Classification of Therapeutic Hand Poses Using Convolutional Neural Networks

Measurement of finger active range of motion (ARoM) is essential to quantify outcomes accurately after hand surgery and during rehabilitation. Currently, finger ARoM is measured by a hand-held goniometer, which introduces measurement error. Moreover, this method is time-consuming. To speed up and simplify this process, we developed a system to measure the ARoM automatically. However, to assess the ARoM for all joints, different hand poses are required. The goal, then, is to design a classifier that achieves accurate and automatic discovery of the hand pose. According to the detected pose, the system will apply the appropriate algorithm to measure the ARoM for all fingers. Furthermore, this will enable a camera capture control system to provide the best view by moving the camera as required by each algorithm. A critical part of the system is the classifier because it controls the accuracy and compute time of the measurement. In this paper, we describe a study of different classifiers for hand pose and include results. The best classifier achieves 99% accuracy in classifying 400 test samples from five previously unseen human subjects with a compute time of 8ms per sample.