Ultrasound-Based Sensing Strategy for Control of Upper Extremity Prosthetic

A number of clinical and scientific applications require the ability to sense complex synergies of muscle activity non-invasively and infer volitional motor intent. These include assistive devices for rehabilitation of motor impairments (such as multi-articulated prosthetic hands, robots, and exoskeletons) and investigations of motor control, biomechanics, and human factor studies. Surface electromyography (sEMG), which measures the electrical activity of motor units at the skin surface, has been the predominant method for sensing muscle activity for these applications; however, there are fundamental limitations to sEMG sensing strategies, such as limited specificity and low SNR. While pattern recognition strategies have been used to improve the functionality of multi-electrode myoelectric sensing for prosthetic control applications, these strategies still do not produce robust graded signals for fine control. The use of implantable EMG and targeted muscle reinnervation strategies avoid some of the limitations of sEMG for prosthetic control, but these strategies are invasive and not well-suited for sensing muscle activity in various applications, such as stroke rehabilitation. Therefore, there continues to be a clear need for better non-invasive sensing of muscle activity. Recently, ultrasound imaging of muscle deformation has been shown as a possible alternative to sEMG for analyzing muscle activity. Ultrasound imaging enables the visualization of the cross-sectional anatomy of muscles and tendons. Real-time imaging can be used to track muscle contraction and relaxation. A significant benefit of ultrasound over sEMG is the ability to visualize deep musculature and the synergistic activity of different muscle compartments. This dissertation proposes a new strategy for sensing muscle activity based on real-time ultrasound imaging. The results verified that the ultrasound-based methodology is able to produce robust signals from contiguous functional compartments deep inside the muscle, a capability exceeding that of sEMG. Ultrasound imaging could potentially be attractive as a sensing strategy for upper extremity prosthetic control and as a muscle-computer interface (MCI) for rehabilitation robotics and exoskeletons.;For a practical ultrasound-based MCI that could be integrated into a compact wearable system, the use of a number of single element ultrasound transducers distributed around a region-of-interest is more practical than dense imaging array. Optimal channel selection is a common area of interest in brain-computer interface community and recently has been investigated for sEMG based MCIs in dense electrode setups. The main advantage of the channel optimization is the reduction in computation, reducing power consumption and improving system operation time. Among the different proposed algorithms, distance-based channel/feature subset selection (DFSS) and correlation-based channel/feature subset selection (CFSS) are commonly used to extract optimal channel/feature subsets for sEMG pattern recognition control. DFSS evaluates the class discrimination impact of each feature/channel using a distance measure such as Fisher's criterion (FC) while CFSS uses mutual information (MI) that each feature/channel represents about different classes. In this dissertation, I investigate the effect of ultrasound channel reduction on classification performance using distance-based channel subset selection (DCSS) and correlation-based channels subset selection (CCSS) where FC and MI are used as a measure of class discrimination, respectively. These two techniques are similar to the DFSS and CFSS except that they are feature selection independent. The results have shown that the number of channels can be reduced significantly (from 128 to 4) without sacrificing performance, measured by classification accuracy (CA). Furthermore, the selection of the discriminative spatial locations is highly subject specific which enables optimal system design on an individual basis.