GA-TongueNet: Tongue Image Segmentation Network Using Innovative DiFP and MDi for Stable Generalization Ability

Zhiyu Dong¹Le Zhao^1*Yajun Fan¹Haihua Ma¹Changle Shao¹Yiran Zhang¹Peng Li^2*

• ¹College of Information Science and Engineering, Henan University of Technology, Zhengzhou, Henan Province, China
• ²Institute for Complexity Science, Henan University of Technology, Zhengzhou, China

Tongue is directly or indirectly connected to many internal organs in Traditional Chinese Medicine (TCM). In computer-aided diagnosis, tongue image segmentation is the first step in tongue diagnosis, and the precision of this segmentation is decisive in determining the accuracy of the tongue diagnosis results. Due to challenges such as insufficient available sample size and complex background, the generalization and robustness of current tongue segmentation algorithms are usually poor, which seriously hinders the practicality of tongue diagnosis. In this article, a GA-TongueNet, namely Tongue Segmentation Network for Stable Generalization Ability, based on self-attention architecture is proposed, which is a tongue segmentation network that can simultaneously have strong generalization ability and accuracy under small samples and diverse background conditions. Firstly, GA-TongueNet is built upon the transformer architecture, embedding the dilated feature pyramid (DiFP) module and the multi-dilated convolution (MDi) module proposed in this article. Secondly, the DiFP module is integrated to comprehend both the overall tongue image structure and intricate local details, while the MDi module is specifically designed to preserve a high feature resolution. Therefore, the network adeptly captures longrange dependencies, extracts high-level semantic content, and retains low-level detail information from tongue images. Moreover, it maintains decent precision and stable generalization capabilities, even when dealing with limited sample sizes. Experimental results show that the accuracy and generalization ability of GA-TongueNet in complex environments are significantly better than various existing semantic segmentation algorithms based on Convolutional Neural Networks (CNN) and Transformer architectures.