Abstract
BACKGROUND: As China's economy enters a new period of development, traditional Chinese culture is receiving more and more attention, and the status and role of traditional Chinese medicine (TCM) in economic and social development are also becoming more and more important, and TCM has become a unique health resource. In recent years, TCM has developed into a complete theoretical system and gradually spread abroad. Chinese herbal medicine (CHM) is an important part of TCM; however, its development and spread is seriously hampered by the fact that due to the wide variety and the large number of low-quality products in the market, professionals are required to identify TCMs accurately. Thus, a set of intelligent tools needs to be developed to classify CHMs. This study aimed to develop an intelligent classification tool for CHMs to address the challenges of accurately identifying and distinguishing between the vast variety of herbal products in the market. METHODS: Deep learning techniques have achieved great success in the field of image recognition. We selected the ResNet50 framework for the intelligent classification of CHMs. The incorporation of a residual structure in the network allowed us to construct deeper networks, while the feature extraction capability of the network was also greatly improved. In our approach, the output layer of the network was fine-tuned to adapt to our dataset, and the stochastic gradient descent optimizer was selected to update the network parameters. RESULTS: The accuracy of our method was verified on a constructed dataset of 24 classified CHMs, and the results showed that the recognition accuracy of our method was high across all categories; thus, our method could meet clinical needs. Overall, our method had a recognition accuracy of over 98% for the whole dataset, and more specifically, it had a recognition accuracy of 100% for 13 herbs. These results showed the efficiency and robustness of our method. It also had an accuracy rate of 100% for the private dataset for the sand nut category, which further shows the robustness of our method in classifying complex and diverse samples, and its ability to maintain a high level of accuracy even for complex herbs. CONCLUSIONS: Our deep learning-based method is able to accurately classify complex CHM properties. Due to its stability and robustness, our method is effective in different environments and conditions. It could help to promote the development of herbal medicine identification and chemical composition analysis, and improve traditional classification methods and personalized medicine.