Leveraging Ensemble on Self-Supervised Techniques to Enhance Model Performance: A Regime Analysis.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Swallowing Disorders

Background:

• Dysphagia, or difficulty swallowing, presents varied health risks, often due to food or liquid (bolus) entering the airway.
• Manual analysis of videofluoroscopy images for bolus localization is labor-intensive and time-consuming.
• Previous automated approaches primarily utilized supervised learning for bolus segmentation.

Purpose of the Study:

• To develop and evaluate a self-supervised learning network for enhanced bolus segmentation in videofluoroscopy.
• To improve the efficiency and accuracy of swallowing assessments.
• To introduce a novel, automated method for swallowing function evaluation.

Main Methods:

• A self-supervised learning network was constructed using a contrastive random walk model for the pretext task.
• The U-Net++ model served as the downstream task network, with ResNet-18 as the backbone.
• Weights from the pretext task were utilized for initializing the downstream task network.

Main Results:

• The proposed self-supervised network outperformed traditional supervised learning approaches in bolus segmentation.
• A self-supervised learning weighted ensemble model strategy increased the U-Net++ model's F1-score from 79.1% to 81.8% compared to ImageNet initialization.
• The study presents the first known automatic method for efficient swallowing assessment using this approach.

Conclusions:

• Self-supervised learning offers a promising avenue for advancing videofluoroscopy image analysis.
• The developed method enhances the performance of bolus segmentation and swallowing assessment.
• This research pioneers the application of self-supervised learning for automated swallowing function evaluation.