Unsupervised Anomaly Detection in Tongue Diagnosis with Semantic Guided Denoising Diffusion Models

Tongue diagnosis is one of the core diagnostic methods in Traditional Chinese Medicine (TCM), primarily involving the visual inspection of tongue images to assess a patient's health status. However, the subjectivity and environmental differences in tongue diagnosis may lead to potential errors and limitations. In this paper, we introduce an unsupervised tongue coating anomaly detection model based on diffusion models, aiming to address the limitations of traditional supervised learning and existing anomaly detection models. Our approach combines the semantic classification ability of the cross-attention module within the diffusion model with score-based conditional guidance to achieve high-quality image reconstruction and precise identification of discriminative regions. Experimental results have demonstrated that our anomaly detection model exhibits state-of-the-art performance, surpassing the accuracy of existing models.