文献速递鼻科

机器学习辅助鼻内镜诊断鼻窦炎的临床决策支持

Machine Learning-Enhanced Clinical Decision Support for Diagnosing Sinusitis With Nasal Endoscopy.
PubMed 文献 · Int Forum Allergy Rhinol · 2025-10-15

研究基于452例患者的3513张鼻内镜图像训练YOLOv11-nano多类检测分割模型,结合规则逻辑定位中鼻道诊断鼻窦炎。模型对含黏液的鼻甲检测F1>75%,分类敏感度75.0%、特异度76.0%、准确率75.2%,F1 81.8%,接近耳鼻咽喉科医生水平,GPU上>20fps近实时。

要点

  • 多类模型检测鼻甲与黏液F1>75%,分类准确率75.2%。
  • 规则算法诊断鼻窦炎敏感度75.0%、特异度76.0%。
  • GPU上>20fps近实时,适合临床工作流集成。

核心结论

  • 机器学习框架结合规则临床算法可提升鼻内镜诊断鼻窦炎的决策,性能接近耳鼻咽喉科医生。

研究设计与样本

  • 回顾性研究,452例患者共3513张鼻内镜图像由4名医生手动标注中鼻甲、下鼻甲和黏液三类。
  • 训练YOLOv11-nano模型进行多类检测与分割,并开发基于解剖IoU和条件逻辑的中鼻道定位规则算法。
  • 在50例无鼻息肉慢性鼻窦炎(CRSsNP)患者的178张图像上验证并基准测试实时性能。

主要发现

  • 多类检测分割模型对含黏液的鼻甲检测F1分数>75%。
  • 临床算法分类鼻窦炎的敏感度75.0%、特异度76.0%、准确率75.2%,F1分数81.8%。
  • 框架在GPU设备上达到>20fps的近实时性能。

关键数值与推荐

  • 敏感度75.0%、特异度76.0%、准确率75.2%、F1分数81.8%用于鼻窦炎分类。
  • 含黏液鼻甲检测F1分数>75%。
  • 实时性能>20fps on GPU device。

临床怎么用

  • 可辅助非专科医生标准化鼻内镜判读,减少操作者间变异。
  • 近实时处理支持集成到临床工作流,提升诊断效率。

局限性

  • 未来需扩展至不同鼻窦炎表型并在临床环境中前瞻性实时实施。

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

BACKGROUND: Sinusitis is a prevalent disease for which nasal endoscopy (NE) is an optimal diagnostic modality. However, NE accuracy is limited by inter-operator variability in landmark identification and localization of mucus that is necessary for sinusitis diagnosis. We sought to develop a novel multi-class machine learning (ML) framework that detects anatomical landmarks and structures for sinusitis assessment as supported by clinical best practices. METHODS: A total of 3513 NE images from 452 patients were manually annotated by four physicians for three classes: middle turbinate (MT), inferior turbinate (IT), and mucus. A YOLOv11-nano model was trained for multi-class detection and segmentation. We developed a rule-based logic for middle meatus localization, implementing a clinical algorithm that applies anatomy Intersection over Union (IoU) and conditional logic for sinusitis diagnosis. The system was validated on 178 images from 50 patients with chronic rhinosinusitis without polyps (CRSsNP) with benchmarking of real-time performance. RESULTS: The multi-class detection and segmentation model achieved > 75% F1 score for detecting turbinates with mucus. The clinical algorithm achieved 75.0% sensitivity, 76.0% specificity, and 75.2% accuracy for sinusitis classification, with a F1 score of 81.8%, approaching the accuracy of a trained otolaryngologist. The framework achieved near real-time performance at > 20fps on GPU device, demonstrating suitability for integration into live clinical workflows. CONCLUSION: This novel ML-driven diagnostic framework with a rule-based clinical algorithm enhances decision-making for diagnosing sinusitis with NE. By reducing inter-operator variability, achieving performance comparable to otolaryngologists, and enabling real-time processing for non-specialists, this work holds potential for standardizing care and improving patient outcomes. Future research will focus on expanding to different sinusitis phenotypes and prospective real-time implementation in clinical settings.

来源:Int Forum Allergy Rhinol,查看原文。版权归原作者所有。

引用

  1. Gyawali D, Mundy T, Hosseini M, et al. Machine Learning-Enhanced Clinical Decision Support for Diagnosing Sinusitis With Nasal Endoscopy. Int Forum Allergy Rhinol. 2025. PMID: 41091528. doi:10.1002/alr.70045

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。