Deep Learning Model Development with U-net Architecture for Glottis Segmentation

With the development of technology for detailed examination of the vibration of the vocal cords, the use of highspeed endoscopic camera images has become widespread. Studies conducted to segment the glottis, which is located between the vocal cords of particular interest, from these images use classical segmentation methods such as histogram, active contour, and region enlargement. In this study, a deep learning model was developed for glottis segmentation using the U-Net architecture, which has shown superior performance in biomedical image segmentation in recent years, and the performance of this model was developed on an IRCAM database consisting of 256x256 3000 images, which were manually labeled. It has been compared with classical segmentation methods. However, how the glottis size affects the segmentation performance was examined. As a result of the study, it was observed that system performances decreased as the glottal area size decreased, especially active contour performance decreased significantly, but this effect remained limited in the U-Net model. The developed model showed the highest performance with 0.88 sensitivity and 0.99 accuracy on average regardless of the size of the glottal area. In terms of recall, the histogram-based method yielded the highest result with 0.96. In terms of DICE, U-Net gave the best result with a score of 0.83.