Automatic Pharyngeal Phase Recognition in Untrimmed Videofluoroscopic Swallowing Study Using Transfer Learning with...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Clinical Diagnostics

Background:

• Video fluoroscopic swallowing study (VFSS) is the gold standard for dysphagia evaluation.
• Manual analysis of VFSS videos is time-consuming and labor-intensive.
• Automating VFSS analysis can improve diagnostic efficiency.

Purpose of the Study:

• To develop a deep learning approach for automatic pharyngeal phase annotation in VFSS videos.
• To reduce the manual effort required for VFSS image analysis.
• To enhance the efficiency of dysphagia diagnosis.

Main Methods:

• A single-frame baseline deep convolutional neural network (CNN) architecture was employed.
• Transfer learning with fine-tuning of the VGG-16 model was applied.
• The model was trained to identify pharyngeal phase frames in VFSS videos.

Main Results:

• The fine-tuned VGG-16 model (VGG16-FT5) demonstrated superior performance.
• Achieved 93.20% accuracy, 84.57% sensitivity, and 94.36% specificity.
• High AUC (0.8947) and Kappa (0.7093) values indicate robust performance.

Conclusions:

• The proposed deep CNN framework enables high-performance, automated VFSS video analysis.
• Fine-tuning techniques and explainable AI (grad CAM) are crucial for performance.
• This approach significantly enhances the automation of VFSS analysis for dysphagia evaluation.