Estimating Musculoskeletal Torque from Ultrasound Images with Regression and Sensor Alignment

Predicting a userâs joint torque is beneficial for improving the design and control of gait-assist robotic devices, ultimately aiding in restoring motor function for individuals with neurological disorders. Joint torque has been shown to correlate with muscle structural characteristics, and brightness-mode (B-mode) ultrasound (US), as a non-invasive, low-cost, and high-sampling-rate technique, has been widely utilized to capture muscle morphology. However, conventional ultrasound-based methods are prone to noise and errors arising from algorithmic limitations and sensor-shift during or across experiments, especially in walking tasks. Although a few machine learning methods (e.g., convolutional neural networks) have attempted to directly predict joint torque from US images, these approaches often require expensive training processes and still fail to address sensor-shift issues. In this study, we demonstrate that regression-based techniques can accurately predict human ankle joint torques during walking, and a virtual sensor alignment method we devised can effectively resolve the sensor-shift problem by adjusting the region of interest. Furthermore, our findings show that models trained at specific walking speeds can be applied to different speeds with the sensor alignment approach. Experimental data from three able-bodied participants validated our methods, and the insights gained from the designs and findings may help guide the development of lightweight predictive models with fast computation suited for gait-assist robotic devices.