Automated Laryngeal Invasion Detector of Boluses in Videofluoroscopic Swallowing Study Videos Using Action...

Area of Science:

• Medical Imaging
• Swallowing Disorders
• Artificial Intelligence

Background:

• Laryngeal invasion during swallowing poses significant clinical challenges.
• Accurate detection of laryngeal invasion is crucial for patient management.
• Current diagnostic methods may be time-consuming for clinicians.

Purpose of the Study:

• To develop an automated detector for laryngeal invasion using 3D stream networks.
• To evaluate the model's performance against existing image classification architectures.
• To establish a robust and accurate tool for identifying laryngeal invasion in VFSS videos.

Main Methods:

• Utilized two 3D stream networks for action recognition in videofluoroscopic swallowing study (VFSS) videos.
• Trained the model to detect laryngeal invasion, defined as Penetration-Aspiration Scale (PAS) scores of 2 or higher.
• Compared the proposed model's accuracy, precision, recall, and F1 scores with established image classification models.

Main Results:

• The automated detector achieved an accuracy of 92.10% in identifying laryngeal invasion (PAS ≥ 2).
• Precision, recall, and F1 scores for laryngeal invasion detection were all 0.9470.
• The model significantly outperformed other image classification models, including ResNet101, Swin-Transformer, EfficientNet-B2, and HRNet-W32.

Conclusions:

• This study presents the first automated detector for laryngeal invasion in VFSS videos based on video action recognition networks.
• The developed model demonstrates high and balanced performance, making it a potential effective screening tool.
• The automated detector can assist clinicians by providing preliminary analysis of VFSS videos, reducing their workload.