目的 糖尿病性视网膜病变是目前一种比较严重的致盲眼病，因此，对糖尿病性视网膜病理图像的自动分类具有重要的临床应用价值。基于人工分类视网膜图像的方法存在判别性特征提取困难、分类性能差、耗时费力且很难得到客观统一的医疗诊断等问题，本文提出了一种基于卷积神经网络和分类器的视网膜病理图像自动分类系统。方法 首先，结合现有的视网膜图像的特点，对图像进行去噪、数据扩增、归一化等预处理操作；其次，在AlexNet网络的基础上，在网络的每一个卷积层和全连接层之前引入一个批归一化层，得到一个网络层次更复杂的深度卷积神经网络BNnet。BNnet网络用于视网膜图像的特征提取网络，对其训练时采用迁移学习的策略利用ILSVRC2012数据集对BNnet网络进行预训练，再将训练得到的模型迁移到视网膜图像上再学习，提取用于视网膜分类的深度特征；最后，将提取的特征输入一个由全连接层组成的深度分类器将视网膜图像分为正常的视网膜图像、轻微病变的视网膜图像、中度病变的视网膜图像等五类。结果 实验结果表明，本文方法的分类准确率可达93%，优于传统的直接训练方法，且具有较好的鲁棒性和泛化性。结论 本文提出的视网膜病理图像分类框架有效的避免了人工特征提取和图像分类的局限性，同时也解决了样本数据不足而导致的过拟合问题。
Objective Diabetic retinopathy is a serious eye disease causing blindness. The retinal pathological image is an important criterion on which a doctor diagnoses eye disease basing, and the accurate classification of retinal images is a crucial basis for doctors to develop personalized treatment plans. The automated classification of diabetic retinopathy images has significant clinical values. The traditional image classification methods based on manually extracted features have a series of problems, including complex retinal image processing, the discriminative features extraction difficulties, poor classification performance, time-consuming and hard to make objective and consistent diagnoses. In this paper, an improved deep convolutional neural network based on AlexNet and a deep classifier are proposed to realize automated diabetic retinopathy image classification. Method Firstly, since retinal images contain much noise, training retinal samples is insufficient, and the differences between retinal pathological images at adjacent stages is small. These result in a major problem for convolutional neural networks to be employed retinal images classification, and retinal images should be preprocessed before they are used as the training samples. This preprocessing process mainly includes the retinal image denoising, enhancement and normalization. The problems of small number of retinal images and data imbalance at different stages are solved by data enhancement. Secondly, the feature extraction network is designed based on the network structure of AlexNet. Since the data distribution will be changed during the training process, A batch normalization layer is introduced before every convolutional layer and fully connected layer of AlexNet network to produce a new deep network which we called the BNnet network. The introduction of batch normalization layer can accelerate the convergence of the network, improve the recognition rate of the obtained model and reduce the need for dropout layer. The BNnet network is a complex deep convolutional neural network, which not only serves as a feature extraction network for retinal images, but also effectively suppresses the problem of data distribution changes in the training process. In this work, the BNnet network is pretrained by using the ILSVRC2012 dataset based on the transfer learning strategy and the obtained model is migrated to the enhanced diabetic retinopathy dataset for further study to capture the distinguishing features. Finally, a classifier is designed based on the fully connected layer which can map the learned deep features to the sample label space. The classifier is composed of the fully connected layer, the ReLU layer and dropout layer and is applied to learn to partition a diabetic retinopathy status as no DR, mild DR, moderate DR, severe DR and proliferative DR. Result In order to describe the effects of the different depths of neural network, different training methods, the introduction of batch normalized layer and data preprocessing on the experimental results, we have designed four groups of comparative experiments. The experimental results show that the more layers of the network, the more features learned when the training samples are sufficient, and the classification performance of the pretrained network is better than the traditional direct training method. Besides, the proposed BNnet neural network and training method can capture the differences of various stages of diabetic retinopathy and the recognition rate is up to 93% which outperforms other methods. It is shown that the introduction of the batch normalization layer can control data distribution changes during the training process and improve the classification rate. In the case of insufficient retinal image samples, the adoption of transfer learning (TL) and data enhancement strategies is good for extracting deep discriminative features for classification. Hence, a deep classifier composed of fully connected layers can accurately distinguish the stage where a retinal image is located, indicating that the obtained deep features based on BNnet and transfer learning can provide suitable information for classifiers to accurately classify retinal images into five categories. Conclusion We use deep learning methods to achieve automatically classification of retinal images. A new diabetic retinopathy classification framework is presented in this paper, which mainly benefits from three important components: an image preprocessing stage, deep features extraction stage based on transfer learning strategies and BNnet neural networks and a stable classification stage. Intensive dropout and ReLU is used to suppress the over-fitting problem of deep learning algorithm when the training samples are insufficient. With this framework, experimental results show that the deep features combined with the proposed methods can provide suitable information to enable the construction of the most accurate prediction model for predicting diabetic retinopathy status and effectively avoid the limitations of the manual feature extraction and image classification and this method improves the recognition rate and has a better robustness and generalization. In the future studies, we would like to develop a real-time computer-aided diagnosis system for diabetic retina image based on this approach.