HipSAFE:基于深度学习的猪尸体模型超声髋部骨折检测
本研究在15具猪尸体(30髋)上评估深度学习辅助超声工具HipSAFE在诱导股骨粗隆间骨折前后检测髋部骨折的准确性,使用无超声培训的操作者采集图像,以F1分数为主要指标。EfficientNet-Lite0模型达到F1 0.944、敏感度89.5%、特异度100%,显著高于无经验操作者(F1 0.667)和放射科医生(F1 0.684)。结论认为轻量级深度学习流程可在临床前环境中准确识别超声上的髋部骨折,支持未来院前分诊可行性研究。
要点
- EfficientNet-Lite0模型F1 0.944、敏感度89.5%、特异度100%,优于操作者和放射科医生。
- Grad-CAM提示模型难以区分生长板与骨折特征,可能影响判读可靠性。
- 研究为猪尸体临床前模型,结论需在人体中进一步验证。
关键影像表现/征象
- 在4具尸体8根股骨上由31名操作者采集的78个动态剪辑中评估,多数架构获得高动态剪辑性能。
- EfficientNet-Lite0达到F1 0.944(95%CI 0.920-0.980)、敏感度89.5%(95%CI 75.9%-95.8%)、特异度100.0%(95%CI 91.2%-100.0%)。
- 多数投票集成达到F1 0.932(95%CI 0.903-0.967)。
- 操作者F1为0.667(0.581-0.750),两名诊断放射科医生F1为0.684(95%CI 0.531-0.800)。
- 完整和骨折股骨均存在生长板,Grad-CAM归因图提示模型难以区分生长板与骨折特征。
诊断标准与关键数值
- 以F1分数为主要评价指标,EfficientNet-Lite0在动态剪辑水平F1 0.944(95%CI 0.920-0.980),敏感度89.5%(95%CI 75.9%-95.8%),特异度100.0%(95%CI 91.2%-100.0%)。
- 多数投票集成F1 0.932(95%CI 0.903-0.967)。
- 操作者F1 0.667(0.581-0.750),放射科医生F1 0.684(95%CI 0.531-0.800)。
- 动态剪辑预测采用移动平均和Youden指数阈值校准。
临床应用与阅片要点
- 深度学习辅助超声有望用于院前髋部骨折分诊,减少不必要的转运和缩短手术时间。
- 阅片时注意生长板与骨折特征的鉴别,模型可能混淆两者。
- 需在人体中验证模型性能,尤其不同操作者和设备条件下的泛化能力。
局限性
- 研究为猪尸体临床前模型,结论不能直接推广到人体。
- 样本量有限,仅15具尸体30髋,测试集包含4具尸体8根股骨。
- 模型在区分生长板与骨折特征方面存在困难。
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
OBJECTIVE: Hip fractures require timely diagnosis and surgical intervention, and current approaches require diagnosis with pelvic radiographs in emergency departments. Portable, diagnostically reliable point-of-care ultrasound (US) could reduce unnecessary transfers and shorten time to surgery. The objective of this work was to evaluate, in a porcine cadaver model, the accuracy of HipSAFE, a deep-learning-assisted US tool intended to reduce operator dependence and support pre-hospital hip fracture triage by nurses and paramedics.
METHODS: The bilateral hind limbs of 15 juvenile porcine cadavers (30 hips) were imaged by operators with no sonographic training before and after inducing a displaced iatrogenic pertrochanteric hip fracture. The limbs were split at the cadaver level into training (n = 16), validation (n = 6) and held-out test (n = 8) sets. Frame-level classifiers (MobileNetV3, EfficientNet-Lite and ResNet) were pre-trained on ImageNet and fine-tuned. Cine clip predictions used a moving average with Youden index threshold calibration. Model F1-scores were compared against the diagnostic radiologists using a femur-level paired bootstrap test with Holm correction.
RESULTS: Cine clip performance was reported on the 78 clips collected by 31 operators on 4 cadavers (8 femurs). Most architectures achieved high cine clip performance. EfficientNet-Lite0 reached an F1-score of 0.944 (95% confidence interval [CI], 0.920-0.980), sensitivity of 89.5% (95% CI, 75.9%-95.8%) and specificity of 100.0% (95% CI, 91.2%-100.0%); the majority voting ensemble reached an F1-score of 0.932 (95% CI, 0.903-0.967). The operators achieved an F1-score of 0.667 (0.581--0.750) and the two diagnostic radiologists achieved an F1-score of 0.684 (95% CI, 0.531-0.800). Although growth plates were present in intact and fractured femurs, the Grad-CAM attribution maps suggest that the models had difficulty distinguishing growth plates from fracture features.
CONCLUSION: In this pre-clinical setting, a lightweight, deep-learning pipeline accurately identified hip fractures on US imaging, motivating future work on the feasibility of US with deep learning for pre-hospital hip trauma triage.
引用
- Yee NJ, Soenjaya Y, Kates Rose N, et al. HipSAFE: detecting hip fractures on ultrasound imaging using deep learning in a porcine cadaveric model. Ultrasound Med Biol. 2026. PMID: 42946582. doi:10.1016/j.ultrasmedbio.2026.09.023
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。