摘要
为解决基于卷积神经网络(CNN)的糖尿病视网膜病变(DR)图像分类算法上普遍存在模型参数难以训练、易过拟合的问题,本文提出一种基于迁移学习和支持向量机(SVM)分类器的DR图像分类算法。首先,对DR图像进行预处理和数据扩增;其次,采用迁移学习方法预初始化深度学习分类算法中的经典框架VGGNet-16网络的模型参数,固定浅层网络参数不变,微调深层网络参数;最后,提取VGGNet-16最后一个隐藏层的特征向量训练支持向量机(SVM)分类器判定DR图像是否病变。实验结果表明,在Kaggle-DR公共数据集共35126张DR图像进行实验,在随机抽取的3500张作为测试集,分类准确率为0.931、敏感性为0.933、特异性为0.928,并能加快网络收敛和提高模型的泛化性。
Aiming at the problem that convolutional neural network based algorithm for diabetic retinal image classification is hard to training a huge parameters of model and easy to overfit.We propose an transfer learning and SVM based algorithm for DR image classification.Firstly,the images were preprocessing and data augumentation.Then,we take convolutioanal neural net-work（VGGNet-16）pre-train on ImageNet,flx most of earlier layers,and only train its higher-level portion .Finally,we extract deep features from the last-fully-connected layer of the fine-tuning VGGNet-16.These features are then feeded into SVM to dis-criminate two-classes DR images.In the test set of Kaggle-DR dataset,the proposed approaches have an mean classification ac-curacy of 0.931 sensitivity of 0.933 and specificity of 0.928.The experimental results demonstrate that our method has the char-acteristics of fast network convergence and strong generation.
作者
王晓权
郑绍华
潘林
Wang Xiaoquan;Zheng Shaohua;Pan Lin(College of Physics and Information Engineering, Fuzhou University, Fuzhou 350016, China)
出处
《信息通信》
2018年第4期96-100,共5页
Information & Communications
基金
福建省自然科学面上基金(2016J01297)基于信息融合的糖尿病视网膜病变计算机辅助诊断系统关键技术研究
