Unsupervised Generative Adversarial Network for Plantar Pressure Image-to-Image Translation

Area of Science:

• Biomedical Engineering
• Computer Science
• Medical Imaging

Background:

• Plantar pressure analysis is crucial for assessing human health and identifying gait abnormalities.
• Existing methods primarily focus on classifying healthy versus unhealthy plantar pressure patterns.
• A gap exists in personalized generation of healthy gait patterns for individual patients.

Purpose of the Study:

• To develop an unsupervised generative adversarial network (GAN) for producing patient-specific healthy plantar pressure images.
• To address the challenge of training without paired healthy and unhealthy gait data.
• To enhance the accuracy and applicability of gait analysis through personalized image generation.

Main Methods:

• Utilized an encoder-decoder generator within a GAN framework.
• Employed a convolution-based discriminator and evaluation network.
• Introduced a novel loss function term to preserve individual gait style.
• Trained the model in an unsupervised generative adversarial learning setting.

Main Results:

• Achieved a high performance rate of 99.8% on the CAD WALK database.
• Successfully generated personalized healthy plantar pressure images.
• Demonstrated the effectiveness of the proposed unsupervised GAN approach.
• Validated the method on a dataset including patients with hallux valgus disease.

Conclusions:

• The proposed unsupervised GAN effectively generates patient-specific healthy plantar pressure images.
• This approach offers a significant advancement in personalized gait analysis and health assessment.
• The method shows promise for clinical applications in diagnosing and managing foot-related conditions.