An automatic laryngoscopic image segmentation system based on SAM prompt engineering: from glottis annotation to...

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Pathology

Background:

• Laryngeal high-speed video (HSV) analysis aids in diagnosing laryngeal diseases.
• Glottis region segmentation is crucial for evaluating vocal fold vibration and detecting disorders.
• Vocal fold segmentation remains an underexplored area in current research.

Purpose of the Study:

• To develop a novel automatic vocal fold segmentation system.
• To leverage glottis information and prompt engineering for vocal fold segmentation.
• To achieve accurate segmentation without requiring supervised vocal fold annotations.

Main Methods:

• Utilized prompt engineering techniques with the Segment Anything Model (SAM).
• Extracted vocal fold features from U-Net-generated glottis masks, enhanced with contrast adjustment and morphological closing.
• Integrated YOLO-v5 for laryngeal region bounding box prompts and derived point prompts from grayscale intensity extrema.

Main Results:

• Achieved a Dice Coefficient of 0.91 for vocal fold segmentation.
• Demonstrated competitive performance compared to fully supervised methods.
• Validated the effectiveness of glottis-based prompts and the absence of need for labeled vocal fold training data.

Conclusions:

• Accurate vocal fold segmentation is feasible using only glottis-based prompts.
• The proposed method offers a viable alternative to supervised learning for vocal fold segmentation.
• The developed system and code are released to promote further research in laryngeal image analysis.