(A701) A Deep Learning Approach to Laryngeal Motion Tracking Under Partial Occlusion in Rats

Introduction: Clinical evaluation of vocal fold motion presents numerous challenges: reliance on subjective assessments that can vary between clinicians, inter-rater variability, and difficulty detecting subtle abnormalities that may hold diagnostic significance. These challenges are compounded by environmental factors: poor visualization due to anatomical occlusions or saliva, which can obscure critical structures. Similarly, in preclinical animal models, laryngoscopy is limited by the same issues, with the added complexity of small-animal anatomy.

Methods: This study focused on assessing vocal fold motion in healthy rats and those with right-sided vocal paresis using a novel automated tracking model. Laryngoscopy videos (N = 42) were collected in a laboratory setting, and 20 nonconsecutive frames per video were manually labeled with key glottic landmarks for model training. The model was iteratively trained and refined with manual corrections to improve tracking accuracy, particularly in frames with occlusion. To evaluate performance, the trained model analyzed 32 videos, with results compared against clinician ratings using a validated 5-point Likert scale. Outcomes included phase synchronization and movement amplitude, converted for direct comparison with clinician scores, demonstrating the model’s robust tracking capabilities even under challenging conditions.

Results: The trained neural network and algorithm effectively analyzed vocal fold motion in videos from 8 rats before, after, and during recovery from right-sided RLN transection. Both human raters and the model consistently rated pre-injury movement as normal (Likert scale: 4) and identified impaired right-sided movement post-injury, with similar ratings (human: 0, model: 0.45 ± 0.37). During recovery, ratings remained comparable, demonstrating the model’s ability to reliably track and evaluate vocal fold function, even under suboptimal video conditions.

Conclusions: The model provides efficient, reliable vocal fold tracking, overcoming occlusion challenges and enhancing automated analysis for laryngology research in rats.