Hyoid Bone Tracking in a Videofluoroscopic Swallowing Study Using a Deep-Learning-Based Segmentation Network.

Area of Science:

• Biomedical Engineering
• Medical Imaging Analysis
• Swallowing Disorders Research

Background:

• Kinematic analysis of hyoid bone movement in videofluoroscopic swallowing studies (VFSS) is crucial for dysphagia assessment.
• Current manual calibration methods are time-consuming and exhibit variable reliability.
• Existing computer-assisted analyses for hyoid bone tracking show performance limitations.

Purpose of the Study:

• To design a robust deep learning network for automated hyoid bone movement tracking in VFSS.
• To improve the efficiency and accuracy of hyoid bone identification and trajectory prediction.
• To develop a system that eliminates the need for human intervention in hyoid bone analysis.

Main Methods:

• A deep learning model, BiFPN-U-Net(T), was developed for detection and trajectory prediction.
• The model was trained using a dataset of 69,389 frames from 197 VFSS files.
• Performance was evaluated using metrics including area under the curve (AUC), object detection accuracy, Dice similarity, and mean average precision (mAP).

Main Results:

• The model achieved an AUC of 0.998 for pixelwise accuracy and 99.5% object detection accuracy.
• Dice similarity reached 90.9%, with a mean average precision of 95.9% for bounding box detection.
• The system demonstrated superior performance in detecting and tracking the hyoid bone and reference objects compared to previous models.
• Accurate estimation of hyoid bone movement distance was achieved.

Conclusions:

• The proposed deep learning model enables efficient and accurate automatic detection and tracking of the hyoid bone in VFSS.
• This technology has the potential to significantly improve the clinical assessment of dysphagia.
• The automated approach offers a reliable alternative to manual calibration, reducing analysis time and variability.