文献速递儿科影像骨肌DR

儿童肱骨髁上骨折的双路径AI:多分类影像分类与诊断荟萃分析

Dual-Approach AI for Pediatric Supracondylar Fractures: Multiclass Radiograph Classification with Explainable AI and Diagnostic Meta-analysis of AI-Based Computational Approaches.
PubMed 文献 · Acad Radiol · 2026-09-05

本研究基于1082例儿童肘关节X线片,用YOLOv11 Nano开展无骨折与Gartland I–III型多分类,并结合骨分割、可解释AI及诊断准确性荟萃分析。荟萃分析显示合并敏感度92.0%、特异度85.0%,但异质性较高,仍需前瞻性多中心验证。

要点

  • AI可区分Gartland I–III型,III型激活集中在骨折断端边界,可视化可提示阅片重点区域。
  • 骨分割能改善骨折分类表现,多分类提升约3.9个百分点。
  • AI对Gartland I型样本极少,日常工作不能单独用其排除无移位骨折。

核心结论

  • YOLOv11 Nano首次实现儿童肱骨髁上骨折Gartland亚型的多分类深度学习,诊断表现较强。
  • 荟萃分析显示纳入AI方法总体诊断表现中等,证据基础有限且异质性明显,结果仍属初步。
  • 临床部署前需要前瞻性多中心验证。

研究设计与样本

  • 回顾性队列研究,共1082例儿童肘关节X线片,时间为2004至2018年。
  • 多分类类别为无骨折、Gartland I型、II型、III型,样本量分别为576、21、125、360。
  • 模型为YOLOv11 Nano,采用随机划分、分层5折交叉验证、留一交叉验证三种策略。
  • 诊断荟萃分析遵循PRISMA-DTA,纳入4项研究共2232幅图像,使用随机效应模型合成结果。

关键影像表现/征象

  • 可解释AI显示无骨折或I型时注意力集中在骨骺区,II型时骨骺区激活范围扩大。
  • III型骨折时AI关注区域主要位于骨折断端边界。

诊断标准与关键数值

  • 多分类准确率在随机划分、分层k折、留一交叉验证中分别为90.98%、93.12%、90.81%。
  • 二分类准确率在随机划分、分层k折、留一交叉验证中分别为88.89%、92.83%、88.46%。
  • 加入骨分割后多分类准确率一致提高3.9个百分点,二分类准确率提高6.2至10.2个百分点。
  • 荟萃分析合并敏感度92.0%(95% CI:87.0–96.0%),特异度85.0%(95% CI:77.0–92.0%),I²=92.9%。
  • YOLOv11特异度最高为98.0%,诊断比值比365.00。

临床应用与阅片要点

  • Gartland III型以断端边界为AI关注焦点,可用于辅助观察明显移位骨折。
  • Gartland I型样本仅21例,应用于无移位骨折时需谨慎解读阴性结果。
  • 使用骨分割前的X线片可改善AI分类表现,提示实际流程中可先做骨轮廓提取。

局限性

  • 为回顾性研究,且Gartland I型样本量很小,可能限制无移位骨折的泛化能力。
  • 荟萃分析仅纳入4项研究且异质性高,I²=92.9%,合并结果仍需谨慎解释。
  • 纳入研究包含单一放射组学研究,与CNN比较的差异仅有数值趋势,证据不充分。

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: Pediatric supracondylar fractures (SCFs) are the most common elbow injury in children, yet radiographic diagnosis remains challenging due to complex developmental anatomy, with initially missed fracture rates of 17-77%. Prior artificial intelligence (AI) studies have been limited to binary classification frameworks without Gartland subtype differentiation, and no diagnostic test accuracy meta-analysis specific to supracondylar fractures exists. This study aimed to develop the first multiclass deep learning model for SCF detection and Gartland classification and to systematically compare diagnostic performance across AI-based computational diagnostic approaches, including deep learning architectures and radiomics-based machine learning, through meta-analysis.

MATERIALS AND METHODS: A retrospective cohort of 1082 pediatric elbow radiographs (2004-2018) was analyzed using YOLOv11 Nano for multiclass classification (No-SCF, Gartland Type I-III; class distribution: No-SCF n=576, Type I n=21, Type II n=125, Type III n=360). Data augmentation mitigated class imbalance, though the limited Type I sample may constrain generalizability for non-displaced fractures. Three patient-level validation strategies were employed: random split (80:10:10), stratified 5-fold cross-validation, and leave-one-out cross-validation. NormEnsemble-HiResCAM explainable AI (XAI) provided visualization of decision-making patterns. A Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies (PRISMA-DTA)-compliant meta-analysis synthesized data from four studies (2232 images) encompassing convolutional neural network (CNN)-based and radiomics-based approaches using random-effects modeling.

RESULTS: Multiclass accuracy without bone segmentation was 90.98%, 93.12%, and 90.81% across random split, stratified k-fold, and leave-one-out cross-validation, respectively; binary accuracy was 88.89%, 92.83%, and 88.46%. Bone segmentation integration consistently improved accuracy across all strategies (multiclass +3.9 throughout; binary +6.2 to +10.2% points). Meta-analysis revealed pooled sensitivity of 92.0% (95% confidence interval [CI]: 87.0-96.0%) and specificity of 85.0% (95% CI: 77.0-92.0%; I²=92.9%). CNN-based algorithms showed numerically higher specificity than the single included radiomics study, though this comparison warrants caution. YOLOv11 achieved the highest specificity (98.0%) and diagnostic odds ratio (365.00). XAI visualizations demonstrated anatomically stratified attention across fracture grades: physeal-centered activation for normal/Type I, expanded physeal activation for Type II, and fragment-boundary foci for Type III.

CONCLUSION: YOLOv11 Nano provided the first multiclass deep learning framework for pediatric SCF Gartland subtype classification with strong diagnostic performance. Meta-analysis demonstrated moderate diagnostic performance across included studies, though findings remain preliminary given the limited evidence base and substantial heterogeneity. Prospective multicenter validation is required before clinical deployment.

来源:Acad Radiol,查看原文。版权归原作者所有。

引用

  1. Lin, Setiawan, Hikmahtiyar, et al. Dual-Approach AI for Pediatric Supracondylar Fractures: Multiclass Radiograph Classification with Explainable AI and Diagnostic Meta-analysis of AI-Based Computational Approaches. Acad Radiol. 2026. PMID: 42701043. doi:10.1016/j.acra.2026.08.048

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。