基于深度卷积神经网络的糖尿病视网膜病变分期及病灶检测

Staging and lesion detection of diabetic retinopathy based on deep convolution neural network

针对糖尿病视网膜病变(DR)图像分辨率过大、病灶特征过于分散难以获取以及正负难易样本不平衡而导致DR分期精确率一直无法得到有效提高的问题,提出了改进的基于快速区域的卷积神经网络(Faster R-CNN)和子图分割相结合的DR分期方法。首先,使用子图分割解决视盘区域对于病灶识别的干扰问题;其次,在特征提取阶段使用深度残差网络以解决病灶在高分辨率眼底图像中占比小而导致的特征难以获取的问题;最后,在感兴趣区域(ROI)生成时采用在线困难样本挖掘(OHEM)方法解决正负难易样本不平衡的问题。在国际公开数据集EyePACS进行DR分期实验,所提方法在DR病分期中精确率0期达到94.83%,1期达到86.84%,2期达到94.00%,3期达到87.21%,4期达到82.96%。实验结果表明,改进后的Faster R-CNN能对DR图像高效分期并自动标注出病灶。

For Diabetic Retinopathy(DR),the image resolution is too high,the lesion features are too scattered to obtain,and the positive,negative,hard and easy samples are imbalanced,thus the DR staging accuracy cannot be effectively improved.Therefore,a DR staging method based on the combination of improved Faster Region-based Convolutional Neural Network(Faster R-CNN)and subgraph segmentation was proposed.First,subgraph segmentation was used to solve the interference problem of the optic disc region to lesion recognition.Second,a deep residual network was used in the feature extraction process to solve the problem of difficulty of obtaining features due to the small proportion of the lesions in the high-resolution fundus image.Finally,the Online Hard Example Mining(OHEM)method was used to solve the problem of imbalance between positive,negative,hard and easy samples during the generation of Region of Interest(ROI).In the DR staging experiments on EyePACS,an internationally open dataset,the accuracy of the proposed method in DR staging reached 94.83%in stage 0,86.84%in stage 1,94.00%in stage 2,87.21%in stage 3 and 82.96%in phase 4.Experimental results show that the improved Faster R-CNN can efficiently stage DR images and automatically label the lesions.