Classification of Therapeutic Hand Poses Using Convolutional Neural Networks

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Computer Vision

Background:

• Accurate measurement of finger active range of motion (ARoM) is crucial for assessing outcomes in hand surgery and rehabilitation.
• Current goniometer-based ARoM measurement is prone to error and is time-consuming.

Purpose of the Study:

• To develop an automated system for ARoM measurement, enhancing speed and accuracy.
• To design and evaluate a hand pose classifier for automatic ARoM measurement system integration.

Main Methods:

• Investigated and compared various classifiers for automatic hand pose detection.
• Integrated the best-performing classifier into an automated ARoM measurement system.
• Evaluated system performance on unseen subjects and diverse hand poses.

Main Results:

• The optimal classifier achieved 99% accuracy in classifying hand poses across 400 test samples.
• The system demonstrated a rapid compute time of 8ms per sample.
• The automated system enables dynamic camera control for optimal joint visualization.

Conclusions:

• An accurate and efficient automated system for finger ARoM measurement has been developed.
• The proposed hand pose classification method significantly improves measurement accuracy and reduces computation time.
• This technology has the potential to revolutionize hand rehabilitation assessment.