Autonomous Multi-modality Burn Wound Characterization using Artificial Intelligence.

Area of Science:

• Medical Artificial Intelligence
• Biomedical Imaging
• Computational Pathology

Background:

• Burn wounds account for 5-20% of combat casualties, with early treatment improving survival rates by 36%.
• Accurate burn injury classification is critical for effective treatment and mortality reduction.
• Current methods for burn characterization can be improved through advanced computational approaches.

Purpose of the Study:

• To develop and validate an autonomous artificial intelligence (AI) system for precise burn wound characterization.
• To investigate the utility of B-mode ultrasound and RGB imaging for burn depth classification.
• To enhance AI model performance by integrating explainable AI (XAI) with expert knowledge.

Main Methods:

• A two-part dataset comprising 10,085 B-mode ultrasound frames from porcine subjects and 338 RGB images was utilized.
• An AI framework incorporating an explanation system was developed to integrate expert knowledge and validate model features.
• Statistical texture features were extracted from ultrasound frames to improve classifier accuracy.

Main Results:

• The AI system achieved >80% accuracy and F1 average for burn depth classification.
• The segmentation module demonstrated a mean global accuracy >84% and a mean intersection-over-union score >0.74.
• Textural features from ultrasound data were confirmed to enhance burn depth classification accuracy.

Conclusions:

• This study demonstrates the feasibility of automated burn characterization using AI.
• Combining human expertise with explainable AI can significantly improve AI system performance in burn analysis.
• The developed framework provides a foundation for future deep learning applications in burn wound assessment.