Enhancing chronic wound assessment through agreement analysis and tissue segmentation

Area of Science:

• Medical imaging analysis
• Computational pathology
• Artificial intelligence in healthcare

Background:

• Manual chronic wound assessment is subjective and time-consuming.
• Accurate tissue segmentation is vital for monitoring wound healing dynamics.
• Existing methods lack consistency and efficiency.

Purpose of the Study:

• To develop and evaluate an automated deep learning framework for chronic wound tissue segmentation and quantification.
• To compare the performance of convolutional neural network (CNN) and transformer-based models.
• To assess the impact of transfer learning and post-processing on segmentation accuracy.

Main Methods:

• Explored DeepLabV3-R50 (CNN) and SegFormer-B0 (transformer) for tissue segmentation.
• Investigated transfer learning from existing open wound datasets.
• Integrated segmentation models into a framework with wound and marker detection.
• Applied post-processing techniques to refine segmentation masks.

Main Results:

• DeepLabV3-R50 achieved a mean Intersection over Union (IoU) of 62.95% and Dice score of 76.82% for independent tissue segmentation.
• The complete framework with DeepLabV3-R50 yielded a mean IoU of 59.67% and Dice score of 74.38%.
• Achieved mean absolute errors of 14.33% (granulation), 14.31% (slough), and 8.84% (eschar) for tissue proportion estimation.

Conclusions:

• Deep learning, particularly DeepLabV3-R50 with transfer learning, offers a promising automated solution for wound tissue segmentation.
• The proposed framework enhances consistency and efficiency in wound bed characterization.
• Results highlight the potential to streamline chronic wound monitoring and improve clinical decision-making.