一种基于改进Attention U-net的联合视杯视盘分割方法  

A JOINT SEGMENTATION METHOD FOR OPTIC DISC AND OPTIC CUP BASED ON MODIFIED ATTENTION U-NET

作　　者：秦运输 王行甫[1] Qin Yunshu;Wang Xingfu(School of Computer Science and Technology,University of Science and Technology of China,Hefei 230031,Anhui,China)

机构地区：[1]中国科学技术大学计算机科学与技术学院,安徽合肥230031

出　　处：《计算机应用与软件》2021年第3期181-189,共9页Computer Applications and Software

摘　　要：青光眼是当前世界范围内致盲的主要病因之一,其发病过程没有明显的特征。视杯盘比是青光眼诊断中最主要的评估指标之一,这使得视杯视盘的分割成为了目前青光眼诊断的关键。已有的视杯视盘分割方法大多基于手工提取的特征,低效且精度不高。提出一种名为MAR2U-net的深度神经网络架构用于青光眼视杯视盘的联合分割。它是基于Attention U-net的一种改进架构,通过在Attention U-net的基础之上引入递归残差卷积模块来提取更加深层次的特征,并结合多尺度的输入和多标签的Focal Tversky损失函数来提升模型的联合分割性能。实验结果表明,该方法在REFUGE数据集上的分割效果较已有方法取得了显著提升,为实现大规模的青光眼诊断筛查提供了基础。Glaucoma is one of the main causes of blindness in the world,and the pathogenesis has no obvious characteristics.As one of the most important evaluation indexes in the diagnosis of glaucoma,optic disc and cup ratio is the key to the diagnosis of glaucoma.Most of the existing optic disc and cup segmentation methods are based on the manually extracted features,which are inefficient and have low accuracy.This paper proposes a deep neural network named MAR2U-net for the joint segmentation of optic disc and optic cup.It is an improved architecture based on Attention U-net.On the basis of Attention U-net,recursive residual convolution module was introduced to extract deeper features,and the joint segmentation performance of the model was improved by combining multi-scale inputs and multi-label Focal Tversky loss function.The experimental results show that the segmentation effect of proposed method on REFUGE data set has been significantly improved compared with the existing methods,which provides a foundation for large-scale glaucoma diagnosis.

关 键 词：青光眼检测 视杯与视盘 分割 ATTENTION U-net 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]