Online prediction of the hand configuration using ultrasound images of the forearm

Abstract

Previous studies have demonstrated a linear relationship between first degree features extracted from ultrasound images taken from the forearm and the angles at which the fingers of the hand are. Using this information it is possible to train a system that in real time predicts the position of the fingers while receiving a feed of ultrasound images of the forearm. With the target of reaching a robust system capable of running without without decay in the performance overtime several approaches were taken. A more dynamic training system was implemented using incremental linear regression. This allows the patient to switch back and forth from the training part of the experiment to the prediction phase. While predicting, a 3D hand model that moves based exclusively on the information extracted from the ultrasound images is displayed on the screen.