TongueViT: A Vision Transformer-Based Framework for Automated Tongue Image Classification

Abstract

Tongue diagnosis is a serious non-invasive technique used in both traditional and modern medical diagnosis to examine the internal health status. Manual interpretation however is subjective and suffers inconsistencies. In this regard, paper introduce TongueViT, a new Vision Transformer (ViT)-based sys-tem of automated classification of tongue images. Compared to traditional CNN-based methods, TongueViT with the self-attention mechanism of transformers can exploit global contextual features to discriminate accurately between affected and normal tongue images. This model was trained and tested on a preprocessed dataset containing two categories of affected and healthy with stupendous classifi-cation accuracy of 99%. The results shown that TongueViT massively outcompetes the traditional models both in terms of performance and interpretability. Atten-tion maps also demonstrate that the model looks at clinically interesting areas of the tongue, making it more probable to be translated into real-world diagnostic assistance. These results confirm the usefulness of transformer-based models in biomedical image analysis and open the prospect of AI-aided tongue diagnostics in telemedicine and clinical screening.