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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Otolaryngology

Background:

• Accurate vocal fold (VF) pose estimation is vital for diagnosing larynx diseases and VF paralysis.
• Current methods like anterior glottic angle (AGA) estimation are subjective, time-consuming, and require expertise.
• Videoendoscopic examination is standard for assessing VF motility.

Purpose of the Study:

• To develop and validate a deep learning framework for precise VF pose estimation from clinical laryngoscopy frames.
• To improve the objective and efficient assessment of VF motility and related parameters like AGA.
• To overcome limitations of existing methods, especially in challenging clinical scenarios.

Main Methods:

• A deep learning framework utilizing heatmap regression on three key anatomical points was developed.
• The framework estimates VF pose and computes AGA from predicted keypoint coordinates.
• A new dataset of 471 laryngoscopy frames from 124 patients was used for evaluation.

Main Results:

• The proposed framework achieved the lowest root mean square error (RMSE) for VF pose estimation (5.09 pixels).
• Heatmap regression demonstrated the lowest mean average error (MAE) for AGA evaluation.
• The method showed superior performance in challenging images, including those with pathology, noise, or occlusion.

Conclusions:

• Keypoint heatmap regression enables accurate VF pose estimation, minimizing errors compared to state-of-the-art methods.
• The framework offers a more objective and efficient approach to VF motility assessment.
• This deep learning solution enhances diagnostic capabilities for larynx diseases in clinical practice.