Enhancing Burn Severity Assessment With Deep Learning: A Comparative Analysis and Computational Efficiency Evaluation

Burn injuries present a substantial global public health challenge. The conventional approach, relying on visual inspection to compute total body surface area (TBSA) for assessing burn severity, encounters the inherent limitations of proper estimation. These limitations prompted the development of computer-based applications, particularly machine learning, and deep learning models, to enhance the performance. This paper presents a comprehensive analytical study of eight deep learning techniques designed for assessing burn severity in terms of four characteristics: inflammation, scar, uniformity, and pigmentation, in small datasets of 2-dimensional (2D) images captured using digital (smartphone) camera. The models are Convolutional Neural Network (CNN), attention-based CNN, decision-level fusion (DF) based on CNN models, DF with attention-based CNN models, autoencoder-NN (Neural Network), and hybrid VGG16-Machine Learning (ML) with Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost). Each model is validated with eight datasets collected and annotated by our team at the Children’s Hospital of Michigan in two phases to classify the severity of burns in terms of inflammation, scar, uniformity, and pigmentation. The average test accuracy across the eight datasets using CNN, attention-based CNN, DF with CNN models, DF with attention-based CNN models, autoencoder-NN, VGG16-RF, VGG16-SVM, and VGG16-XGBoost are $0.87\pm 0.04$ , $0.93\pm 0.04$ , $0.90\pm 0.01$ , $0.95\pm 0.02$ , $0.87\pm 0.03$ , $0.63\pm 0.03$ , $0.79\pm 0.02$ , $0.79\pm 0.01$ , correspondingly. The research also computes and compares the computational complexity of each model in terms of FLoating point Operations Per Second (FLOPS) and Multiply-ACcumulate operations (MACs). Compared with the base CNN model, the decision-level fusion with attention mechanism model outperforms with a gain of 9.19% in test accuracy and an increase of 3321.53% in FLOPS. Considering the priority and constraint of the task, the attention-based CNN model can also be preferable as it achieves an accuracy gain of 6.90% and significantly less computational increase expense (8.62%) compared with the base CNN. The code for the best performing decision-level fusion with attention mechanism model is provided on GitHub link at https://github.com/Saeka2022/Burn-Assessment .