糖尿病视网膜病变（DR）是导致中老年失明最常见的原因之一，精确预测DR分类有助于及早干预。传统DR分类依赖局部图像特征且不够精细。本研究在课题组前期神经网络和医学图像预测疾病研究基础上，采用卷积神经网络整体提取DR眼底图像特征建立多分类模型。针对无标记DR眼底图像，利用降噪自编码器无监督方法对图像预训练。通过卷积神经网络直接与图像像素进行卷积提取图像特征，利用网络权值共享简化模型提高效率，探讨特征提取和分类器联合反馈优化从DR眼底图像中学习特征；收集不少于5000例新发和既往DR患者信息，整合DR眼底图像特征、患者基本信息、糖尿病史和生理生化指标，利用卷积神经网络的输出层进行分类预测，研究无监督卷积神经网络模型构建糖尿病发生DR的六分类预测模型；通过验证样本进行十折交叉内部验证、随访的样本资料进行外部验证。为DR眼底图像的多分类识别提供方法学支撑，为糖尿病导致DR的分类预测提供科学依据。

Up to now, Diabetic retinopathy(DR) is one of the most common causes of blindness in the elderly, correct prediction of DR classification is helpful for early intervention. Traditional DR prediction relies on image features design, which has few classification. Based on our previous experiences in neural networks and disease prediction with medical images, we assume that convolutional neural networks is advanced and suitable methods to solve the problem with fundus images. Denoising autoencoder unsupervised method will be used to study DR fundus images features, convolutional neural network will be used to simplify the network model to improve the efficiency of weights, and explore the characteristics of the extraction and optimization of the joint feedback optimization of learning features from DR fundus images,the feature of fundus image is extracted automatically by convolution neural network and image pixels, the image will be divided into six categories by the consistency analysis. In the study, one shall collect no less than 5000 cases of recent and previous DR patient’s informations，based on the features of DR fundus images, the basic information of patients, the history of diabetes and the physiological and biochemical indexes, the output layer of convolutional neural network will be used to classified predict DR.A multi-class classification model of DR will be studied based on unsupervised convolution neural network model.Ten fold cross validation methods and sample datas for follow-up will be used to perform validation. The study will provide scientific evidence for the fundus image characteristics and prediction of DR, and supply methodological reference for multi classification of medical color images.