A novel framework for rosacea detection using Swin Transformers and explainable artificial intelligence

Accurate and efficient diagnosis of skin rosacea is crucial in dermatological healthcare, yet remains challenging due to the need for precise classification and interpretability. This study aims to develop a robust and explainable approach for skin rosacea classification using advanced deep learning and explainable artificial intelligence (XAI) techniques. Our innovation lies in combining Swin Transformer models (Swin-Tiny, Swin-Small, and Swin-Base) with pre-trained convolutional neural network (CNN) architectures such as VGG19, ResNet50, and DenseNet201, addressing the limitations of traditional CNN-based methods. The proposed Swin-Base transformer achieves a state-of-the-art accuracy of 99%, outperforming all tested models across multiple evaluation metrics, including accuracy, precision, recall, F1 score, specificity, and others, on a dataset of 10,539 images. To enhance interpretability, we employ XAI techniques, including Score-CAM, GradCAM, and GradCAM++, which provide visual explanations and insights into the decision-making process. These findings underscore the potential of integrating cutting-edge transformer architectures with explainability techniques to build reliable, interpretable, and high-performing models for clinical dermatology, ultimately improving patient care.