卷积神经网络在糖网病眼底图像分类中的应用研究

发布时间：2018-03-23 21:57

本文选题：糖尿病视网膜病变　切入点：眼底图像　出处：《深圳大学》2017年硕士论文

【摘要】：糖尿病视网膜病变(Diabetic Retinopathy,DR)是由于糖尿病引起视网膜血管壁受损,导致视网膜上出现微血管瘤、硬性渗出物和出血等病灶,使视觉功能下降,是主要致盲疾病之一。目前在临床上,眼科医生对DR诊断最常用的方法是观察眼底视网膜,通过对视网膜上异常病灶的分析进行诊断。但是,由于DR病变种类多样,形态各异,程度不一,导致眼科医生诊断困难。特别是在进行大规模DR筛查中,应用计算机辅助诊断技术则可以帮助眼科医生快速、有效的实现诊断。目前主要的DR眼底图像分类方法是基于眼底图像病灶检测和基于眼底图像全图特征进行分类的方法。但是这些传统的方法首先需要人工设计特征,然后进行分类器训练,效率较低并且适应性差。因为卷积神经网络(Convolutional Neural Network,CNN)结合了特征提取与分类,能够自动学习特征,从而避免了繁琐的人工特征提取工作。本文针对眼底图像样本量偏少的情况,使用基于CNN的迁移学习方法对眼底图像进行分类研究。论文主要的研究内容包含两部分:一部分是特征迁移学习,利用大规模自然图像分类中多种卷积神经网络预训练模型作为特征提取器对眼底图像提取特征,使用支持向量机进行分类。另一部分是微调迁移学习,首先对眼底图像进行数据扩增处理,然后使用大规模自然图像分类中多种卷积神经网络预训练模型参数初始化网络参数,最后利用眼底图像训练网络,其中微调包括所有层微调和局部层微调两种方式。实验数据包括DR1和Messidor两个公开数据集,以及一个自有数据集,并采用五倍交叉验证对各种方法进行对比。在DR1数据集上得到的最优分类正确率为94.52%,在Messidor上得到的最优分类正确率为92.01%,在自有数据上得到的最优分类正确率为97.93%。实验结果表明了使用CNN和迁移学习对DR眼底图像分类的有效性。另外,论文对卷积网络得到的特征进行了可视化分析,验证了CNN特征提取从低层次具体信息到高层次抽象信息的过程。还对错分样本进行了分析,发现错分的原因可能是采集的数据质量差、数据类型不全和数据数量不足等。最后,实现了DR眼底图像分类原型系统,可以为眼科医生提供辅助诊断。
[Abstract]:Diabetic retinopathy (DRR) is one of the main causes of blindness due to the damage of retinal vascular wall caused by diabetes, the appearance of microhemangioma, rigid exudate and hemorrhage on the retina, and the decrease of visual function. The most common method used by ophthalmologists in the diagnosis of Dr is to observe the retinal fundus and diagnose the abnormal lesions in the retina. However, because of the variety, shape and degree of Dr lesions, This makes it difficult for ophthalmologists to diagnose. Especially in large-scale Dr screening, the use of computer-aided diagnostic techniques can help ophthalmologists quickly. At present, the main methods of Dr fundus image classification are based on the focus detection of the fundus image and the classification method based on the whole image feature of the fundus image. But these traditional methods need to design the features manually. Then the classifier is trained, which is low efficiency and low adaptability, because convolution neural network Convolutional Neural Network CNNs combine feature extraction and classification, and can automatically learn features. In order to avoid the tedious work of artificial feature extraction. This paper uses CNN based transfer learning method to classify the fundus images. The main contents of this paper include two parts: one is feature transfer learning, and the other is feature transfer learning. Several convolution neural network pretraining models in large-scale natural image classification are used as feature extractors to extract features from eye fundus images, and support vector machines are used to classify them. First, the image of the fundus is processed by data amplification, then the parameters of the neural network pre-training model are initialized by using a variety of convolution neural network parameters in the classification of large scale natural images. Finally, the network is used to train the image of the fundus. The experimental data include two kinds of open data sets, DR1 and Messidor, and a private data set. The best classification accuracy is 94.522 on DR1 data set, 92.01 on Messidor and 97.93 on own data. The experimental results show the effectiveness of using CNN and transfer learning to classify Dr fundus images. In this paper, the features obtained from convolutional network are analyzed visually, and the process of feature extraction from low level concrete information to high level abstract information is verified by CNN feature extraction. The reasons for the misclassification may be poor data quality, incomplete data type and insufficient data quantity. Finally, the prototype system of Dr fundus image classification is implemented, which can provide ophthalmologist with auxiliary diagnosis.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;R587.2;R771.3

【相似文献】