Burn image segmentation based on Mask Regions with Convolutional Neural Network deep learning framework: more...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate burn area and depth assessment are critical for effective treatment and patient outcomes.
• Traditional burn diagnosis methods lack repeatability and comparability, leading to diagnostic inconsistencies.
• Deep learning offers a potential solution for semi-automating burn diagnosis and reducing human error.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automated burn wound segmentation.
• To improve the accuracy and reliability of burn diagnosis compared to existing methods.
• To facilitate more precise calculations of total body surface area (TBSA) burned.

Main Methods:

• A deep learning segmentation framework based on Mask Regions with Convolutional Neural Network (Mask R-CNN) was designed.
• The model was trained on 1000 labeled images and evaluated on 150 images, with comparisons of different backbone networks.
• Performance was assessed using the Dice coefficient (DC), comparing Residual Network-101 with Atrous Convolution in Feature Pyramid Network (R101FA), Residual Network-101 with Atrous Convolution (R101A), and InceptionV2-Residual Network with Atrous Convolution (IV2RA).

Main Results:

• The R101FA backbone network achieved the highest accuracy of 84.51% in burn wound segmentation.
• R101FA demonstrated superior segmentation performance across superficial, superficial partial-thickness, and deep partial-thickness burns.
• R101A showed the best segmentation results for full-thickness burns.

Conclusions:

• The proposed deep learning framework provides accurate and robust burn wound segmentation across various burn depths.
• This AI-driven approach is more convenient and suitable for clinical application than traditional methods.
• The framework aids in more accurate total body surface area (TBSA) calculations for burn patients.