Convolutional Neural Networks in Chronic Wound Segmentation and Tissue Classification Using Real-World Images

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Wound care technology

Background:

• Chronic wounds represent a significant global health burden.
• Objective diagnostic and monitoring tools are crucial for effective wound management.
• Artificial intelligence presents a promising avenue for advancing wound care.

Purpose of the Study:

• To train a convolutional neural network (CNN) for automatic segmentation of wound area and tissues from real-world clinical images.
• To evaluate the performance of the AI model in identifying different wound components.

Main Methods:

• Utilized a U-Net convolutional neural network architecture with fully supervised learning.
• Trained the model on 362 real-world images of chronic wounds (venous, arterial, vasculitis, pyoderma gangrenosum).
• Employed data augmentation and pretraining to enhance model performance.

Main Results:

• Achieved high accuracy in wound area segmentation (Dice Similarity Coefficient [DSC] = 0.927, Intersection over Union [IoU] = 0.868).
• Demonstrated fair performance for identifying fibrinous exudate (DSC = 0.750, IoU = 0.659) and granulation tissue (DSC = 0.696, IoU = 0.601).
• Lower performance was observed for necrosis segmentation (DSC = 0.503, IoU = 0.502), likely due to limited image data.

Conclusions:

• A neural network can be effectively trained to analyze clinical wound images for segmentation tasks.
• AI models can identify various wound structures beyond just the total wound area.
• Performance in identifying specific wound structures is contingent on the quantity of training data available for each structure.