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

BACKGROUND

Skin burns result from thermal or chemical damage to the skin, requiring timely and accurate assessment for effective treatment. Determining the degree of burns is crucial for appropriate clinical decisions, especially for interventions like grafting. However, visual burn degree classification is challenging for non-specialists, underscoring the need for Artificial Intelligence(AI)-powered tools to assist in burn assessment. Current AI models face challenges related to biases, validation, and limited data availability, and there is no standardized system for skin burn classification.

METHOD

This study develops an AI-based approach to classify burn images into three degrees for initial assessment recommendations and employs binary classification for grafting determination. The methodology includes a robust data collection and preprocessing pipeline, which enhances existing datasets by filtering publicly available labeled data, annotating unlabeled data, and collaborating with specialists for local data annotation. A diverse, high-quality dataset is built by integrating images from various sources to ensure generalization across populations, with a particular focus on Egyptian skin tones. Deep learning models, including ResNet50, DenseNet, MobileNet, VGG16, and ShuffleNet, are employed to classify burn images. A cascading classifier approach is used for burn degree classification, which is divided into two stages: first-degree vs. others and second-degree vs. third-degree.

RESULTS

The highest performance is achieved using a modified ResNet50 model, which attains an accuracy of 94.03% and an F1 score of 0.94 for grafting classification, outperforming state-of-the-art models by at least 5% during cross-validation. For the burn degree classification task, the cascading classifier approach yields an accuracy of 63.23% and an F1 score of 0.63. These results demonstrate the effectiveness of deep learning models when supported by a carefully curated, diverse dataset, specifically that the model was tested using multiple source data from different clinical settings.

CONCLUSIONS

This study highlights the potential of AI models in skin burn classification and clinical decision support, particularly with diverse, high-quality datasets. Our approach, which addresses dataset challenges and incorporates novel methods like cascading classifiers and grafting classification, represents a significant step toward developing a standardized AI system for skin burn assessment. The findings show that deep learning models can significantly improve classification performance, paving the way for more reliable, clinically applicable burn assessment tools.