Integrating multi-source data for skin burn classification using deep learning

Comprehensive Summary

Determining the degree of skin burns is difficult for both general practitioners and specialists. This study aimed to utilize AI models in the classification of skin burns using high-quality datasets. Experts first divided skin burns into 4 classes (I, IIa, IIb, III). These images were then preprocessed using ROI extraction, conversion to the CIELAB colour space, and image enhancement to improve quality and reduce noise. These preprocessed images, with their clinical classifications, were used to train the deep learning models ResNet50, DenseNet, MobileNet, VGG16, and ShuffleNet, to learn patterns associated with each class. The models' performance was then tested on new images. Two approaches were compared: multiclass classification (distinguish between all four classes) and binary classification (distinguish between Class I vs others and Class II vs Class III). The binary classification (accuracy 94.03%, F1 score 0.9384), performed better than the multiclass classification (accuracy 63.23%, F1-score of 0.6271).

Outcomes and Implications

While larger and more diverse datasets are needed, this study demonstrates the first steps towards developing a standardized AI system for skin burn assessment. This system would make diagnosis of burns faster and more accessible for practitioners while improving the overall patient experience.