BPBSAM: Body part-specific burn severity assessment model.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Computational pathology

Background:

• Automated burn diagnosis remains a significant challenge despite advancements in science and technology.
• Burns pose a serious health threat, causing thousands of deaths annually.
• Existing automated approaches often struggle with the nuances of burn severity assessment.

Purpose of the Study:

• To develop a body part-specific burns severity assessment model (BPBSAM) using deep convolutional neural networks.
• To address the challenge of limited availability of labeled burn images.
• To improve the accuracy and reliability of automated burn diagnosis.

Main Methods:

• Utilized a deep convolutional neural network (CNN) for feature extraction from burn images.
• Employed body part-specific support vector machines (SVMs) for burn severity estimation.
• Leveraged larger datasets of non-burn images for pre-training CNNs to overcome data scarcity.

Main Results:

• The BPBSAM achieved an overall average F1 score of 77.8% and accuracy of 84.85% on the BI dataset.
• Achieved higher accuracy on the UBI dataset with an average F1 score of 87.2% and accuracy of 91.53%.
• Demonstrated superior performance over generic methods, with average accuracy improvements of up to 10.61%.

Conclusions:

• The proposed body part-specific model significantly enhances burn severity assessment performance, particularly with limited datasets.
• The customized approach shows potential for addressing burn region segmentation challenges.
• Fine-tuning pre-trained networks on non-burn body part images proved to be a robust and reliable strategy.