视网膜OCT图像中液体分割方法的研究  

Research on segmentation method of OCT retinal image fluid

作　　者：王腾 陈明惠[1] 柯舒婷 袁媛[1] 赖湘玲 黄鐸文 刘渡新 马昕宏 WANG Teng;CHEN Minghui;KE Shuting;YUAN yuan;LAI xiangling;HUANG Duowen;LIU Duxin;MA Xinhong(Shanghai Engineering Research Center of Interventional Medical,Shanghai Institute for Interventional medical devices,Department of Biomedical Engineering,School of Health Science and Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学健康科学与工程学院生物医学工程系上海介入医疗器械工程技术研究中心教育部医学光学工程中心,上海200093

出　　处：《光学技术》2022年第3期364-371,共8页Optical Technique

基　　金：上海市科委产学研医项目(15DZ1940400)。

摘　　要：光学相干层析成像(OCT)广泛应用于眼科,用于观察视网膜的形态,对病变区域的检测和诊断评估具有重要意义。由于液体的积聚引发的一系列视网膜疾病,针对视网膜OCT图像中的液体检测和分割问题,设计了一种具有全局上下文特征信息的神经网络,利用多尺度的特征提取与融合的方法,提出了一种多尺度的并行提取与高度融合的U型网络模型PH-UNet,这是一种新的用于视网膜OCT图像液体区域分割的深度卷积神经网络。PH-UNet网络通过捕获多尺度的上下文信息,更好地利用信息提取和融合的方法对视网膜OCT图像中液体区域进行端到端的分割。将提出的模型对MICCAI RETOUCH数据集中三种视网膜液体视网膜内液体(IRF)、视网膜下液体(SRF)和色素上皮脱离(PED)进行分割并与其他经典的分割网络模型进行了比较,在precision精确率、dice相似系数、mIoU平均交并比三种指标上取得了最佳效果证明了其优越性。Optical coherence tomography(OCT)is widely used in ophthalmology to observe the morphology of the retina,and is of great significance for the detection and diagnostic evaluation of lesions.Due to series of retinal diseases caused by liquid accumulation,a neural network with global context feature information is designed for liquid detection and segmentation in retinal OCT images.By means of multi-scale feature extraction and fusion,a multi-scale parallel extraction and highly integrated U network model PH-UNet is proposed,which is a new deep convolutional network for liquid area segmentation in OCT images.PH-UNet network captures multi-scale contextual information,and better utilizes information extraction and fusion methods to perform end-to-end segmentation of the liquid area of the OCT image.The proposed model is segmented on three types of retinal fluid.The intraretinal fluid(irf),subretinal fluid(srf),and pigment epithelial detachment(ped)are segmented and compared with other classic segmentation network models.The best results have been achieved on the three indexes of precision,dice and mIoU,which proves its superiority.

关 键 词：光学相干层析成像 液体分割 空洞卷积 特征融合 多尺度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]