A Human-Centered Machine-Learning Approach for Muscle-Tendon Junction Tracking in Ultrasound Images.

Area of Science:

• Biomechanics
• Medical Imaging
• Machine Learning

Background:

• Gait analysis in biomechanics relies on observing limb muscles and tendons.
• Accurate tracking of anatomical landmarks like muscle-tendon junctions is crucial for gait research.
• Current methods for tracking these junctions can be time-consuming.

Purpose of the Study:

• To develop a reliable and time-efficient machine learning approach for tracking muscle-tendon junctions in ultrasound videos.
• To support clinical biomechanists in performing gait analysis more effectively.
• To provide a publicly accessible tool for researchers.

Main Methods:

• A deep learning-based method was developed to track muscle-tendon junctions.
• An extensive dataset of 66,864 annotated ultrasound images was created, covering diverse movements, muscles, subjects, and ultrasound systems.
• Data was collected across independent laboratories and curated by researchers with varied experience levels.

Main Results:

• The proposed deep learning model demonstrated performance comparable to human specialists in identifying muscle-tendon junction positions.
• The method achieves time-efficient tracking, with prediction times as fast as 0.078 seconds per frame, approximately 100 times faster than manual labeling.
• A diverse test set, independently verified by four specialists, was used for evaluation.

Conclusions:

• The developed machine learning approach offers a reliable and significantly faster alternative for tracking muscle-tendon junctions in gait analysis.
• The public availability of the code, models, and dataset promotes further research and application in biomechanics.
• The free online service democratizes access to advanced gait analysis tools.