Pediatric BurnNet: Robust multi-class segmentation and severity recognition under real-world imaging conditions

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pediatric Burn Care

Background:

• Accurate burn wound assessment is crucial for effective treatment.
• Current methods for burn depth grading can be subjective and time-consuming.
• The need for objective, rapid assessment tools in diverse clinical settings is increasing.

Purpose of the Study:

• To develop and validate a deep learning model for simultaneous segmentation and depth grading of pediatric burn wounds.
• To ensure the model performs accurately under complex, real-world imaging conditions.
• To assess the model's potential for deployment on mobile devices for clinical use.

Main Methods:

• Retrospective collection of 4785 smartphone/camera images of pediatric burns over 5 years.
• Annotation of 14,355 burn regions across three depth categories (superficial second-degree, deep second-degree, third-degree).
• Development of an attention-enhanced DeepLabv3-ResNet101 model (Dfusion) with specific architectural improvements and weighted cross-entropy loss, validated using ten-fold cross-validation.

Main Results:

• The Dfusion model achieved a mean Dice coefficient of 0.8766 and intersection-over-union of 0.8052 for segmentation.
• Classification accuracy reached 97.65%, with an F1-score of 85.33% for burn depth grading.
• Dfusion significantly outperformed baseline models (e.g., DeepLabv3, U-Net-ResNet101) in Dice and IoU metrics, with fast inference speeds suitable for mobile deployment.

Conclusions:

• The Dfusion model offers accurate, end-to-end segmentation and depth grading of pediatric burn wounds in uncontrolled environments.
• Its performance and computational efficiency support mobile device deployment, providing rapid, objective clinical assistance.
• This technology can enhance triage and treatment planning in pediatric burn care, especially in resource-limited settings.