基于全局与局部信息融合的视网膜血管分割  

Retinal Vessel Segmentation Based on Global and Local Information Fusion

作　　者：王倩 辛月兰 WANG Qian;XIN Yue-lan(School of Physical and Electronic Information Engineering,Qinghai Normal University,2.State Key Laboratory of Tibetan Inteligent Information Processing and Application,Xining Qinghai 810001,China)

机构地区：[1]青海师范大学物理与电子信息工程学院,青海西宁810001 [2]省部共建藏语智能信息处理及应用国家重点实验室,青海西宁810001

出　　处：《计算机仿真》2024年第11期244-250,共7页Computer Simulation

基　　金：国家自然科学基金项目(61662062);青海省自然科学基金面上项目(2022-ZJ-929)。

摘　　要：针对视网膜血管分割方法中U-Net网络捕获全局上下文信息效率不高的问题,提出了一种融合全局信息和局部信息的方法,即在提取特征的同时提取局部细节特征和全局上下文特征,并将二者有效融合。首先,在U-Net网络编码器部分添加一个与原卷积层并行的金字塔视觉Transformer分支,用于提取全局上下文信息;其次,在编码器底部引入了BiFusion模块,将U-Net网络卷积层提取到的局部信息与Transformer分支提取到的全局信息进行有效融合,以获取到更加完整的特征信息。在FIVES与OCTA-500数据集上的实验结果表明,上述网络的DICE系数分别达到了90.09%和83.28%,并且视网膜分割效果也了明显提高。Aiming at the problem of low efficiency of U-Net network in capturing global contextual information in retinal vessel segmentation methods,a method of integrating global and local information is proposed,which extracts local detail features and global contextual features while extracting features,and effectively fuses the two.Firstly,a pyramid vision Transformer branch parallel to the original convolution layer is added to the U-Net network encoder to extract global context information.Secondly,BiFusion module is introduced at the bottom of the encoder to effectively integrate the local information extracted from the U-Net convolutional layer with the global information extracted from Transformer branch,so as to obtain more complete feature information.The experimental results on FIVES and OCTA-500 data sets showed that the DICE coefficients of the network reached 90.09%and 83.28%,respectively,and the retinal segmentation effect was significantly improved.

关 键 词：全局信息 局部信息 网络 分支 视网膜血管分割 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]