摘要目的 构建基于视觉注意力的糖尿病视网膜病变(DR)辅助诊断的智能分析系统,实现DR眼底影像的自动筛查及分级.方法 从数据建模及数据分析竞赛平台(Kaggle)中的Diabetic Retinopathy Detection竞赛上下载得到35126张DR眼底图片,并从Messidor网站上下载得到1200张彩色眼底照片.首先,针对现有的DR眼底图像的特征,对视网膜图像进行一系列预处理;然后,在VGG16网络的基础上引入视觉注意力SENet模块,以提高病灶特征的显著性,生成一个网络结构较为复杂的深度卷积神经网络(CNN)SEVGG,该网络基本上继承了VGG16的一些结构参数,而SENet模块参数则根据基本网络及训练数据集进行相应的调整;最后,应用SEVGG网络模型对DR眼底图像进行筛查,并根据不同时期DR的病变程度把眼底图像分成不同等级.配置训练平台及环境并进行算法性能验证实验.结果 将本研究中提出的方法在不同的公开标准数据集上进行检验,最终在基于图像的分类上实现了较高的准确率,其中Kaggle数据集中5分类准确率可达83％,病变检测的敏感性为99.86％,特异性为99.63％,Messidor数据集中4分类准确率可达88％,病变检测的敏感性为98.17％,特异性为96.39％.引入视觉注意力对于病灶点的关注更加显著,有助于DR的精准检测.结论 应用视觉注意力的DR眼底影像筛查及分级方法有效避免了传统人工特征提取和眼底图像分类的一些缺点,且对于病灶点的识别更加精确,不仅优于之前的方法,而且具有较好的鲁棒性及泛化性.

abstractsObjective To construct an intelligent analysis system based on visual attention for diabetic retinopathy ( DR) assistant diagnosis and to realize the automatic screening and grading of fundus images of DR. Methods Total of 35126 DR fundus images were downloaded from the Diabetic Retinopathy Detection competition in the Data Modeling and Data Analysis Competition Platform (Kaggle),and 1200 fundus images were downloaded from the Messidor website. Firstly,according to the characteristics of DR fundus images,a series of preprocessing was carried out for retina images. Then,on the basis of VGG16 network,visual attention SENet module was introduced to improve the saliency of lesion features,and a deep convolution neural network SEVGG with complex network structure was generated. The network basically inherited some structural parameters of VGG16,and the parameters of SENet module were adjusted according to the basic network and training data set. Finally, the SEVGG network model was used to screen the DR fundus image,and the fundus image was divided into different levels according to the degree of lesions of DR in different periods. Configure the training platform and environment and perform algorithm performance verification experiments. Results The method proposed in this study was tested on different open standard datasets,and finally achieved high accuracy in image-based classification. The accuracy of 5 classification in Kaggle dataset was 83％,the sensitivity of lesion detection was 99. 86％ and the specificity was 99. 63％. The accuracy rate of the 4 classification in the Messidor data set was up to 88％,the sensitivity of the lesion detection was 98. 17％,and the specificity was 96. 39％. The introduction of visual attention was more significant for the focus of the lesion,which helped the accurate detection of DR. Conclusions This method effectively avoids some shortcomings of traditional artificial feature extraction and fundus image classification,and is more accurate for lesion recognition. It is not only superior to the previous method,but also shows better robustness and generalization.