A Real-time Scale-robust Network for Glottis Segmentation in Nasal Transnasal Intubation

Area of Science:

• Medical Imaging
• Computer Vision
• Anesthesiology

Background:

• Nasotracheal intubation (NTI) is vital for airway management.
• Machine-assisted NTI enhances procedural efficiency but faces visual detection challenges.
• Glottis visibility varies, and traditional methods are computationally intensive.

Purpose of the Study:

• To develop an optimized glottis segmentation framework for vision-assisted NTI.
• To address challenges like scale variability and suboptimal illumination in NTI.
• To improve the real-time detection accuracy and computational efficiency of NTI navigation.

Main Methods:

• Designed a lightweight, multi-receptive field feature extraction module for scale variation robustness.
• Developed an advanced label assignment method and redefined sample numbers to reduce intra-class differences.
• Implemented a novel glottis segmentation network (GlottisNet) for vision-assisted NTI.

Main Results:

• Achieved a segmentation mDice of 92.9% on three datasets.
• Demonstrated superior performance compared to state-of-the-art algorithms.
• Obtained a compact model size (19 MB) and high inference speed (>170 fps).

Conclusions:

• The proposed GlottisNet framework enhances NTI efficacy through accurate and efficient glottis segmentation.
• The network's design effectively handles scale variations and complex visual conditions.
• GlottisNet offers a promising solution for real-time, high-precision detection in portable NTI devices.