Classification Techniques of Tongue Manifestation Based on Deep Learning

With the ongoing advancement and refinement of medical theory, the exploration of Traditional Chinese Medicine (TCM) tongue inspection is gradually becoming more profound. However, the subjectivity inherent in tongue inspection leads to a lack of objectivity and standardization, significantly impeding its clinical application. This paper introduces the utilization of deep learning models for tongue inspection and presents a two-stage classification technique for tongue manifestations based on deep learning models. Initially, the UNet image segmentation model was employed to segment the original tongue images for more precise classification. Subsequently, the ResNet101 model was utilized to establish a deep learning-based lingual classification model. The outcomes indicate that this approach offers fast processing speeds and high classification accuracy, effectively supporting TCM tongue inspection. Recognizing the scarcity of accessible datasets, this study also developed a new tongue classification dataset by selecting expert TCM texts and employing web crawler technology to acquire tongue images and corresponding labels. Evaluation of ablation experiments reveals that training the model on the new tongue manifestation classification dataset significantly boosts performance compared to the original open-source dataset. When compared to the ResNet50 and VGG 19 network models, the ResNet101 model demonstrates superior training and testing outcomes. Furthermore, the segmented dataset outperforms the unsegmented dataset, achieving an accuracy and recall rate of 87.6% and 85.2%, respectively. The model and dataset utilized in this study are openly accessible, and the training and testing process were carried out on the domestic accelerated card light DCU Z100L.
¹