An Automatic Laryngoscopic Image Segmentation System Based on SAM Prompt Engineering: From Glottis Annotation to Vocal Fold Segmentation

Yucong Zhang¹Yuchen Song²Juan Liu^1*Ming Li^2*

• ¹Wuhan University, Wuhan, Hubei Province, China
• ²Duke Kunshan University, Kunshan, China

The laryngeal high-speed video (HSV) is a commonly used method for diagnosing laryngeal diseases. Among various approaches, the segmentation of glottis areas on laryngeal images shows great potential in analyzing vocal fold vibration patterns and diagnosing vocal fold disorder. However, few works have been done on vocal fold segmentation. In this study, we present an innovative approach to automatic vocal fold segmentation using only the glottis information. Our system designs prompt engineering techniques customized for the Segment Anything Model (SAM), leveraging glottis data to enhance segmentation accuracy. By combining vocal fold information extracted from U-Net masks-enhanced through brightness contrast adjustment and morphological closing-with a coarse bounding box of the larynx region generated by the YOLO-v5 model, we generate an effective bounding box prompt.Additionally, we introduce a point prompt derived from the local extrema in the first derivative of gray-scale intensity along glottis-intersecting lines, providing auxiliary information on the vocal fold location. Experimental results show that our method that does not need labeled vocal fold training data achieves comparable performance with the fully supervised method, reaching a Dice Coefficient of 0.91. Exemplary features extracted on the segmented masks are included to further show the effectiveness of our work. We release our codes at https: //github.com/yucongzh/Laryngoscopic-Image-Segmentation-Toolkit.