Estimating Musculoskeletal Torque from Ultrasound Images with Regression and Sensor Alignment

Area of Science:

• Biomechanics and Robotics
• Biomedical Engineering
• Rehabilitation Technology

Background:

• Predicting joint torque aids in developing advanced gait-assist robotic devices for motor function restoration.
• Ultrasound (US) imaging captures muscle morphology, correlating with joint torque, but faces noise and sensor-shift challenges.
• Existing machine learning methods for torque prediction from US images are costly and don't resolve sensor-shift.

Purpose of the Study:

• To accurately predict human ankle joint torques during walking using regression-based techniques.
• To develop and validate a virtual sensor alignment method to address sensor-shift errors in ultrasound data.
• To assess the generalizability of trained models across different walking speeds.

Main Methods:

• Utilized brightness-mode (B-mode) ultrasound to capture muscle morphology during walking tasks.
• Employed regression-based techniques for direct prediction of joint torques from ultrasound data.
• Devised a virtual sensor alignment method to correct for sensor-shift by adjusting the region of interest.

Main Results:

• Demonstrated accurate prediction of human ankle joint torques during walking.
• The virtual sensor alignment method effectively resolved sensor-shift issues.
• Models trained at specific walking speeds were successfully applied to different speeds using the alignment approach.

Conclusions:

• Regression techniques combined with a novel sensor alignment method provide accurate and robust joint torque prediction.
• The developed approach overcomes limitations of previous methods, offering a pathway for improved gait-assist robotic devices.
• Findings support the development of lightweight, computationally efficient predictive models for real-time robotic applications.