Novel Videographic-Free Framework for Tracking Anatomical Structures Using Swallowing Accelerometer Signals and...

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Machine Learning

Background:

• High-resolution cervical auscultation (HRCA) is a noninvasive swallowing assessment tool using throat accelerometry.
• Traditional methods like videofluoroscopic swallowing studies pose radiation risks and have accessibility limitations.
• Accurate tracking of anatomical landmarks in HRCA remains a challenge for machine learning models.

Purpose of the Study:

• To develop a deep learning multi-task model for precise anatomical landmark tracking during swallowing using HRCA.
• To address the challenge of accurate displacement detection of multiple anatomical structures.

Main Methods:

• Proposed a deep learning multi-task model utilizing transformer encoders for sequential data processing.
• The model was designed to track the displacement of the hyoid bone, laryngeal base, and hyolaryngeal approximation (HLA).
• Evaluated model performance based on relative overlapping (ROP) area for bone and base tracking and accuracy for HLA distance prediction.

Main Results:

• Achieved over 85% average ROP for hyoid bone tracking, surpassing state-of-the-art by over 30%.
• Accurately tracked the laryngeal base with an average ROP exceeding 80%.
• Predicted HLA distance with over 95% average accuracy across all frames.

Conclusions:

• The proposed deep learning model significantly enhances the accuracy of anatomical landmark tracking in HRCA.
• The multi-task learning approach effectively encodes spatial information and the interplay between correlated structures.
• Findings support the integration of HRCA with advanced AI for noninvasive, comprehensive swallowing assessments.