Wound Tissue Classification: A Comparative Analysis of Deep Neural Network Models

Wound tissue classification is an important task in medical imaging, with applications ranging from wound assessment to treatment planning. In this study, we investigated different neural network architectures and loss functions to improve the accuracy and efficiency of wound tissue classification. The study included eight different neural network architectures, including classic U-Net, MobileNet U-Net, Attention U-Net, Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net), Attention Recurrent Residual U-Net (R2AU-Net), Residual Network (ResNet-50), EfficientNet and SegFormer. Each architecture was trained separately with four different loss functions: categorical cross-entropy, weighted categorical cross-entropy, focal loss and soft dice loss. The comparative analysis revealed that the SegFormer architecture in conjunction with soft dice loss function achieved the most promising results on all classification metrics. The results of the study highlight the potential for further research in this area and emphasise the need for more comprehensive datasets to improve model performance.