Network for Real-time Laryngeal Lesions Video Object Detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Early diagnosis of nasopharyngeal-laryngeal tumors is crucial for patient outcomes.
• Deep learning excels at static image lesion detection but struggles with video quality issues like motion blur and artifacts.
• Existing methods are suboptimal for detecting lesions in challenging endoscopic video conditions.

Purpose of the Study:

• To develop a novel deep learning network, DynSTPN, for accurate lesion detection in nasopharyngeal-laryngeal endoscopic videos.
• To address challenges posed by motion blur, uneven exposure, and artifacts in endoscopic video analysis.
• To improve the real-time detection capabilities for clinical nasopharyngeal-laryngeal examinations.

Main Methods:

• Proposed a two-stage video lesion detection network, DynSTPN.
• Implemented a dynamic prompt generator using spatio-temporal features from reference frames to mitigate quality degradation.
• Introduced an adaptive differentiable gating mechanism to integrate reference frame information for enhanced inference frame analysis.

Main Results:

• DynSTPN achieved a superior detection accuracy of 79.6% and speed of 29.4 FPS on the NLLVOD dataset, meeting real-time clinical requirements.
• Outperformed SOTA static image detector YOLOv12-M on the NLLVOD dataset.
• Demonstrated a strong balance between detection accuracy and efficiency on the ImageNet VID dataset compared to SOTA methods.

Conclusions:

• DynSTPN effectively leverages video reference frames to enhance lesion detection performance in challenging endoscopic scenarios.
• The proposed method significantly improves accuracy and efficiency over existing static and video-based approaches.
• DynSTPN shows enhanced clinical applicability for real-time nasopharyngeal-laryngeal tumor diagnosis.