Metric learning based multi-branch network for tongue manifestation recognition(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2024年第4期
- Page:
- 521-528
- Research Field:
- 医学人工智能
- Publishing date:
Info
- Title:
- Metric learning based multi-branch network for tongue manifestation recognition
- Author(s):
- REN Siyu1; WU Rui2; LUO Qinglin3; XIAO Kaihui 4; WANG Yifan2; LI Jie2
- 1. Teaching Department of the Open University of Chengdu, Chengdu 610000, China 2. School of Intelligent Technology and Engineering, Chongqing University of Science and Technology, Chongqing 401331, China 3. Chongqing Kerui Pharmaceutical Co., Ltd., Chongqing 400060, China 4. Department of Traditional Chinese Medicine, Beijing Friendship Hospital, Capital Medical University, Beijing 100050, China
- Keywords:
- Keywords: tongue manifestation recognition multi-branch network architecture feature coding loss function multi-label residual mapping
- PACS:
- R318；R241.25
- DOI:
- DOI:10.3969/j.issn.1005-202X.2024.04.018
- Abstract:
- Abstract: Based on metric learning, a novel method of assisting doctors in identifying tongue manifestation is proposed to improve the efficiency and accuracy of tongue manifestation recognition. A total of 111 tongue images are collected, and the data are randomly divided into training set and test set at a ratio of 7:3. Subsequently, a metric learning based multi-branch tongue manifestation recognition network is designed. The deep learning network is divided into 2 parts. The first part is the shared weight layer which employs metric learning based loss function in tongue manifestation feature coding to obtain accurate features. In order to reduce the difficulty of tongue manifestation recognition and improve the accuracy, the latter part is split into 4 branches for tongue manifestation recognition which correspond to the classification of tongue manifestation in traditional Chinese medicine. Additionally, a multi-label residual mapping is constructed to increase inter-class distance and reduce intra-class distance, so as to enhance the accuracy of final recognition. The proposed method achieves a recognition accuracy of 84.8% on the test set of tongue manifestation dataset, indicating that multi-branch network architecture can lower the difficulties in tongue manifestation recognition, especially for the tongue shape and coating nature with multiple feature categories. The loss function in tongue manifestation feature coding can effectively extract tongue features, while multi-label residual mapping can reduce the interference between different categories, which improves the recognition accuracy.
Last Update: 2024-04-25