Spatial attention-based residual network for human burn identification and classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Manual burn diagnosis is complex and time-consuming.
• Existing machine learning (ML) and deep learning (DL) models for burn diagnosis have limitations, including reliance on handcrafted features or challenges in designing robust models.
• Shallow DL methods often lack long-range feature dependency, impacting diagnostic efficiency.

Purpose of the Study:

• To develop and evaluate an attention-based deep convolutional neural network (CNN) model for accurate human burn diagnosis.
• To improve upon existing DL models by addressing feature dependency and enhancing classification accuracy for burn degree and depth.

Main Methods:

• Implemented and compared several deep CNN models (ResNeXt, VGG16, AlexNet) for burn diagnosis.
• Developed an attention-based model, BuRnGANeXt50, which divides feature maps into categories to highlight channel dependencies and uses a spatial attention map.
• Optimized BuRnGANeXt50's kernel and convolutional layers for burn diagnosis and evaluated its performance on the Burns_BIP_US_database.

Main Results:

• Previous shallow CNN models showed less reliable results, necessitating improved attention mechanisms.
• The proposed BuRnGANeXt50 model achieved high sensitivity: 97.22% for classifying burn degree and 99.14% for classifying burn depth.
• The model demonstrated improved feature dependency preservation through attention mechanisms.

Conclusions:

• The developed BuRnGANeXt50 model offers a reliable and efficient solution for human burn diagnosis.
• This model can facilitate rapid screening of burn patients and is deployable on cloud or local systems.
• The study emphasizes the importance of attention modules in deep learning for medical image analysis and provides reproducible code.