MCTFWC: a multiscale CNN-transformer fusion based model for wound image classification

Chronic wounds affects millions of individuals globally, creating substantial challenges for healthcare systems due to high medical costs and the complexity of wound management. Accurate classification of wound types is crucial for effective diagnosis and treatment planning, enabling clinicians to determine optimal care procedures efficiently. However, most existing deep learning methods focus on binary classification and struggle with multiclass wound classification, particularly in distinguishing complex wound types such as pressure and diabetic wounds. This study proposes a novel dual-branch architecture that combines a Convolutional Neural Network (CNN) and a Transformer to enhance wound classification accuracy. The CNN branch, based on a ResNet backbone, captures local features, while the Transformer branch, structured with a pyramidal multi-scale design, encodes global contextual information at multiple resolutions. A Feature Fusion Attention Block (FFAB) at each encoding level enhances feature integration through spatial and channel attention mechanisms. The proposed model achieves superior performance in both binary and multiclass wound classification, with average accuracies of 91.20% and 89.80%, respectively, outperforming state-of-the-art methods. Ablation studies confirm the effectiveness of key components, including the FFAB and parallelized CNN-Transformer structure. This work demonstrates the potential of the proposed approach to significantly improve diagnostic accuracy and efficiency, offering a robust solution for wound image classification in intelligent medical systems.