Integrating Dynamic Representation and Multi-Priors for Transnasal Intubation via Visual Foundation Model

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate glottis localization is crucial for safe nasotracheal intubation.
• Foundation models like Segment Anything Model have limitations in medical image segmentation due to rigid mechanisms and poor generalization.

Purpose of the Study:

• To develop Glottis-SAM, a novel, lightweight, and task-adaptive segmentation framework for precise glottis localization.
• To address the limitations of existing models in segmenting complex glottal anatomies.

Main Methods:

• Proposed Glottis-SAM, integrating dynamic representation learning and multi-prior contextual modeling.
• Introduced a hierarchical low-rank adaptation for efficient fine-tuning of visual foundation models.
• Designed a feature aggregation module with dual-path dynamic feature pyramids for enhanced semantic fusion.

Main Results:

• Achieved state-of-the-art segmentation accuracy with 72.6% mDice.
• Demonstrated a compact model size (55.2 MB) and high inference speed (44.3 FPS).
• Validated robustness and efficiency on three diverse datasets.

Conclusions:

• Glottis-SAM offers a robust and efficient solution for glottis segmentation.
• The framework shows potential for real-time deployment in visual guidance systems for nasotracheal intubation.
• Addresses key challenges in medical image segmentation for improved patient safety.