Integrating multi-source data for skin burn classification using deep learning.

Area of Science:

• Medical Artificial Intelligence
• Dermatology
• Computer Vision

Background:

• Accurate burn degree classification is vital for effective treatment and clinical decisions.
• Visual burn assessment is challenging for non-specialists, necessitating AI-assisted tools.
• Existing AI models for burns face limitations in data diversity, bias, and standardization.

Purpose of the Study:

• To develop an AI-based system for classifying burn degrees and determining grafting requirements.
• To address limitations in current AI burn assessment tools through improved datasets and methodologies.
• To enhance clinical decision support for burn management.

Main Methods:

• Developed a robust data pipeline, enhancing datasets with diverse, annotated images, including specific focus on Egyptian skin tones.
• Employed deep learning models (ResNet50, DenseNet, MobileNet, VGG16, ShuffleNet) for image classification.
• Utilized a cascading classifier for burn degree assessment and binary classification for grafting determination.

Main Results:

• A modified ResNet50 model achieved 94.03% accuracy and 0.94 F1 score for grafting classification, surpassing other models.
• The cascading classifier achieved 63.23% accuracy and 0.63 F1 score for burn degree classification.
• Demonstrated the effectiveness of deep learning with diverse, curated datasets across varied clinical settings.

Conclusions:

• AI models show significant potential for skin burn classification and clinical decision support.
• The developed approach, incorporating diverse datasets and novel methods, advances standardized AI for burn assessment.
• Deep learning models, when trained on high-quality, diverse data, can lead to more reliable burn assessment tools.