U-net改进的视网膜血管图像分割算法  

Improved retinal vascular image segmentation algorithm based on U-net

作　　者：王原 马瑜 江妍 梁远哲 马鼎 李霞 WANG Yuan;MA Yu;JIANG Yan;LIANG Yuan-zhe;MA Ding;LI Xia(School of Physics and Electrical Engineering,Ningxia University,Yinchuan 750021,China)

机构地区：[1]宁夏大学物理与电子电气工程学院,宁夏银川750021

出　　处：《计算机工程与设计》2021年第10期2884-2893,共10页Computer Engineering and Design

基　　金：宁夏自然科学基金项目(NZ16009);宁夏高等学校科学研究基金项目(NGY2016015);2018年宁夏研究生教育教学改革研究与实践基金项目(YJG201811);宁夏大学研究生创新研究基金项目(GIP2019060)。

摘　　要：传统抽取算法在病灶、细节区域存在分割不精准的问题。改进算法对U-net深度神经网络进行改进,在网络底层加入Dense-net网络中的稠密连接方式,使用BConvLSTM来组合编码器和解码器的特征信息,结合AC-net思想提出MultiAc模块,在U-net下采样和上采样过程加入该模块,帮助网络学习更复杂的特征信息,在预测过程中提高精确率。通过在DRIVE、STARE、CHASE_DB1这3个公开眼底数据库的实验,分割结果的客观评价指标与主观视觉验证了改进算法在分割精度方面的有效性。The traditional extraction algorithm is not accurate in the segmentation of lesions and details.The improved algorithm improved the U-net depth neural network.Dense connection mode of Dense-net network was added to the underlying join,BConv-LSTM was used to combine the characteristic information of the encoder and the decoder,MultiAc module was put forward combined with AC-net thoughts,and the MultiAc module was used in U-net down-sampling and up-sampling process,which helped the network learn more complex characteristic information to improve accuracy rate in the process of prediction.Through experiments in three public fundus databases including DRIVE,STARE and CHASE_DB1,objective evaluation indexes and subjective vision of segmentation results verified the effectiveness of the improved algorithm in segmentation accuracy.

关 键 词：视网膜血管提取 U型网络 稠密卷积网络 双向卷积长短时记忆网络 不对称卷积网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]