Generating a novel synthetic dataset for rehabilitation exercises using pose-guided conditioned diffusion models: A...

Area of Science:

• Biomedical Engineering
• Computer Science
• Rehabilitation Science

Background:

• Machine learning (ML) offers personalized insights for rehabilitation therapy monitoring and evaluation.
• A significant challenge in developing robust ML models for rehabilitation is the scarcity of data.

Purpose of the Study:

• To introduce a novel synthetic dataset for rehabilitation exercises using pose-guided person image generation.
• To address the data scarcity issue in machine learning for physical therapy.

Main Methods:

• Leveraged conditioned diffusion models for pose-guided person image generation.
• Processed a pre-labeled dataset of 6 rehabilitation exercises to generate realistic human movement images.
• Generated 22,352 images, enhancing variability in subject attributes and background environments.

Main Results:

• Generated images accurately captured spatial consistency of human joint relationships.
• Achieved highly favorable quantitative metrics for image assessment.
• Maintained excellent intra-class and inter-class consistency in motion data (distance correlation > 0.90).

Conclusions:

• The novel synthetic dataset effectively augments limited real-world data for ML in rehabilitation.
• This approach enhances the value of existing datasets by generating high-fidelity synthetic images.
• Facilitates improved ML model development for personalized rehabilitation therapy.