文献速递骨肌技术与安全DR

基于YOLO26x的X线骨折检测及对放射科医师表现影响

YOLO26x-based automated fracture detection on radiographs and its impact on radiologist performance: A multi-reader multi-case study.
PubMed 文献 · Eur J Radiol · 2026-06-17

本研究基于YOLO26x架构开发并验证一个使用1280×1280像素高分辨率输入的四肢骨折X线检测模型。测试集共500张图像(250例骨折阳性,250例阴性;250例来自机构,250例来自开放源),模型图像级AUC达0.847。三名放射科医师(年资3年、6年、21年)参与的MRMC研究显示,AI辅助显著提高两名医师的诊断准确性,整体中位判读时间从12.9秒降至8.6秒,读者间一致性Fleiss κ从0.432提升至0.642。

要点

  • 模型图像级AUC为0.847,F1-score为0.799,机构与开放源子集表现相近。
  • AI辅助后三名医师中两名准确性显著提高,中位判读时间减少33.2%。
  • 读者间Fleiss κ从0.432升至0.642,所有医师自评诊断信心均显著提升。

核心结论

  • YOLO26x模型在四肢骨折X线检测中图像级AUC达0.847(95% CI:0.811-0.881),在机构与开放源数据集间表现稳定(AUC分别为0.843和0.852)。
  • AI辅助显著提高3名医师中2名的诊断准确性(RAD1:76.2%→83.4%,p=0.001;RAD3:75.6%→83.4%,p<0.001),整体中位判读时间减少33.2%(12.9→8.6秒,p<0.001)。
  • 读者间一致性明显改善(Fleiss κ:0.432→0.642),所有医师自评诊断信心均显著提高(均p<0.001)。

研究设计与样本

  • 研究类型为模型开发验证结合MRMC多读者多病例研究,采用顺序设计(先无AI读片,间隔3周后再有AI辅助读片)。
  • 训练集共8690张四肢X线片,来源于机构PACS及开放源数据库,采用迁移学习训练YOLO26x目标检测模型。
  • 测试集共500张图像,均衡设计(250例骨折阳性,250例阴性;250例机构来源,250例开放源来源)。
  • 三名放射科医师参与MRMC研究,年资分别为3年、6年和21年。
  • 统计方法包括McNemar检验(Holm-Bonferroni校正)、Wilcoxon符号秩检验、Cohen κ和Fleiss κ统计量及来源分层分析。

关键影像表现/征象

  • 模型采用1280×1280像素高分辨率输入,以保留细微骨折细节特征。
  • 来源分层分析显示模型在机构来源子集(AUC 0.843)与独立开放源子集(AUC 0.852)检测表现相当。
  • 决策改变分析显示净正效应:在1500对读者-病例中,有利改变197次,不利改变104次。

诊断标准与关键数值

  • 模型在主要工作阈值下图像级AUC-ROC为0.847(95% CI:0.811-0.881),F1-score为0.799(95% CI:0.760-0.834)。
  • AI辅助后RAD1准确性从76.2%升至83.4%(p=0.001,Holm校正后显著),RAD3从75.6%升至83.4%(p<0.001,Holm校正后显著)。
  • AI辅助后敏感度提高幅度为2.4-16.4个百分点。
  • 整体中位判读时间从12.9秒降至8.6秒(p<0.001),所有医师自评信心提升均p<0.001。
  • Fleiss κ一致性在AI辅助后从0.432升至0.642。

临床应用与阅片要点

  • 该AI系统可作为急诊放射科骨折检测的辅助决策工具,帮助减少漏诊。
  • AI辅助可缩短判读时间,适合高流量急诊环境中的效率提升。
  • AI辅助提高了低年资与高年资放射科医师的诊断一致性和信心,有利于报告标准化。

局限性

  • 原文指出在常规临床应用前仍有待前瞻性多中心验证和工作流程整合研究。
  • MRMC研究仅纳入三名放射科医师,样本量有限。
  • 测试集采用骨折阳性与阴性各250例的均衡设计,与真实临床发病率和先验概率不同。

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 AND AIMS: Missed fractures are among the most common diagnostic errors in emergency radiology and may lead to delayed treatment and adverse outcomes. Deep learning-based approaches have shown promise for automated fracture detection; however, many existing models are restricted to specific anatomical regions, operate on downsampled images, or focus primarily on image classification rather than robust localization. This study aimed to develop and validate a YOLO26x-based deep learning model trained using high-resolution input (1280 × 1280 pixels) for detection of appendicular fractures on radiographs and to evaluate its impact on radiologist diagnostic performance, reading time, and diagnostic confidence in a multi-reader multi-case (MRMC) study.

METHODS: A total of 8,690 appendicular radiographs from an institutional picture archiving and communication system (PACS) and an open-source repository were used to train a YOLO26x-based object detection model using transfer learning. High-resolution input (1280 × 1280 pixels) was used to preserve fine fracture detail. The test set comprised 500 images with a balanced design (250 fracture-positive, 250 fracture-negative; 250 institutional, 250 open-source), ensuring equal representation of both data sources and providing a prevalence-controlled evaluation of diagnostic performance. Model performance was assessed at both image and object levels. A multi-reader multi-case (MRMC) study involving three radiologists (3, 6, and 21 years of experience) evaluated the effect of model assistance on diagnostic accuracy, efficiency, and self-reported confidence using a sequential unassisted-then-assisted design with a 3-week washout interval between sessions. Statistical comparisons included McNemar's test with Holm-Bonferroni correction, Wilcoxon signed-rank tests, Cohen's and Fleiss' κ statistics, and source-stratified analyses.

RESULTS: The model achieved an image-level AUC-ROC of 0.847 (95 % CI: 0.811-0.881) with an F1-score of 0.799 (95 % CI: 0.760-0.834) at the primary operating threshold. Source-stratified analysis demonstrated comparable performance across institutional (AUC 0.843) and independent open-source (AUC 0.852) subsets. Model-assisted reading significantly improved accuracy for two of three readers (RAD1: 76.2 % → 83.4 %, p = 0.001; RAD3: 75.6 % → 83.4 %, p < 0.001; Holm-corrected), with sensitivity gains of 2.4-16.4 percentage points. Inter-reader agreement improved substantially (Fleiss' κ: 0.432 → 0.642). Median interpretation time decreased by 33.2 % overall (12.9 → 8.6 s; p < 0.001), and self-reported confidence increased significantly across all readers (all p < 0.001). Decision-change analysis demonstrated a net positive effect, with beneficial changes substantially outnumbering detrimental ones (197 vs 104 across 1,500 reader-case pairs).

CONCLUSIONS: The proposed YOLO26x model trained using high-resolution input demonstrated robust image-level fracture detection with consistent performance across institutional and independent open-source data sources. Model assistance improved diagnostic accuracy, efficiency, inter-reader agreement, and diagnostic confidence. These findings support the potential of high-resolution deep learning-based systems as clinically practical decision-support tools in emergency radiology, while prospective multicenter validation and workflow integration studies remain warranted prior to routine clinical implementation.

IMPACT AND IMPLICATIONS: AI-assisted fracture detection using high-resolution radiograph input may enhance diagnostic consistency, improve interpretation efficiency, and help reduce missed-injury rates in high-volume emergency settings. With additional validation and workflow integration, such systems may support triage prioritization and clinical decision-making in emergency radiology.

来源:Eur J Radiol,查看原文。版权归原作者所有。

引用

  1. Pamuk, Yüce, Öztürk, et al. YOLO26x-based automated fracture detection on radiographs and its impact on radiologist performance: A multi-reader multi-case study. Eur J Radiol. 2026. PMID: 42330673. doi:10.1016/j.ejrad.2026.113025

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。