Abstract:

In view of the lack of large-scale publicly marked glaucoma data set, in order to solve the problems of insufficient learning ability of small samples and low classification accuracy, a glaucoma fundus image recognition system based on transfer learning is proposed. The obtained glaucoma fundus image is preprocessed, such as denoising, deleting redundant background, Extracting Region of Interest（ROI）, image enhancement and so on. Then, based on the VGG16 network, the fully connected layer is redesigned to obtain a simplified deep neural network model Reduce-VGGNet（R-VGGNet）. In the training process of R-VGGNet network, the convolution layer and pooling layer inherit the VGG16 model and pre-train the weight parameters on the ImageNet data set, and the parameters of the fully connected layer are adaptively adjusted according to the glaucoma data set. Finally, different network structures and different training strategies are tested for performance and comparative tests of different classification methods. The experimental results show that the recognition method based on R-VGGNet network model improves the accuracy of distinguishing glaucoma patients. The accuracy rate can reach to 91.7%, which provides a good solution for the diagnosis and treatment of clinicians.

Key words: glaucoma, transfer learning, extracting Region of Interest（ROI） network, VGG16 network, Reduce-VGGNet（R-VGGNet）

摘要：

针对目前缺少大型公开已标记的青光眼数据集，为了解决小样本学习能力不足、分类精度低等问题，提出一套基于迁移学习的青光眼眼底图像识别系统。对获取的青光眼眼底图像进行去噪、删除多余背景、提取感兴趣区域（ROI）、图像增强等预处理操作。在VGG16网络的基础上，对全连接层进行重新设计，得到一个简化的深度神经网络模型Reduce-VGGNet（R-VGGNet）。R-VGGNet网络在训练过程中，其卷积层与池化层继承VGG16模型在ImageNet数据集上预训练得到权值参数，全连接层的参数则根据青光眼数据集进行自适应调整。针对不同的网络结构和不同的训练策略进行了性能测试以及不同分类方法的对比实验。实验结果表明：基于R-VGGNet网络模型的识别方法提高了判别青光眼患者的准确率，可达91.7%，为临床医生诊断治疗提供了良好的解决方案。

关键词: 青光眼, 迁移学习, 提取感兴趣区域（ROI）, VGG16网络, 简化的深度神经网络（R-VGGNet）