Abstract: Tongue diagnosis is one of the characteristic diagnostic methods in traditional Chinese medicine (TCM). The book named Medical Mirror profoundly explains the intimate relationship between tongue images and human health status, as well as organ diseases. Therefore, the important task of modern intelligent tongue diagnosis in TCM is the processing for features of tongue image. To address these difficulties in learning the distribution rules of tongue image features and extracting these features, this paper proposes a novel method for extracting tongue image features based on the generative adversarial network (TongueIFE-GAN). This method constructs an extraction network for the latent features of tongue images through the adversarial idea. It consists of two important components, namely the discriminator and generator, thereby establishing a mapping relationship between the image reconstruction quality and the feature extraction effect of tongue images. To enhance the interpret ability of the deep algorithm, the class activation mapping mechanism (CAM) is integrated into the network discriminator to further optimize the feature processing performance of the encoder. Moreover, it provides a visual explanation of the sensitive image areas that the TongueIFE-GAN model focuses on during feature extraction. Meanwhile, driven by the tasks of tongue image segmentation and classification, the new model can self-optimize its ability to extract tongue features. Through various experiments, the experimental results demonstrate that the tongue image segmentation and classification task based on TongueIFE-GAN model has better segmentation performance and classification accuracy than the benchmark model and comparison methods on IoU and Dice indices. TongueIFE-GAN employs adversarial thinking to devise a novel feature extraction method and attention visualization mechanism, offering a fresh perspective on tongue image feature research.

Key words: generative adversarial network, tongue image, class activation mapping, tongue image segmentation, tongue classification

摘要： 舌诊是中医特色诊法之一，《医镜》深刻阐释了舌象与人体健康状况、脏腑病变的密切关系。因此，现代中医智能舌诊的重要工作便是舌象图像特征处理，针对舌象图像特征分布规律的学习和特征提取困难的问题，提出一种基于生成对抗网络的舌象图像特征提取新方法（TongueIFE-GAN）。该方法通过对抗思想构建一种面向舌象图像潜在特征的提取网络，它包括判别器和生成器两个重要组成部分，藉此建立图像重构质量与舌象图像的特征提取效果映射关系。为增强深度算法的可解释性，在网络判别器中融入类激活映射机制，进一步优化编码器的特征处理性能，并对TongueIFE-GAN模型在提取特征时关注的图像敏感区域作出可视化解释。同时，在舌象图像分割、分类任务驱动下，新模型可自优化舌象特征提取的能力。通过多组实验，结果表明，基于TongueIFE-GAN模型的舌象分割和分类任务，其分割性能IoU与Dice指标值，以及分类准确率均优于基准模型和对比方法。TongueIFE-GAN利用对抗思想构建新型的特征提取以及注意力可视化机制的研究方法，可为舌象图像特征研究提供一种新的思路。

关键词: 生成对抗网络, 舌象图像, 类激活映射, 舌象图像分割, 舌象分类