A TCM Tongue Color and Coating Color Collaborative Classification Method Based on Relation Network
- VernacularTitle:基于关系网络的中医舌色苔色协同分类方法
- Author:
Enci WANG
1
;
Li ZHUO
;
Yanping LI
;
Xin WANG
;
Yang YANG
;
Wei WEI
Author Information
- Publication Type:Journal Article
- Keywords: TCM tongue diagnosis; Collaborative classification of tongue color and coating color; Deep learning; Relational network; Feature fusion
- From: World Science and Technology-Modernization of Traditional Chinese Medicine 2025;27(5):1207-1218
- CountryChina
- Language:Chinese
- Abstract: Objective There are so many tongue diagnosis features that need to be analyzed in Traditional Chinese Medicine(TCM),including tongue color,coating color,tongue morphology,and tongue dynamics,with about 30 different categories.Different tongue diagnosis features are often analyzed by individual methods,which fails to fully utilize the correlation between diagnostic features and greatly increases the overall implementation complexity of the TCM analysis system.Methods Therefore,in this paper,a TCM tongue color and coating color collaborative classification method based on relation network has been proposed.Firstly,a dual branch lightweight convolutional neural network has been constructed.By designing low and high layer feature fusion modules and combining Coordinate Attention mechanism,high classification accuracy can be achieved with lower model complexity.Secondly,a network for the tongue color and coating color relationship leaning has been designed to model the correlation between the two labels.By the network,the nonlinear information can be learned,which is used as a supplement for collaborative classification.Results The experimental results on two self-established TCM tongue diagnosis feature analysis datasets show that,the accuracy of tongue color classification reaches 95.17%and 93.67%respectively,while that of coating color classification reaches 91.11%and 90.53%respectively.Conclusion The proposed collaborative classification method can improve the classification accuracy of the two tasks,but simultaneously,the overall model complexity is reduced by almost half.
