摘要
糖尿病性视网膜病变严重威胁患者视力,现有的手工分级糖尿病性视网膜病变眼底图像方法费时费力,针对此问题提出一种基于Inception-v4结构的集成网络模型。该模型由5个基于Inception-v4的架构网络集成。通过训练和正则化随机性,即使用相同的训练数据集和ImageNet初始化学习不同的图像特征。在Kaggle-DR数据集上集成Inception-v4网络的AUC值达到0.992;在Messidor-2数据集上集成Inception-v4网络的AUC值达到0.977,灵敏度达到0.923,特异性达到0.947。与在Messidor-2数据集上进行测试的单一网络相比,集成网络表现比单一网络要好,说明集成Inception-v4网络对提高糖尿病性视网膜病变的筛查效率具有一定意义。
Diabetic retinopathy is a serious threat to the vision of the patients,and the existing manual retinal image method of diabetic retinopathy is time-consuming and laborious.An ensemble network model based on Inception-v4 structure is proposed to solve the problems of manual classification.In the current study,we adopt an ensemble of five classification model instances each of which was on the basis of improved Inception-v4 architecture.Different image features can also be learned from the Inception-v4 network through training and regularization randomness,using the same training data set and ImageNet initialization.On the Kaggle-DR data set,the AUC value of the ensemble Inception-v4 network reached 0.992;on the Messidor-2 data set,the AUC value of the ensemble Inception-v4 network was 0.977,the sensitivity 0.923,and the specificity 0.947.The ensemble network performs slightly better than the single network tested on the Messidor-2 dataset.The results indicate that the ensemble Inception-v4 network proposed has a certain significance for improving the screening efficiency of diabetic retinopathy.
作者
王宇光
李峰
WANG Yu-guang;LI Feng(School of Mechanical Engineering,University of Shanghai for Science and Technology;School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《软件导刊》
2021年第5期164-167,共4页
Software Guide
基金
国家自然科学基金项目(51675321,61905144)。
