Abstract:

Diabetic Retinopathy（DR） is a leading cause of blindness. In order to diagnose diabetic retinopathy efficiently, deep learning algorithm is applied to detect diabetic retinopathy images. As the plain convolution neural networks can only classify images and the scale of input must be fixed, Region-based Fully Convolutional Networks（R-FCN） object detection algorithm is proposed to classify retinal fundus images and detect lesion regions with any scales. To solve the problem that the original R-FCN algorithm is difficult to detect small objects （minimal retinal hemorrhage and microaneurysm）, the R-FCN algorithm is modified by adding the FPN （Feature Pyramid Networks）, upgrading the backbone network, and modifying the RPN （Region Proposal Network）. The results show that the modified R-FCN algorithm can achieve high accuracy in the classification of retinal fundus images （healthy, mild, moderate, severe and proliferative） and the detection of lesion regions （microaneurysm, retinal hemorrhage and vitreous hemorrhage）.

Key words: deep learning, object detection, convolution neural networks, computer vision, diabetic retinopathy

摘要：

糖尿病眼底病变（Diabetic Retinopathy，DR）是糖尿病患者常见的致盲疾病，可使用深度学习算法对患者的糖尿病眼底图片进行图像识别，实现对糖尿病眼底病变的辅助诊断。针对以往普通卷积神经网络只能进行分类和输入尺寸固定的问题，提出了基于目标检测的区域全卷积网络（Region-based Fully Convolutional Networks，R-FCN）算法，实现同时对任意尺寸输入的糖尿病眼底图片的分类和病变区域检测。针对原始R-FCN算法对小目标（极小的出血点和血管瘤）检测困难的问题，对R-FCN算法做了一定的改进，加入特征金字塔网络（Feature Pyramid Networks，FPN）结构，升级主干网络，修改区域建议网络（Region Proposal Network，RPN）。实现结果表明，改进后的R-FCN算法能以很高的正确率实现对糖尿病眼底图片的五级分类（健康、轻度、中度、重度、增殖）和病变区域检测（血管瘤、眼底出血、玻璃体出血）。

关键词: 深度学习, 目标检测, 卷积神经网络, 计算机视觉, 糖尿病眼底病变