Spot Recognition in Tongue Images based on ResNet50

Abstract

Tongue diagnosis plays a crucial role in traditional Chinese medicine (TCM) by offering valuable insights for clinical syndrome differentiation and health assessment through tongue image analysis. Among these features, tongue spots, as a common abnormal manifestation, may indicate underlying pathological changes. Nevertheless, manual tongue diagnosis in TCM heavily relies on subjective visual observation, posing challenges in standardization. To tackle this issue, our study introduces an automated method for recognizing tongue spots using deep residual networks and transfer learning. Leveraging a real clinical dataset of 5,371 tongue images, we preprocessed the images, applied data augmentation, and fine-tuned a ResNet50 model pretrained on ImageNet for binary classification of tongue spot features. We employed five-fold cross-validation to assess the proposed approach. The results demonstrated that ResNet50 achieved an average accuracy of 97.48% (±0.53%) and a macro-F1 score of 53.84% (±6.64%), surpassing the ResNet18 baseline model. These findings suggest that deep residual learning effectively captures distinctive features from tongue images, showing promise for automated tongue feature recognition. Our study offers practical insights for the advancement of intelligent TCM-assisted diagnostic systems and may pave the way for further investigations into detailed tongue image analysis.