基于改进U-Net的视网膜血管分割

Image Segmentation of Retinal Blood Vessels Based on Improved U-Net

针对视网膜血管细小部分分割困难与不准确等问题,在原U-Net网络中加入改进的Inception模块与级联空洞卷积融合的结构、注意力机制和可逆金字塔结构等模块,提出一种改进U-Net视网膜血管分割图像的算法。用改进的Inception与不同空洞率的空洞卷积、循环卷积网络进行融合,替换原来的卷积层。在编码部分和解码部分之间加入可逆特征金字塔结构,提高了前后层之间的联系,在模型的最底层采用级联空洞卷积的结构扩大感受野的特征,从多个尺度更好地提取细小的血管信息。在连接层之间加入改进后的注意力机制,让模型更加有目的地去提取特征信息,通过混合的损失函数来提高模型的分割效率。改进后的模型在公开数据集DRIVE、STARE的准确率为98.47%、98.76%,较原来U-Net模型有较大提升。

To address the difficulty and inaccuracy in segmentation of small retinal blood vessels,an improved U-Net algorithm for retinal blood vessel segmentation is proposed,in which a structure combing an improved Inception module and cascaded dilated convolution,improved attention mechanism and a reversible pyramid structure are added to the original U-Net network.The improved Inception module is fused with the dilated convolution with different dilation rates and the cyclic convolution network to replace the original convolution layer.A reversible pyramid structure is added between the encoding and decoding parts to improve the connection between the front and rear layers.The cascaded dilated convolution structure is used at the bottom of the model to expand the special diagnosis of the receptive field and extract fine vascular information from multiple scales.The improved attention mechanism is added between the connection layers to enable the model to extract feature information more purposefully.Finally,the segmentation efficiency of the model is improved by the mixed loss function.The accuracy of the improved model on open data sets DRIVE and STARE is 98.47%and 98.76%respectively,which shows a significant improvement over the original U-Net model.