Deep Learning-Assisted Burn Wound Diagnosis: Diagnostic Model Development Study.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Burn Management

Background:

• Accurate percentage total body surface area (%TBSA) estimation is critical for burn patient management, influencing fluid resuscitation, nutrition, and mortality prediction.
• Estimating irregular burn areas by visual inspection is challenging and prone to discrepancies among clinicians.
• Existing methods for %TBSA assessment often lack precision, necessitating improved diagnostic tools.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for precise burn wound detection, segmentation, and %TBSA calculation.
• To leverage artificial intelligence for pixel-to-pixel accurate assessment of burn wound surface area.
• To provide a more objective and reliable tool for burn size estimation in clinical practice.

Main Methods:

• A two-step deep learning approach using U-Net and Mask R-CNN architectures with different backbones (ResNet101).
• Burn wound images were segmented, and hand images were used as a reference (1 hand = 0.8% TBSA) for %TBSA calculation.
• The models were trained and validated on datasets of labeled burn wound and hand images.

Main Results:

• Mask R-CNN with ResNet101 achieved the highest segmentation accuracy for burn wounds (Dice coefficient of 0.9496).
• Both U-Net and Mask R-CNN demonstrated high accuracy in segmenting hand images (Dice coefficients ~0.99).
• The Mask R-CNN model showed a smaller average deviation (0.115% TBSA) from the ground truth compared to burn surgeons in a test diagnosis.

Conclusions:

• This study presents a novel deep learning application for diagnosing all burn depths and calculating %TBSA.
• The developed AI models offer a more accurate and consistent method for estimating burn size.
• The findings aim to assist healthcare professionals in providing more precise and effective care for burn victims.