基于深度学习的阳虚质与阴虚质舌象分类研究  

Study of Tongue Image Classification for Yang Deficiency and Yin Deficiency Constitutions Based on Deep Learning

作　　者：董易杭 王建勋 王晶 赵梅 周恒宇 王林雁 张晓晴 DONG Yihang;WANG Jianxun;WANG Jing;ZHAO Mei;ZHOU Hengyu;WANG Linyan;ZHANG Xiaoqing(School of Life Science,Beijing University of Chinese Medicine,Beijing 100029,China;Shanghai Daosheng Medical Technology Co.,LTD.,Shanghai 201203,China)

机构地区：[1]北京中医药大学生命科学学院,北京100029 [2]上海道生医疗科技有限公司,上海201203

出　　处：《中华中医药学刊》2024年第7期27-30,I0003,I0004,共6页Chinese Archives of Traditional Chinese Medicine

基　　金：国家重点研发计划项目(2022YFC3502301)。

摘　　要：目的尝试通过自动化分析舌象图像,填补传统中医体质辨识方法的不足之处,推动体质辨识的现代化和智能化。方法通过使用体质调查问卷和DS01-A型舌诊仪采集受试者体质信息和舌象信息,最终纳入阳虚质舌象260张,阴虚质舌象114张。在训练阳虚质与阴虚质舌象分类模型之前,进行了数据增强和舌体分割。采用了U-net网络来分割舌象图像。分类模型是基于ResNet-34网络结构进行训练,并使用了交叉熵损失和Dice损失进行优化。结果在模型评价方面,研究使用精度、损失函数、召回率、F1分数等指标进行性能评估。实验结果显示,ResNet-34模型在验证集中达到了88%的精度,并且在训练数据上表现良好。与其他模型(ResNet-18、ResNet-50和RegNet)相比,ResNet-34模型表现最佳。结论使用深度学习方法可以有效地识别阳虚质和阴虚质的舌象特征,为中医体质现代化和智能化分类提供了新的可能性。Objective This study aims to fill the gaps in traditional Chinese medicine(TCM)constitution identification methods by automating the analysis of tongue image data,promoting the modernization and intelligence of constitution identification.Methods It collected participant constitution information and tongue image data using constitution questionnaires and a DS01-A tongue diagnosis instrument,ultimately including 260 images of Yang deficiency constitution tongue features and 114 images of Yin deficiency constitution tongue features.Before training the classification model for Yang deficiency and Yin deficiency tongue features,data augmentation and tongue image segmentation were performed.In this study,it used a U-net network for tongue image segmentation.The classification model was trained based on the ResNet-34 network architecture and optimized using both cross-entropy and Dice loss functions.Results In terms of model evaluation,this study employed metrics such as accuracy,loss functions,recall and F1 score.Experimental results demonstrated that the ResNet-34 model achieved an 88%accuracy rate on the validation dataset and performed well on the training data.Compared to other models(ResNet-18,ResNet-50,and RegNet),the ResNet-34 model exhibited the best performance.Conclusion These findings suggest that deep learning methods can effectively identify tongue features associated with Yang deficiency and Yin deficiency constitutions,opening up new possibilities for modernizing and automating TCM constitution classification.

关 键 词：中医体质 舌象 深度学习 中医药现代化 

分 类 号：R241.25[医药卫生—中医诊断学]