糖尿病性视网膜图像的深度神经网络分类方法  

Diabetic retinal image classification method based on deep neural network

作　　者：丁蓬莉 李清勇[1] 张振[1] 李峰[1] 

机构地区：[1]北京交通大学轨道交通数据分析与挖掘北京市重点实验室,北京100044

出　　处：《计算机应用》2017年第3期699-704,共6页journal of Computer Applications

基　　金：北京市自然科学基金资助项目(4142043);中央高校基本科研业务费专项基金资助项目(2014JBZ003)~~

摘　　要：针对传统的视网膜图像处理步骤复杂、泛化性差、缺少完整的自动识别系统等问题,提出了一套完整的基于深度神经网络的视网膜图像自动识别系统。首先,对图像进行去噪、归一化、数据扩增等预处理;然后,设计了紧凑的神经网络模型——CompactNet,CompactNet继承了AlexNet的浅层结构参数,深层网络参数则根据训练数据进行自适应调整;最后,针对不同的训练方法和不同的网络结构进行了性能测试。实验结果表明,CompactNet网络的微调方法要优于传统的网络训练方法,其分类指标可以达到0.87,与传统直接训练相比高出0.27;对于LeNet,AlexNet和CompactNet三种网络模型,CompactNet网络模型的分类准确率最高;并且通过实验证实了数据扩增等预处理方法的必要性。Aiming at the problems of complex retinal image processing, poor generalization and lack of complete automatic recognition system, a complete retinal image automatic recognition system based on deep neural network was proposed. Firstly, the image was denoised, normalized, and data preprocessed. Then, a compact neural network model named CompactNet was designed. The structure parameters of CompactNet were inherited from AlexNet. The deep network parameters were adjusted adaptively based on the training data. Finally, the performance experiments were conducted on different training methods and various network structures. The experimental results demonstrate that the fine-tuning method of CompactNet is better than the traditional network training method, the classification index can reach 0.87, 0.27 higher than the traditional direct training. By comparing LeNet, AlexNet and CompactNet, CompactNet network model has the highest classification accuracy, and the necessity of preprocessing methods such as data amplification is confirmed by experiments.

关 键 词：糖尿病性视网膜图像 深度学习 卷积神经网络 图像分类 微调 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]