融合双眼特征的糖网病图像识别方法
Identification of Diabetic Retinopathy Stages by Correlating Binocular Cues
投稿时间:2020-04-21  修订日期:2020-04-21
DOI:
中文关键词: 糖网病视网膜病变自动检测  卷积神经网络  图像分类  特征融合
英文关键词: Diabetic Retinopathy classification  Fundus images  Binocular feature fusion  Deep network  Kappa metric
基金项目:湖北省自然科学基金
作者单位E-mail
王娇 中南民族大学 生物医学工程 wangjiaoscmu@outlook.com 
方全 中南民族大学 生物医学工程  
罗芬 中南民族大学 生物医学工程  
唐奇伶 中南民族大学 生物医学工程 qltang@mail.scuec.edu.cn 
摘要点击次数: 287
全文下载次数: 0
中文摘要:
      本文针对糖尿病视网膜病变的眼底图像数据,根据传统的卷积神经网络模型设计出一种能够针对此病程进行分类研究的网络模型。首先引入一种全新的基于双眼的特征融合模型来比较不同模型结构的处理效果,通过实验证明了此模型在对糖尿病视网膜病变眼底图像进行分类时的有效性和优越性。实验数据为Kaggle公司提供的公开数据集,模型在此数据上的最优分类准确率为77.54%,相较于其他模型都有所提高。其次通过对模型的损失函数和学习率进行调整,分析得出在不同评价准则下的最优损失函数和最佳学习率。最后对卷积神经网络的特征层进行可视化,验证卷积神经网络是从低层的具体特征逐渐融合为高层的抽象信息的过程,并对错分的图像样本进行分析,总结了造成错误分类的两种主要原因:微血管瘤特征的不明显和相机伪影构成的图像噪声。
英文摘要:
      In this paper, we build a double input-double output deep network to identify Diabetic Retinopathy (DR) stages based on the correlation between the left and right retinal lesions of DR patients, which can correlate and integrate the binocular cues to improve the classification performance of DR. This proposed network takes the binocular fundus images as input, extracts the respective features by the convolutional networks, fuses the cues of both eyes in the advanced feature layer, and then produce the respective result for each input by using double outputs. Our pipeline can not only extract the characteristic of a single-eye fundus image, but also exploit the association information of double-eye fundus images. In addition, we define a novel loss function by combining the cross entropy and Kappa score, in which the cross-entropy error guides the learning of the model weights in terms of the distance between the predicted distribution and the true distribution and Kappa score perform the consistency check of the model predictions and the class labels based on the confusion matrix. Finally, we verify the effectiveness and superiority of our method on the Kaggle Diabetic Retinopathy image dataset, and investigate the misclassified samples and analyze the reasons of the misclassification.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭