Scalable and objective wound infection screening from clinical images using deep learning

Background: Wound infection is a common and clinically significant complication that can delay healing, increase healthcare costs, and contribute to inappropriate antimicrobial use. Rapid, objective, and scalable screening tools are urgently needed, particularly in resource-limited or non-specialist clinical settings. This study aimed to develop and evaluate a deep learning-based framework for automated wound infection detection using clinical wound images, with a focus on improving diagnostic consistency and supporting public health-oriented wound management. Methods: A dataset of 4,000 diverse clinical wound images was used to train and evaluate multiple deep learning models. The Swin Transformer architecture was compared with conventional convolutional neural networks. Model performance was assessed using accuracy, area under the receiver operating characteristic curve, and F1-score. To evaluate real-world applicability, model predictions were further compared with assessments made by non-specialist clinicians. Results: The Swin Transformer outperformed conventional convolutional neural networks, achieving an accuracy of 0.9025 (95% CI: 0.8695-0.9279), an area under the receiver operating characteristic curve of 0.9546, and an F1-score of 0.9042. Compared with non-specialist clinicians, the model reduced diagnostic variability and enabled earlier and more consistent recognition of wound infections. Conclusion: Deep learning applied to clinical wound images provides a scalable and objective approach for wound infection screening. Such tools have the potential to support earlier detection, reduce diagnostic variability, and improve wound management and antimicrobial stewardship, particularly in public health and resource-limited settings.