Exploring Real-Time Tracking of Vocal Fold Polyps in Video-Stroboscopy Using Deep Learning.

Area of Science:

• Medical imaging
• Artificial intelligence
• Otolaryngology

Background:

• Vocal fold polyps are lesions that can affect voice quality.
• Accurate and timely detection of vocal fold polyps is crucial for diagnosis and treatment.
• Current detection methods may benefit from AI-driven enhancements.

Purpose of the Study:

• 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).
• To assess the added benefit of temporal tracking on the detection performance of vocal fold polyps.
• To explore the potential of real-time AI-assisted detection of vocal fold lesions.

Main Methods:

• A dataset of 12,742 annotated frames from 55 laryngoscopy videos was used.
• Pretrained YOLO11 and YOLO12 models were fine-tuned for polyp detection.
• A temporal tracking algorithm was developed to improve detection consistency across frames.

Main Results:

• YOLO12 significantly outperformed YOLO11 in polyp detection metrics.
• YOLO12 achieved a precision of 83.1% and an F1 score of 67.6% on the test set.
• Temporal tracking with YOLO12 increased mean average precision (mAP@0.5) to 70.4% at near real-time speeds (21.4 fps).

Conclusions:

• YOLO12 combined with temporal tracking offers enhanced performance for vocal fold polyp detection.
• The system demonstrates near real-time capabilities, facilitating potential clinical applications.
• AI-assisted detection holds promise for improving the identification of vocal fold lesions.