Burn Images Segmentation Based on Burn-GAN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Burn injuries necessitate accurate wound assessment for effective treatment planning.
• Automated wound segmentation using deep learning shows promise but is limited by data scarcity.
• Precise calculation of total burn surface area (%TBSA) is crucial for clinical decision-making.

Purpose of the Study:

• To develop an automated framework for generating annotated burn wound image datasets.
• To enhance the accuracy and efficiency of burn wound segmentation models.
• To overcome the limitations of data collection for training deep learning models.

Main Methods:

• Proposed Burn-GAN framework utilizes Style-GAN for wound generation.
• Employs Color Adjusted Seamless Cloning (CASC) for realistic wound-skin fusion.
• Simulates 3D burn scenes and derives annotations via coordinate transformation.
• Incorporates nonsaturating loss with R2 regularization (NSLR2) for training.

Main Results:

• The Burn-GAN framework successfully generated diverse, annotated burn image datasets.
• Segmentation network achieved 90.75% precision and 96.88% Pixel Accuracy (PA).
• Dice Coefficient (DC) improved from 84.5% to 89.3% with the generated data.
• The framework demonstrated significant improvements in segmentation accuracy and efficiency.

Conclusions:

• The developed Burn-GAN framework effectively addresses the challenge of limited training data for burn wound segmentation.
• Automated data generation significantly improves segmentation performance, aiding in accurate %TBSA calculation.
• This approach offers a time-efficient and accurate solution for clinical applications in burn management.