基于U-Net网络改进算法的视网膜血管分割研究北大核心CSCD

针对视网膜图像血管细小,细节特征丢失、梯度下降、爆炸而导致分割效果差的问题,本文提出了一种引入残差块、循环卷积模块和空间通道挤压激励模块的U-Net视网膜血管图像分割模型。首先通过使用一系列随机增强来扩展训练集并对数据集进行预处理,然后在U-Net模型中引入残差块,避免随着网络深度增加,分割准确率达到饱和然后迅速退化以及优化计算成本;并将U-Net网络的底部替换为循环卷积模块,提取图像低层次的特征,并不断的进行特征积累,增强上下文之间的语义信息,获得更有效的分割模型;最后在卷积层之间嵌入空间通道挤压激励模块,通过找到特征较好的通道,强调这一通道,压缩不相关的通道使得网络模型能够加强关键语义特征信息的学习,通过训练过程学习到有效的特征信息,同时增强抗干扰能力。通过在DRIVE数据集上的验证结果可得,本文所提模型的准确率为98.42%,灵敏度达到了82.36%,特异值达到了98.86%。通过和其他网络分割方法比较,本文所提分割方法具有更优的分割效果。

Research on retinal vessel segmentation based on U-Net network improved algorit hm

To solve the problem of poor segmentation effect caused by small blood vessels in retinal images,loss of detailed features, gradient descent and explosion,a U-Net retinal vascular image segmentation model with residual block,cyclic convolution module and spatial channel extrusion excitation module is proposed.First,the training set is expanded by using a series of random enhancements,and then residual blocks are introduced into the U-Net model to avoid the segmentation accuracy reaching saturation and then rapidly degrading as the network depth increases.The bottom of the U-Net is replaced with a circular convolution module,the low-level features of the image are extracted,and features are continuously accumulated,the semantic information between contexts is enhanced,and a more effective segmentation model is obtained.Finally,the concurrent spatial and channel squeeze and channel excitation module is embedded between the convolutional layers.The excitation module finds the channel with stronger characteristic signal, and emphasizes this channel, compresses irrelevant channels,and reduces the interference of irrelevant characteristic information.Through the verification results on the DRIVE data set,the accuracy of the model proposed in this paper is 98.42%,the sensitivity reaches 82.36%,and the specific value reaches 98.86%.Compared with other network segmentation methods,the segmentation method proposed in this article has better segmentation effect.