A Deep-Learning Approach for Vocal Fold Pose Estimation in Videoendoscopy

Accurate vocal fold (VF) pose estimation is crucial for diagnosing larynx diseases that can eventually lead to VF paralysis. The videoendoscopic examination is used to assess VF motility, usually estimating the change in the anterior glottic angle (AGA). This is a subjective and time-consuming procedure requiring extensive expertise. This research proposes a deep learning framework to estimate VF pose from laryngoscopy frames acquired in the actual clinical practice. The framework performs heatmap regression relying on three anatomically relevant keypoints as a prior for AGA computation, which is estimated from the coordinates of the predicted points. The assessment of the proposed framework is performed using a newly collected dataset of 471 laryngoscopy frames from 124 patients, 28 of whom with cancer. The framework was tested in various configurations and compared with other state-of-the-art approaches (direct keypoints regression and glottal segmentation) for both pose estimation, and AGA evaluation. The proposed framework obtained the lowest root mean square error (RMSE) computed on all the keypoints (5.09, 6.56, and 6.40 pixels, respectively) among all the models tested for VF pose estimation. Also for the AGA evaluation, heatmap regression reached the lowest mean average error (MAE) ( $5.{87}^{\circ }$ ). Results show that relying on keypoints heatmap regression allows to perform VF pose estimation with a small error, overcoming drawbacks of state-of-the-art algorithms, especially in challenging images such as pathologic subjects, presence of noise, and occlusion.