基于深度学习的舌象特征研究  

Study on Tongue Image Features Based on Deep Learning

作　　者：崔涛 何佳俊 何华 李睿 赵亮 垢德双 CUI Tao;HE Jiajun;HE Hua;LI Rui;ZHAO Liang;KOU Deshuang(Tianjin Medical Device Evaluation and Inspection Center,Tianjin 300191,China;Beijing Shennong Intelligent Technology Co.Ltd.,Beijing 102208,China;Medvalley TCM Technology(Tianjin)Co.Ltd.,Tianjin 300392,China)

机构地区：[1]天津市医疗器械审评查验中心,天津300191 [2]北京神农智能技术有限公司,北京102208 [3]慧医谷中医药科技(天津)股份有限公司,天津300392

出　　处：《医学信息学杂志》2024年第7期81-87,共7页Journal of Medical Informatics

基　　金：国家重点研发计划(项目编号:2023YFC3606200)。

摘　　要：目的/意义应用深度学习技术达到舌象分析自动化的目的,为中医舌象标准化提供参考依据,进一步推进中医诊疗技术现代化进程。方法/过程构建以区域关联性为基础、引入标签松弛技术的语义分割损失函数,显式约束舌象分割模型学习局部区域各像素关联性的同时,对错误标签具有一定的容错能力;从舌象特征中内含的颜色相关性潜在先验出发,在模型构建阶段将舌象特征仅解耦为两类下游多标签分类任务,在加速模型拟合的同时有效降低模型复杂度。结果/结论在自建数据集上验证算法有效性,舌象分割MIoU指标为96.57%,舌象分析宏F 1值、平均准确率分别为88.58%、82.59%。Purpose/Significance To apply deep learning technology to achieve the purpose of tongue image analysis automation,so as to provide references for the standardization of tongue image of traditional Chinese medicine(TCM),and further promote the modernization of TCM diagnosis and treatment technology.Method/Process It develops a new semantic segmentation loss function with region-based correlation and label relaxation to enhance the capability of tongue image segmentation model to learn pixel relationships and handle mislabeled data.Additionally,leveraging inherent color-related priors in tongue image features,the model is simplified by decomposing it into two multi-label classification tasks,thereby accelerating model training and reducing its complexity.Result/Conclusion The proposed algorithm is proven effective on a self-constructed dataset,attaining a high 96.57%MIoU in tongue segmentation,and demonstrating strong performance with a macro F 1-score of 88.58%and average accuracy of 82.59%.

关 键 词：迁移学习 舌象特征 深度学习 舌象分割 中医舌诊 

分 类 号：R-058[医药卫生]