Advancing modified barium swallow pre-sorting with deep learning: a new paradigm for the first step analysis in X-ray...

Area of Science:

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

Background:

• Modified Barium Swallow (MBS) exams are crucial for evaluating swallowing function.
• MBS exams generate diverse video segments, including diagnostic planes (AP, lateral) and non-diagnostic scout views.
• Manual sorting and labeling of MBS video files complicate the pre-analysis workflow.

Purpose of the Study:

• To introduce a deep learning approach for automating the categorization of swallow videos in MBS exams.
• To distinguish between diagnostic video planes (AP, lateral) and identify non-diagnostic scout videos.
• To streamline the MBS review workflow and enhance pre-analysis efficiency.

Main Methods:

• Development of deep learning algorithms for video segment categorization.
• Training on a dataset of 3,740 video segments (986,808 frames) from 285 MBS exams.
• Utilizing a multi-task learning approach to improve differentiation accuracy.

Main Results:

• Frame-level accuracy of 99.68% for differentiating AP and lateral planes; 100% video-level accuracy.
• Frame-level accuracy of 90.26% for distinguishing scout from bolus swallowing videos; 93.86% video-level accuracy.
• Multi-task learning improved video-level scout/bolus differentiation to 96.35%.

Conclusions:

• Deep learning effectively automates MBS video classification, significantly boosting processing efficiency.
• Leveraging inter-frame connectivity enhances model performance in MBS video analysis.
• Reduced manual sorting allows clinicians to focus more on interpretation and patient care.