注意力残差链式融合的彩色眼底图像硬性渗出检测  

Hard exudation detection of color fundus images with chained fusion of attentional residua

作　　者：陈志远 金炜[1] 李纲[1] CHEN Zhi-yuan;JIN Wei;LI Gang(Faculty of Electrical Engineering and Computer Science,Ningbo University,Ningbo 315211,China)

机构地区：[1]宁波大学信息科学与工程学院,浙江宁波315211

出　　处：《光电子．激光》2020年第10期1044-1053,共10页Journal of Optoelectronics·Laser

基　　金：浙江省自然科学基金(LY20H180003);宁波市自然科学基金(2019A610104)资助项目。

摘　　要：硬性渗出物(Hard Exudate,HE)对糖尿病视网膜病变(Diabetic Retinopathy,DR)的早期诊断具有重要意义,从彩色眼底图像中准确检测出HE是诊断和筛查DR的重要步骤。针对以往自动检测方法中存在的分割效果不佳和对小目标误检率高的问题,本文在经典U-Net生成网络的基础上,引入注意力残差链式融合机制,构造出一种适用于视网膜HE检测的注意力残差链式融合生成对抗网络(Chain fusion of attention residuals GAN,CFAR-GAN)。该网络在生成网络编码过程的每个子模块后添加一个残差网络(Residual Network,ResNet)结构,并采用带有残差连接的卷积层链(Residual convolutional layer path,Res path)建立不同层间的跳跃连接,同时将全局最大池化注意力机制用于刻画不同深度特征的权重,以防止训练过程的过拟合从而提高网络的泛化能力。将CFAR-GAN用e-ophtha EX数据库的训练集进行训练,在其测试集上,检测敏感性、PPV和F-score分别为92.5%、88.7%和90.6%;将训练好的网络在另一个独立的DIARETDB1数据库上进行测试,敏感性、特异性和准确性分别为100%、98.5%和99.1%,表明本文所提出的方法具有理想的泛化能力,这对于准确高效地检测眼底图像中的HE,实现DR早期的自动诊断具有积极意义。Hard exudate(HE)is of great importance for the early diagnosis of diabetic retinopathy(DR),and the accurate detection of HE from the fundus image is an important step in the diagnosis and screening of DR.To solve the problems of poor segmentation effect and high false detection rate of small targets in previous automatic detection methods,based on the classic U-Net generation network,this paper introduces the chain fusion mechanism of attention residuals,and constructs a chain fusion of attention residuals generation countermeasures network(CFAR-GAN)suitable for retinal HE detection.A network in the generated code after each module of the process of adding a residual network(ResNet)structure,and uses the convolution with residual connection layer chain(Res path)jump between different layers of connections,at the same time the global biggest pooling attention mechanism is used for depicting the weight of different depth characteristics,in order to prevent the training process of the fitting so as to improve the generalization ability of the network.CFAR-GAN was trained with the training set of e-ophtha EX database.In the test set,the detection sensitivity,PPV and F-score were 92.5%,88.7%and 90.6%,respectively.The trained network was tested on another independent DIARETDB1 database,and the sensitivity,specificity and accuracy were 100%,98.5%and 99.1%,respectively,indicating that the method proposed in this paper has ideal generalization ability,which is of positive significance for accurate and efficient detection of HE in fundus images and early automatic diagnosis of DR.

关 键 词：眼底图像 硬性渗出 检测 注意力 残差链式 

分 类 号：R318[医药卫生—生物医学工程]