UltraMN:多任务深度学习实时正中神经超声监测框架
本研究开发多任务深度学习框架 UltraMN,整合标准切面分类与组织分割。纳入446例健康受试者,双侧采集8段视频,共3568段视频、249985幅图像,覆盖四个标准成像切面。UltraMN分类准确率95.6%,分割mIoU 97.6%,优于MedMamba和FPT。
要点
- UltraMN可同时完成正中神经标准切面分类与组织分割,适合辅助实时超声监测。
- 四个标准切面的分类精确率、召回率、F1值均超过95.0%,分割mIoU达97.6%。
- 该研究仅基于健康正中神经数据,对腕管综合征等病理状态的可推广性仍需验证。
核心结论
- UltraMN在健康受试者正中神经超声中实现95.6%的标准切面分类准确率和97.6%的分割mIoU。
- UltraMN显著优于MedMamba和FPT,为正中神经疾病管理的临床流程优化奠定基础。
研究设计与样本
- 研究类型为初步可行性研究,对象为446例健康受试者。
- 每例双侧采集8段视频,共3568段视频、249985幅图像,覆盖四个标准成像切面。
- 将UltraMN的分类准确率与MedMamba、FPT模型进行比较,并评估分割精确率、召回率、F1及mIoU。
- 统计分析使用Python 3.12.9完成。
关键影像表现/征象
- UltraCLS模块在四个标准切面均达到超过95.0%的分类精确率、召回率和F1值。
- UltraSEG模块在所有成像切面均表现优越,整体mIoU为97.6%。
诊断标准与关键数值
- UltraCLS分类准确率为95.6%,精确率、召回率、F1值在各标准切面均超过95.0%。
- UltraSEG分割mIoU为97.6%,该数值对应四个标准成像切面的组织分割任务。
- 对比模型为MedMamba和FPT,UltraMN的分类表现显著优于两者。
临床应用与阅片要点
- 该框架可用于辅助正中神经超声检查中的标准切面识别与组织分割,提升实时监测效率。
- 在阅片和报告中可借助UltraMN的切面分类结果确认当前扫查是否位于标准切面。
- 注意该模型未在病理状态下验证,当前不应用于腕管综合征等疾病的诊断替代。
局限性
- 研究仅基于健康受试者正中神经超声数据,未纳入腕管综合征等病理状态。
- 作为初步可行性研究,后续需要在病理病例中进一步测试验证其可推广性。
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: This study aims to develop an advanced deep learning framework to overcome the challenges associated with real-time ultrasound monitoring of the median nerve.
METHOD: We propose UltraMN, a novel multitask learning model integrating standard plane classification (UltraCLS) and tissue segmentation (UltraSEG) for comprehensive analysis. The dataset includes 446 cases, with 8 videos collected bilaterally for each case, resulting in a total of 3568 videos and 249,985 images across four standardized imaging planes (4-SIP). The classification accuracy of the median nerve was compared among UltraCLS, MedMamba, and FPT models, while precision, recall, F1 scores, and mean Intersection over Union (mIoU) for 4-SIP segmentation were evaluated. Statistical analyses were conducted using Python 3.12.9.
RESULTS: UltraMN significantly outperformed MedMamba and FPT. The UltraCLS module achieved a classification accuracy of 95.6%, with precision, recall, and F1 scores exceeding 95.0% across all standard planes. The UltraSEG module achieved an mIoU of 97.6%, demonstrating superior segmentation performance across all imaging planes.
CONCLUSION: UltraMN offers a robust and efficient solution for real-time assessment, achieving high classification accuracy and precise segmentation. As a preliminary feasibility study on healthy subjects, this work is based exclusively on ultrasound data of healthy median nerves-its generalizability to pathological scenarios (e.g., carpal tunnel syndrome) requires further validation. It lays the foundation for enhancing clinical workflows in median nerve disorder management, pending subsequent testing on pathological cases.
引用
- Zhou, Xiang, Guo, et al. UltraMN: Advancing Real-Time Median Nerve Ultrasound Monitoring With a Multitask Deep Learning Framework. Ultrasound Med Biol. 2026. PMID: 41506963. doi:10.1016/j.ultrasmedbio.2025.12.002
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。