This paper presents a deep learning-based automatic skin disease classification model based on a dataset of over 50 diverse skin diseases captured using mobile devices. Unlike previous studies that primarily focused on microscopic image datasets and a limited number of disease types, this study utilizes a diverse skin disease dataset that reflects real-world conditions for a more realistic approach. After evaluating several CNN and Transformer-based architectures, we confirmed that Transformer models, such as the Swin Transformer, effectively capture global contextual information and demonstrate superior performance. Furthermore, we leverage Grad-CAM to enhance the interpretability of model predictions and ensure model transparency by visualizing clinically important regions. This paves the way for AI-based skin disease screening and early diagnosis accessible even in resource-poor settings.
