基于Batch Normalization的UNet视网膜血管分割算法  

Retinal Vessel Segmentation Algorithm Base on BN-UNet

作　　者：董恒 黄香平 傅诏午 邹梓秀 林涌 Dong Heng;Huang Xiangping;Fu Zhaowu;Zou Zixiu;Lin Yong(School of Computing and Information Science,Fuzhou Institute of Technology,Fuzhou 350506,China;School of Future Technology,Fujian Agriculture and Forestry University,Fuzhou 350002,China)

机构地区：[1]福州理工学院计算与信息科学学院,福建福州350506 [2]福建农林大学未来技术学院,福建福州350002

出　　处：《洛阳师范学院学报》2024年第5期33-37,共5页Journal of Luoyang Normal University

基　　金：福建省中青年教师教育科研项目(JZ230069);福州理工学院校级科研项目(FTKY2023014)。

摘　　要：视网膜血管的提取可以通过卷积神经网络学习的方法,但这种方法容易受到初始图像质量的影响,存在网络学习速度慢、效率低的问题.针对这一问题,提出了一种融合Batch Normalization和UNet卷积神经网络的视网膜血管分割算法.以视网膜图像的三原色通道作为网络的输入数据,然后利用融合批归一化的UNet卷积神经网络对分割模型进行训练,进一步优化网络效率.经过测试和验证,结果表明,在DRIVE公开数据集上,该方法能够更好地提取周围血管结构,灵敏度、准确度、特异性和曲线下面积均较好.The retinal blood vessels can be extracted with convolutional neural network,but this method is easily affected by the initial image quality,and there are problems of slow network learning speed and low efficiency.Aiming at this problem,this research proposes a retinal blood vessel segmentation algorithm combining Batch Normalization and UNet convolutional neural network.This algorithm uses the three primary color channels of the retinal image as the input data of the network,and uses the batch-normalized UNet convolutional neural network to train the segmentation model to further optimize the network efficiency.The results show that the algorithm can better extract the surrounding vascular structure,and optimizes the sensitivity,accuracy,specificity,and area under curve(AUC).

关 键 词：血管分割 视网膜图像 UNet 归一化 

分 类 号：TP39[自动化与计算机技术—计算机应用技术]