Deep Learning Based Spatial Source Localization Using High-Density Surface EMG Signals

The identification of the spatial position of activated motor units (MUs) in real-time is crucial in understanding and optimizing muscle function during various activities, which contributes to advancements in various fields such as sports science, biomechanics, and neurological rehabilitation. In this study, we attempt to improve upon an existing ‘Curve Fitting Based Minimum Norm Estimation’ (CFB-MNE) approach, which uses signal processing techniques to localize MUs, with attempts to enhance its efficiency and overall capability by employing Deep Learning (DL) to solve the localization problem. The performance adequacy of several DL Convolutional Neural Network (CNN) models was explored, with their performances regarding spatial localization compared and analyzed. A dataset of high-density surface electromyography (HD-sEMG) signals, closely resembling those acquired from Biceps Brachii (BB) muscles, was generated using a multilayered volume conductor generation model. Inverse solutions were obtained then plotted as a 3D curve to finally extract the 2D images used to train the models. Testing on unseen data yielded excellent results, with all performance metrics exceeding 90%, indicating the reliability of the DL-based models in real-time spatial localization of MUs. These promising outcomes not only demonstrate the effectiveness of our current approach in accurately identifying the spatial position of individual MUs in real-time, but they also pave the way for additional improvements exploiting the abilities of more advanced DL models.