文献速递烧伤创面修复

人工智能在烧伤与创面诊疗中的应用:图像分析、预测与临床整合

Artificial Intelligence in Burn and Wound Care: Image Analysis, Prediction, and Clinical Integration.
PubMed 文献 · J Burn Care Res · 2026-01-01

这是一篇叙述性综述,汇总2015至2025年间的研究,聚焦图像识别与分割、愈合与移植等结局预测、以及人工智能与远程医疗和智能技术的整合三个领域。卷积神经网络的分割Dice系数常超过0.85,烧伤深度或组织类型分类敏感度高于0.90,多模态预测模型的准确度和AUC约为0.80-0.95。早期临床试点显示将人工智能嵌入手机应用、远程医疗流程和传感器敷料具有可行性,但肤色相关的算法偏倚、数据集多样性有限、模型不透明和监管框架演变等挑战仍待解决。

要点

  • 卷积神经网络分割Dice系数常超0.85,烧伤深度分类敏感度高于0.90。
  • 多模态预测模型准确度和AUC约为0.80-0.95,可用于愈合、移植、感染和截肢预测。
  • 智能手机应用、远程医疗和传感器敷料等早期试点已显示整合可行性。

核心结论

  • 人工智能尤其是深度学习计算机视觉,有望为烧伤和复杂创面提供客观、可重复且可规模化的评估。
  • 在安全且公平地大规模部署前,仍需解决肤色算法偏倚、数据集多样性、模型不透明、流程整合和监管框架等问题。

研究设计与样本

  • 叙述性综述,汇总2015至2025年间发表的有关研究。
  • 聚焦图像识别与分割、结局预测建模、人工智能与远程医疗及智能技术整合三个领域。

主要发现

  • 卷积神经网络在多个数据集中分割Dice系数常超过0.85,烧伤深度或组织类型分类敏感度高于0.90。
  • 多模态预测模型准确度和受试者工作特征曲线下面积约为0.80-0.95。
  • 早期临床试点显示将人工智能工具嵌入手机应用、远程医疗流程和传感器敷料具有可行性。

关键数值与推荐

  • 分割Dice系数常超过0.85。
  • 烧伤深度或组织类型分类敏感度高于0.90。
  • 多模态预测模型准确度和AUC约为0.80-0.95。

临床怎么用

  • 可关注人工智能辅助的烧伤深度判断、创面测量和愈合与移植结局预测,作为传统目视评估和人工测量的补充。
  • 可探索将人工智能工具接入手机应用、远程医疗流程和传感器敷料,用于远程监测。

局限性

  • 存在肤色相关的算法偏倚和数据集多样性有限的问题。
  • 模型行为不透明,工作流程整合和监管框架仍在演变。

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)。

原文

Burn injuries and chronic wounds impose a substantial and growing global health and economic burden, particularly in low- and middle-income countries and among aging populations with diabetes, vascular disease, and immobility. Conventional wound assessment depends heavily on visual inspection, manual measurements, and clinician experience, leading to variability in burn-depth estimation, wound sizing, and prognostication. Artificial intelligence, especially deep learning-based computer vision, has emerged as a promising approach to provide objective, reproducible, and scalable evaluation of burns and complex wounds. In this narrative review, we synthesize studies published between 2015 and 2025 focused on 3 domains: image-based wound recognition and segmentation, predictive modeling of outcomes such as healing, graft success, infection, and amputation, and integration of artificial intelligence into telemedicine platforms and smart technologies for remote monitoring. Across multiple datasets, convolutional neural networks achieve segmentation Dice coefficients frequently exceeding 0.85 and burn-depth or tissue-type classification sensitivities above 0.90, while multimodal prediction models reach accuracies and areas under the receiver operating characteristic curve of approximately 0.80-0.95. Early clinical pilots demonstrate the feasibility of embedding artificial intelligence tools into smartphone applications, telehealth workflows, and sensor-enabled dressings. Nonetheless, persistent challenges related to algorithmic bias across skin tones, limited dataset diversity, opaque model behavior, workflow integration, and evolving regulatory frameworks must be addressed before artificial intelligence-enabled wound care systems can be safely and equitably deployed at scale.

来源:J Burn Care Res,查看原文。版权归原作者所有。

引用

  1. Khorsandi J, Mirharooni J, Kahen J, et al. Artificial Intelligence in Burn and Wound Care: Image Analysis, Prediction, and Clinical Integration. J Burn Care Res. 2026. PMID: 42251750. doi:10.1093/jbcr/irag090

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。