Uniform Selection And Representation Matching: A Framework For Classifying Wound Healing Stage.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Wound healing research

Background:

• Accurate wound stage classification is challenging due to under-specified descriptors and limited, noisy labeled data.
• Existing methods struggle with the complexities of wound image analysis, hindering clinical applications.
• The need for robust algorithms to monitor wound healing remotely and support treatment recommendations is critical.

Purpose of the Study:

• To develop an accurate wound stage classification framework using limited and noisy-labeled data.
• To improve the performance of image classification for wound healing assessment.
• To enable applications such as remote wound monitoring and intelligent bandage devices.

Main Methods:

• Proposed the Uniform Selection and Representation Matching (USRM) framework.
• Integrated co-teaching, contrastive learning, representation matching, and uniform selection techniques.
• Utilized an entropy-based selection process to identify low-confidence images and assigned pseudo-labels via representation matching in latent space.

Main Results:

• Achieved a classification accuracy of 90.0% for wound-stage images.
• Demonstrated significant improvement over conventional convolutional neural networks.
• Successfully classified wound stages even with minimal and noisy training data.

Conclusions:

• The USRM framework offers a robust solution for wound stage classification with limited and noisy data.
• This algorithm has potential applications in remote wound healing monitoring, treatment recommendation, and smart medical devices.
• The findings pave the way for more intelligent wound care solutions.