Deep learning-based artificial vision for grasp classification in myoelectric hands

Area of Science:

• Biomedical Engineering
• Computer Science
• Rehabilitation Technology

Background:

• Sensorimotor disorders significantly impact quality of life.
• Computer vision offers potential for advanced assistive technologies.
• Myoelectric prosthetic hands require intuitive control systems.

Purpose of the Study:

• To develop a computer vision system for trans-radial amputees.
• To enable grasping and manipulation of household objects using a myoelectric prosthetic hand.
• To improve the functionality of prosthetic hands through artificial vision.

Main Methods:

• A deep learning-based artificial vision system using a convolutional neural network (CNN) was developed.
• The CNN was trained to classify objects into four grasp patterns without explicit identification or dimension measurement.
• The system was tested offline and in real-time with novel and rotated objects, and finally with trans-radial amputee volunteers.

Main Results:

• Offline classification accuracy reached [Formula: see text] for seen and [Formula: see text] for novel objects.
• Real-time classification accuracy was [Formula: see text] for a set of novel and rotated objects.
• Trans-radial amputees achieved up to [Formula: see text] success rate in picking up and moving objects, with performance improving with training.

Conclusions:

• Deep learning-based computer vision significantly enhances myoelectric prosthetic hand grip functionality.
• This system offers a substantial conceptual improvement for controlling multi-functional prosthetic hands.
• The study demonstrates the feasibility of using artificial vision for intuitive prosthetic control in amputees.