Deep learning-based video analysis for automatically detecting penetration and aspiration in videofluoroscopic...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Swallowing Disorders

Background:

• Videofluoroscopic swallowing study (VFSS) is the gold standard for diagnosing dysphagia.
• VFSS interpretation is time-consuming and requires specialized expertise.
• Automating VFSS analysis can improve diagnostic efficiency.

Purpose of the Study:

• To develop and assess a deep learning model for automatic detection of penetration and aspiration in VFSS.
• To evaluate the diagnostic accuracy of the model in identifying normal swallowing, penetration, and aspiration.

Main Methods:

• A convolutional neural network architecture was employed to analyze automatically extracted images from VFSS videos.
• Key image frames corresponding to hyoid bone positions were selected.
• Techniques to handle class imbalance and optimize model performance were incorporated.

Main Results:

• The model achieved high patient-level diagnostic accuracy: AUC 0.935 for normal swallowing, 0.889 for penetration, and 0.845 for aspiration.
• The model demonstrated strong performance in identifying normal swallowing.
• Sensitivity for detecting penetration and aspiration was notably lower.

Conclusions:

• The developed deep learning model shows potential to reduce VFSS interpretation time by automating the detection of penetration and aspiration.
• Clinicians may be able to focus more on other critical VFSS findings.
• Future research should explore full-frame analysis and multicenter datasets to enhance model generalizability and address limitations.