基于深度学习的舌象自监督聚类-《中国医学物理学杂志》

文章信息/Info

作者:: 罗思言1; 王心舟2; 饶向荣1; 1.中国中医科学院广安门医院肾病科，北京 100053； 2.同济大学电子与信息工程学院，上海 201804

Author(s):: LUO Siyan1; WANG Xinzhou2; RAO Xiangrong1; 1. Department of Nephropathy, Guanganmen Hospital, China Academy of Chinese Medical Sciences, Beijing 100053, China 2. College of Electronic and Information Engineering, Tongji University, Shanghai 201804, China

Keywords:: Keywords: self-supervised learning tongue diagnosis image clustering Gaussian mixture model

摘要:: 目的：解决人工智能舌诊领域数据标注成本较高且带有较强主观性的问题。方法：基于深度学习中的对比学习技术，对舌象进行自监督聚类。该方法首先利用卷积神经网络将不同数据增强模式下的舌象映射到潜在空间，并在学习同类实例之间共同特征的同时尽可能区分非同类实例；随后利用高斯混合模型对网络提取的特征向量进行聚类。结果：在无需引入先验知识的情况下，利用300张舌象仪采集的无标签图像取得了52.54%的聚类纯度。结论：该方法一定程度上将医疗工作者从费事费力的数据标注工作中解放出来。除应用于自动化舌象分类外，该方法还可进一步针对不同病症的特殊舌象症候群进行聚类分析，其提取的舌象特征也可为舌体分割、舌色分类、苔质分区等下游任务提供预训练的参考。

Abstract:: Abstract: Objective To solve the problem of high cost of data annotation and evident subjectivity in the artificial intelligence-based tongue diagnosis. Methods The proposed method adopted the contrastive learning of deep learning for realizing the self-supervised clustering of tongue images. The convolutional neural network was used to map tongue images under different data enhancement patterns into the latent space, and to distinguish non-similar instances as much as possible while learning the common features among similar instances. The feature vectors extracted by the network were subsequently clustered using Gaussian mixture model. Results A clustering accuracy of 53% was obtained using 300 unlabeled images collected by the instrument for tongue image analysis without introducing prior knowledge. Conclusion The method somewhat frees medical practitioners from the laborious and tedious task of data labeling. In addition to its application to automated tongue classification, the method can further be used for the clustering analysis of special tongue syndromes of different diseases and its extracted tongue features can provide pre-trained references for downstream tasks such as tongue segmentation, tongue color classification, and moss texture partitioning.