基于双模型互学习的半监督中医舌诊图像分割方法  

Semi-supervised tongue image segmentation method for traditional chinese medicine based on mutual learning with dual models

作　　者：李方旭 徐望明[1] 徐雪 贾云 LI Fangxu;XU Wangming;XU Xue;JIA Yun(School of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,China;School of Medicine,Wuhan University of Science and Technology,Wuhan 430065,China;Affiliated Hospital,China University of Geosciences(Wuhan),Wuhan 430074,China)

机构地区：[1]武汉科技大学信息科学与工程学院,湖北武汉430081 [2]武汉科技大学医学院,湖北武汉430065 [3]中国地质大学(武汉)医院,湖北武汉430074

出　　处：《液晶与显示》2024年第8期1014-1023,共10页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.51805386);国家重点研发计划(No.3502300,No.3502302)。

摘　　要：舌体图像的准确分割是中医舌诊客观化分析的重要前提,目前广泛采用的全监督分割方法需要对大量像素级标注的样本进行训练,而基于单一模型的半监督分割方法对学习到的错误知识缺乏自我纠正能力。针对这一问题,本文提出一种新颖的基于双模型互学习的半监督舌体图像分割方法。首先,模型A、B分别在有标签数据集上进行监督训练。之后,模型A、B进入互学习阶段,使用本文设计的互学习损失函数,根据双模型对无标签数据预测的分歧而赋予不同的权重。先由模型A对无标签数据集生成伪标签,模型B在有标签数据集和伪标签数据集上进行微调,然后模型B再对无标签数据集生成伪标签,模型A以同样的方式进行微调。双模型微调结束后,选择性能更佳的一个作为最终的舌体图像分割模型。实验结果表明,本文方法的有标签数据比例为1/100、1/50、1/25、1/8时,mIoU分别达到96.67%、97.92%、98.52%、98.85%,优于对比的其他典型半监督方法。本文方法仅需少数标签数据便可达到较高的舌体图像分割精度,可为后续的中医舌色、舌形等舌象分析应用奠定坚实基础,推动中医诊疗数字化进程。Accurate tongue image segmentation is a crucial prerequisite for objective analysis in tongue diagnosis in traditional Chinese medicine(TCM).At present,the widely-used full-supervised segmentation methods require a large number of pixel-level annotated samples for training,and the single-model-based semi-supervised segmentation methods lack the ability to self-correct the learned error knowledge.To address this issue,a novel semi-supervised tongue image segmentation method based on mutual learning with dual models is proposed.Firstly,model A and B undergo supervised training on the labeled datasets.Subsequently,model A and B enter the mutual learning phase,utilizing a designed mutual learning loss function,in which different weights are assigned based on the disagreement between predictions of the two models on the unlabeled data.Model A generates the pseudo-labels for the unlabeled dataset,and model B fine-tunes on both the labeled dataset and the pseudo-labeled dataset.Then,model B generates the pseudo-labels for the unlabeled dataset,and model A fine-tunes in the same manner.After the dual-model fine-tuning process,the model with better performance is selected as the final tongue image segmentation model.Experimental results show that with labeled data proportions of 1/100,1/50,1/25,and 1/8,the mean intersection over union(mIoU)achieved by the proposed method is 96.67%,97.92%,98.52%,and 98.85%,respectively,outperforming other typical semi-supervised methods compared.The proposed method achieves high precision in tongue image segmentation with only a small number of labeled data,laying a solid foundation for subsequent applications in TCM such as tongue color,tongue shape and other tongue image analysis,which can promote the digitization of TCM diagnosis and treatment.

关 键 词：半监督 互学习 舌体图像分割 损失函数 中医数字化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]