Fully Automated Wound Tissue Segmentation Using Deep Learning on Mobile Devices: Cohort Study.

Area of Science:

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

Background:

• Wound tissue composition is crucial for assessing healing progression and guiding treatment.
• Current clinical methods for wound tissue identification are subjective, leading to potential inaccuracies.
• Variability in assessment can result in inappropriate wound care decisions, such as incorrect dressing selection or delayed specialist referrals.

Purpose of the Study:

• To quantify inter- and intrarater variability in manual wound tissue segmentation among clinicians.
• To develop and evaluate deep neural networks for objective wound tissue identification and quantification.

Main Methods:

• A dataset of 58 chronic wound images was used for manual segmentation by 5 clinicians.
• Four tissue types (epithelial, granulation, slough, eschar) were manually labeled at 1-week intervals.
• Two deep convolutional neural network architectures were developed and trained on a large dataset (465,187 and 17,000 image-label pairs) for segmentation tasks.

Main Results:

• Poor to moderate interrater agreement was observed, with significant variability in identifying specific tissue types (e.g., epithelization).
• Deep learning models achieved high performance with mean intersection over union scores of 0.8644 for wound segmentation and 0.7192 for tissue segmentation.
• Clinicians rated 91% of the deep learning model's tissue segmentation results as fair to good.

Conclusions:

• Manual assessment of wound tissue composition by clinicians demonstrates considerable variability.
• Deep learning techniques offer objective and accurate tissue identification and measurements for wound documentation.
• Objective wound assessment using AI has the potential to significantly improve wound care when implemented broadly.