Unsupervised Segmentation of Bolus and Residue in Videofluoroscopy Swallowing Studies

Area of Science:

• Biomedical Engineering
• Medical Imaging Analysis
• Artificial Intelligence in Medicine

Background:

• Bolus tracking in videofluoroscopic swallowing studies (VFSS) is crucial for identifying swallowing abnormalities.
• Current machine learning methods for VFSS analysis require extensive annotated data and struggle with subtle residue detection.

Purpose of the Study:

• To develop an unsupervised machine learning architecture for segmenting both bolus and residue in VFSS images.
• To achieve accurate residue detection without relying on pixel-level annotations.

Main Methods:

• An unsupervised convolutional autoencoder architecture was designed for bolus and residue segmentation.
• Positional encoding was integrated to address locality bias and capture global spatial context.
• The model was trained and validated on VFSS images annotated by certified raters.

Main Results:

• The unsupervised model achieved 61% IoU for bolus segmentation, comparable to supervised methods.
• The model achieved 52% IoU for residue detection, significantly outperforming supervised baselines.
• Statistical analysis confirmed the superior performance of the unsupervised method in residue detection.

Conclusions:

• Unsupervised learning offers a robust pathway for segmenting clinically significant but underrepresented features like residue in swallowing analysis.
• This approach reduces the need for extensive manual annotation, making VFSS analysis more accessible and efficient.
• The proposed method demonstrates the potential of learning from negative space for improved diagnostic capabilities in medical imaging.