Tongue image feature classification and gastric disease diagnosis using deep learning in traditional chinese medicine

The assessment of gastrointestinal health through tongue image analysis is a significant aspect of traditional Chinese medicine. Utilizing computer vision technology for the analysis of tongue image features and disease diagnosis has emerged as a focal point in medical image processing research. However, the current integration of deep learning with traditional Chinese medicine remains relatively limited, particularly in the comprehensive exploration of tongue image features based on traditional Chinese medical diagnostic theories. In this study, a variety of deep learning models were employed to perform classification tasks on the presence of common tongue features such as thick coating, cracks, tooth marks, and the existence of gastric diseases. The deep learning models utilized include CNN, ResNet, AlexNet, and DenseNet. Subsequently, DenseNet was used as the reference model to evaluate the performance of pre-training with the three tongue image features for gastric disease classification. The training and validation were conducted on tongue image datasets collected and annotated at the Department of Traditional Chinese Medicine of Peking University Third Hospital. Experimental results demonstrate that DenseNet achieved an AUROC value of 0.9207 for certain tongue image features. Different networks exhibited favorable performance in metrics such as Accuracy, Precision, and Recall. Moreover, the injection of the three tongue image features as prior information significantly enhanced the model's accuracy in identifying gastric diseases. Our research validates the feasibility of deep learning in intelligent tongue image diagnosis, laying a foundation for the digitization and intelligence of traditional Chinese medicine tongue diagnosis.