摘要
中医舌象与脏器病理特征具有一定相关性,通常用来分析人体脏器病理变化。采用深度学习方法可以提取舌象深层特征,反映脏器病理变化,但现有的深度学习模型结构单一,无法有效提取舌象深层的局部和全局特征,也未结合临床舌象浅层特征,如边缘特征、纹理特征等进行综合分析,降低了模型的分类精度和泛化性能。本文提出了一种基于多特征融合的迁移学习分类网络,采用基于自适应注意力机制迁移学习框架,提取舌象深层特征并融合舌象边缘特征、纹理特征,提高舌象分类精度和泛化性能。实验结果表明,本文方法针对6分类中医证候舌象样本,模型分类精度为0.953±0.031,灵敏度0.952±0.032,F1值0.952±0.032,与典型分类模型相比,具有较高的分类精度和泛化性能。
Tongue image of traditional Chinese medicine has relationship with viscera pathological features, and are usually used to judge the pathological variants of the human organs in the clinical diagnosis of TCM. Deep learning methods can extract in-depth features of tongue image to reflect the pathological variants of the human organs. However, current deep learning model has simple and single functions, and is unable to extract in-depth local and global features of tongue image and to combine the superficial features such as edge, texture, resulting reduction of classification accuracy and generalization. In this article we proposes a novel transfer learning network based on multi-feature fusion, which can effectively extract in-depth features of tongue image using a transfer learning network with adaptive attention mechanism combined with superficial features such as edge, texture feature etc. The experimental results show that the proposed model can achieve precision of 0.953±0.031, sensitivity of 0.952±0.032, and F1-score of 0.952±0.032 on the 6-class tongue images of TCM syndrome, which is significantly higher than the state-of-the-art networks in classification accuracy and generalization performance.
作者
徐雍钦
杨晶东
姜泉
韩曼
宋梦歌
XU Yongqin;YANG Jingdong;JIANG Quan;HAN Man;SONG Mengge(Autonomous Robotics Laboratory,School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Division of Rheumatology,Guang′anmen Hospital,China Academy of Chinese Medical Sciences,Beijing 100053,China)
出处
《智能计算机与应用》
2022年第7期25-34,共10页
Intelligent Computer and Applications
基金
国家自然基金(81973749)
中国中医科学院科技创新工程项目(CI2021A01503)。
关键词
舌象分类
深度学习
注意力机制
特征融合
tongue images
deep learning
attention mechanism
feature fusion
作者简介
徐雍钦(2000-),男,本科生,主要研究方向:人工智能、机器学习与大数据分析等;杨晶东(1973-),男,博士,副教授,主要研究方向:人工智能、机器学习与大数据分析、机器视觉等;姜泉(1961-),女,博士,教授,主任医师,主要研究方向:风湿免疫病的中医、中西医结合临床及基础研究;韩曼(1984-),女,博士,副主任医师,主要研究方向:风湿免疫病的中医、中西医结合临床及基础研究;宋梦歌(1993-),女,博士研究生,主要研究方向:风湿免疫疾病的临床与基础研究。