基于胶囊网络和图神经网络的糖尿病视网膜病变图像分级  

Grading of diabetic retinopathy images based on capsule networks and graph neural network

作　　者：王婧祎 温凯 袁立明 孙婧 徐海霞 温显斌[1,2] WANG Jingyi;WEN Kai;YUAN Liming;SUN Jing;XU Haixia;WEN Xianbin(School of Computer Science and Engineering,Tianjin University of Technology,Tianjin 300384,China;Key Laboratory of Computer Vision and System,Ministry of Education,Tianjin University of Technology,Tianjin 300384,China;Eye Hospital,Tianjin Medical University,Tianjin 300392,China)

机构地区：[1]天津理工大学计算机科学与工程学院,天津300384 [2]天津理工大学计算机视觉与系统教育部重点实验室,天津300384 [3]天津医科大学眼科医院,天津300392

出　　处：《天津理工大学学报》2025年第1期97-103,共7页Journal of Tianjin University of Technology

基　　金：天津市新一代人工智能科技重大专项基金(18ZXZNGX00150);天津市技术创新引导专项基金(21YDTPJC00250)。

摘　　要：糖尿病视网膜病变(diabetic retinopathy, DR)已成为全球失明人数攀升的主要原因之一。目前,基于深度学习的智能分级已成为DR智能诊断研究的热点。现有的基于卷积神经网络(convolutional neural network, CNN)的DR智能分级模型已取得不错效果,这些模型将DR分级看作是图像分类任务,但是,这些模型更多关注深层特征提取,而未考虑DR图像本身特性、不同级之间的关系等。为了克服这些不足,提出一种新的DR智能分级模型。该模型利用胶囊网络的细节捕获能力,将CNN的池化层替换成胶囊网络,提取DR图像的深层细节特征,其次,考虑到相邻级别间DR图像差别小、容易混淆,利用图神经网络捕获DR级别间关系,最后,两路网络输出通过自适应权重完成融合,给出整个网络的分级结果。该模型在两个数据集上进行评估实验,并取得良好效果,进一步论证了该方法的优越性。Diabetic Retinopathy(DR)has become one of the main reasons for the rise in the number of blind people worldwide.At present,the intelligent grading based on deep learning has become a hotspot in DR intelligent diagnosis,The existing DR Intelligent classification models based on convolutional neural networks(CNN)have achieved good results,these models regard DR classification as images classification task,but these models pay more attention on the extraction of deep features,and do not consider the characteristics of the DR image itself,the relationships between different levels,etc.In order to overcome these shortcomings,a new DR intelligent classification model is proposed.First,by using the detail capturing capability of capsule network,the pooling layer of CNN is replaced by capsule network to extract the deep detail features of DR images.Secondly,considering the small difference of DR images between adjacent levels,which is easy to be confused,the graph neural network is used to capture the relationship between DR levels.Finally,the output of the two networks is fused through the adaptive weight,and give the classification results of the whole network.The proposed model is evaluated on two datasets respectively,and good results are obtained,which further demonstrates the superiority of the proposed method.

关 键 词：糖尿病性视网膜病变 分级 图卷积网络 卷积神经网络 胶囊网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]