Ensemble Learning for Burn Severity Classification

The classification of burns into 1st, 2nd, and 3rd degrees is a challenging task due to the subtle variations in texture, color, and depth, with overlapping features such as redness and swelling between 1st- and 2nd-degree burns, or blistering and charring between 2nd- and 3rd-degree burns. Our study proposes an enhanced framework called EffiBurn for the grading of burn severity, achieving state-of-the-art performance by incorporating EfficientNet, hyperparameter tuning, and Deep Convolutional Neural Networks. Preprocessing on the image dataset from the public repository involves contrast adjustment, random flipping, and rotation. Trained models were VGG16, VGG19, GoogleNet, MobileNetV2, and ResNet50 separately, showing suboptimal results. Model ensembling increased the classification accuracy to 80.52%, and upon further tuning, our proposed EfficientNet model, EffiBurn increased the accuracy by 82.49%, significantly higher than the baseline result for EfficientNetB0 at 38.46%. Critical factors in this improvement process were data cleaning and selecting optimal hyperparameters. The EffiBurn framework puts forth the possible relevance of optimized EfficientNet architecture for burn severity classification, thereby providing an efficient solution for burn management.