Optimizing CNN for burn severity classification using hyperband

Abstract

Burn injuries require accurate severity diagnosis to determine appropriate treatment and reduce mortality. However, conventional clinical assessment is susceptible to subjectivity and diagnostic errors. This study aims to optimize a Convolutional Neural Network (CNN) model for classifying burn severity using the Hyperband hyperparameter optimization algorithm. A dataset of 1,040 burn images, collected from primary and secondary sources, was categorized as first-, second-, or third-degree burns. A MobileNet-based CNN model was trained using both default and Hyperband-optimized hyperparameters and evaluated using 10-fold cross-validation. The results demonstrate that hyperparameter optimization substantially improved classification performance. The optimized model achieved an accuracy of 93.57%, compared with 82.14% for the baseline model. Improvements were also observed in precision, recall, and F1-score, indicating better classification performance and reduced prediction errors. These findings demonstrate that Hyperband can effectively optimize CNN hyperparameters for burn severity classification. The proposed approach has the potential to support more consistent and accurate burn assessment. Future research should investigate advanced optimization strategies and larger, more diverse datasets to improve model generalization and robustness further.