文献速递咽喉与嗓音

深度学习在动态喉镜视频中实时追踪声带息肉

Exploring Real-Time Tracking of Vocal Fold Polyps in Video-Stroboscopy Using Deep Learning.
PubMed 文献 · Laryngoscope · 2026-02-02

本研究基于55段喉镜视频的12,742帧标注数据,微调YOLO11与YOLO12模型检测声带息肉,并开发时序追踪算法传播漏检。YOLO12在测试集上精确率83.1%、F1 67.6%、mAP@0.5 64.1%,加入时序追踪后mAP@0.5提升至70.4%,检测速度21.4 fps接近实时。

要点

  • YOLO12检测声带息肉优于YOLO11,测试集精确率83.1%、mAP@0.5 64.1%。
  • 时序追踪将YOLO12的mAP@0.5从64.1%提升至70.4%。
  • 检测速度21.4 fps接近实时30 fps,具实时AI辅助潜力。

核心结论

  • YOLO12结合时序追踪在动态喉镜中检测声带息肉达到mAP@0.5 70.4%且接近实时,展示实时AI辅助检测声带病变的潜力。

研究设计与样本

  • 回顾性数据集包含55段喉镜视频的12,742帧标注图像,以边界框标注声带息肉。
  • 微调预训练YOLO11与YOLO12模型进行帧内息肉检测,并开发时序追踪算法传播漏检。

主要发现

  • YOLO12在所有指标上优于YOLO11。
  • 留出测试集上YOLO12精确率83.1%、F1 67.6%、mAP@0.5 64.1%。
  • YOLO11精确率67.3%、F1 56.2%、mAP@0.5 56.0%。
  • YOLO12加入时序追踪后mAP@0.5升至70.4%,检测速度21.4 fps。

临床怎么用

  • 该技术有望用于动态喉镜中声带息肉的实时AI辅助检测,但需进一步临床验证。

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

OBJECTIVE: To develop and evaluate a deep learning object detection system for identifying vocal fold polyps in stroboscopic video frames using You Only Look Once (YOLO), and to assess the added benefit of temporal tracking on detection performance. METHODS: A retrospective dataset of 12,742 annotated frames from 55 laryngoscopy video recordings was annotated with bounding boxes identifying vocal fold polyps. Pretrained YOLO11 and YOLO12 models were fine-tuned to detect the polyps in the frames. A temporal tracking algorithm was further developed to propagate missed detections across adjacent frames. RESULTS: YOLO12 outperformed YOLO11 across all metrics. On the hold-out test set, YOLO12 reached a precision of 83.1% and an F1 score of 67.6%, with a mean average precision at 0.5 (mAP@0.5) of 64.1%. By comparison, YOLO11 achieved a precision of 67.3% and an F1 score of 56.2%, with a mAP@0.5 of 56.0%. Incorporating temporal tracking increased mAP@0.5 to 70.4% with YOLO 12, while maintaining a detection speed of 21.4 frames per second (fps), close to real time (30 fps). CONCLUSIONS: Using YOLO 12 for vocal fold polyp detection in stroboscopy was enhanced with temporal tracking, achieving a mAP@0.5 of 70.4% with near real time performance. These results demonstrate the potential of real-time AI-assisted detection of vocal fold lesions.

来源:Laryngoscope,查看原文。版权归原作者所有。

引用

  1. Kaza S, Mohanty AS, Serpedin A, et al. Exploring Real-Time Tracking of Vocal Fold Polyps in Video-Stroboscopy Using Deep Learning. Laryngoscope. 2026. PMID: 41623203. doi:10.1002/lary.70396

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。