基于深度学习的OCT图像视网膜积液自动分割  

Automatic segmentation of retinal fluid in OCT images using deep learning

作　　者：魏静 江旻珊[1] 茅前 WEI Jing;JIANG Minshan;MAO Qian(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《光学仪器》2021年第3期29-35,共7页Optical Instruments

摘　　要：糖尿病性黄斑水肿(DME)是视网膜破损导致液体积聚的一种眼底疾病,是老年人视力丧失的主要原因之一。利用眼底光学相干断层扫描(OCT)图像检测黄斑区积液在选择DME治疗方案和评估治疗效果方面发挥着重要的作用。为此,提出了一种改进的U-net网络Res-SE Unet实现在OCT图像上自动分割视网膜内积液(IRF)和视网膜下积液(SRF)。该网络使用Res-SE Block替代标准卷积层,增强网络对有效特征的提取。利用Kermany数据集训练和评估Res-SE Unet模型,通过Dice系数和IoU评价模型分割效果。IRF的平均Dice系数和平均IoU分别为0.84和0.72,SRF的平均Dice系数和平均IoU分别为0.86和0.74,结果表明Res-SE Unet网络可以有效分割IRF和SRF。Diabetic macular edema(DME),a fundus disease caused by retinal damage and fluid accumulation,is one of the main causes of vision loss in the elderly.Detection of macular effusion using fundus OCT images plays an important role in the selection of DME treatment options and evaluation of treatment effects.This paper proposed an improved U-net network,Res-SE Unet,to automatically segment intraretinal fluid(IRF)and subretinal fluid(SRF)in OCT images.The network uses Res-SE Block instead of the standard convolutional layer to enhance the network’s extraction of effective features.This study used Kermany data set to train and test Res-SE Unet and evaluated the model by Dice coefficient and IoU.The average Dice coefficient and average IoU of IRF were 0.84 and 0.72,and the average Dice coefficient and average IoU of SRF were 0.86 and 0.74,respectively,indicating that the Res-SE Unet network can effectively segment IRF and SRF.

关 键 词：糖尿病性黄斑水肿 深度学习 U-net 视网膜内积液(IRF) 视网膜下积液(SRF) 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]