TL-BSLD: a transfer learning-based multi-convolutional neural network framework for burn severity level detection

Background: Burn injuries continue to pose a major global public health challenge, contributing to severe physical trauma and a high number of fatalities each year. Prompt and accurate assessment of burn severity is crucial for determining the appropriate course of treatment and improving patient outcomes. However, traditional diagnostic methods primarily depend on visual inspections by clinicians, which are inherently subjective and susceptible to errors. These limitations often lead to delayed or incorrect diagnoses. To overcome these challenges, the need for an automated, objective, and accurate diagnostic system has become increasingly evident in the field of burn care. Finally, the objective of this study is to develop and evaluate an automated framework for accurate burn severity classification using transfer learning (TL). We further aim to integrate interpretability tools to enhance clinical trust and applicability. Methods: This study introduces the Transfer Learning-Based Burn Severity Level Detection System (TL-BSLD), an innovative framework that applies advanced computer vision and TL techniques to accurately classify the severity of burn injuries. The system utilizes three cutting-edge pre-trained models—VGG-19 architecture, Inception-v3, and NasNet—to analyze burn images and categorize them into four classes: normal, first-degree, second-degree, and third-degree burns. TL enhances the model’s ability to extract relevant features from limited medical data, improving classification performance. Moreover, to ensure transparency and clinical trust, the system incorporates two interpretability tools: Gradient-weighted Class Activation Mapping (Grad-CAM), which highlights important regions in the input image influencing predictions, and SHapley Additive exPlanations (SHAP), which quantifies the contribution of each feature to the model’s decision. Results: The proposed TL-BSLD system underwent extensive evaluation using standard performance metrics. Among the tested models, VGG-19 demonstrated superior performance, achieving an accuracy of 98.3% and an F1-score of 98.5%, indicating the system’s reliability and robustness in classifying burn severity levels. The integration of Grad-CAM and SHAP also provided valuable visual and feature-based insights into the model’s predictions, enhancing its interpretability and clinical applicability. Conclusions: The TL-BSLD presents a significant advancement in automated burn diagnosis by combining high accuracy, rapid image classification, and interpretability. This approach not only improves diagnostic efficiency but also empowers clinicians with deeper insights into artificial intelligence (AI) decision-making. By reducing reliance on subjective assessments and minimizing diagnostic delays, the system demonstrates promising potential to enhance burn care and decision support; however, further validation in larger, prospective clinical studies is necessary before clinical deployment. Future work may explore deployment in real-time clinical settings and adaptation to mobile health platforms for broader accessibility.