Question Can a commercially available AI tool reliably detect fractures across anatomical regions, confounding factors, and individual bones -and are there patterns in diagnostic limitations? Findings The AI tool achieved 89% sensitivity and 88% specificity with consistent accuracy across subgroups. However, accuracy dropped for old fractures and irregular short bones. Clinical relevance Despite broad regulatory approval, AI fracture tools may overlook clinically relevant weaknesses. Our in-depth evaluation highlights limitations, guiding responsible clinical use and future research to support safe AI implementation in radiology and informed medicolegal regulation.
引用
- Bruun FJ, Müller FC, Nybing JU, et al. Independent bone-level diagnostic accuracy study of an AI tool for detecting appendicular skeletal fractures on radiographs. Eur Radiol. 2026. PMID: 41995742. doi:10.1007/s00330-026-12489-5
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。