基于改进U-Net的眼底图像血管分割方法

Automatic Segmentation for Retinal Vessel Based on Improve U-Net

眼底血管图像分析可用于各种眼病的评估和监测。它在降低失明风险方面发挥着重要作用。目前,许多眼底血管分割模型在小血管的分割结果上仍需改进。针对上述问题,我们提出了一种改进U-Net模型的视网膜血管分割方法。首先,利用网状跳跃连接提取浅层到深层的特征映射。然后将特征图拼接融合,最大限度地发挥它们的作用。在特征层中,我们采用空洞卷积来增加接受野。在DRIVE、STARE、CHASE_DB1三个标准数据集上验证了该方法的分割精确度分别为0.9595,0.9716,0.9638。实验结果表明,该方法是一种良好的视网膜血管分割方法,比现有的许多视网膜血管分割方法具有更好的分割效果。

Fundus image vessel analysis can be used to evaluate and monitor a variety of eye diseases.It play an important role in reducing the risk of blindness.At present,many model of fundus vessel segmentation still need to be improved for the segmentation results of small vessels.To address above issue,we propose a novel method based on improved U-Net to segment retinal vessels.Firstly,we use the nested skip connections to extract feature maps from shallow layer to deep layer.Then concatenate and fuse the feature maps to make the utmost of them.And in feature layer,we use atrous convolution to increase the receptive field.By verifying this method on three standard datasets,DRIVE,STARE,CHASE_DB1,the segmentation accuracy of proposed method are 0.9595,0.9716,0.9638 respectively.The result can demonstrate the proposed method has improvement on retinal vessels segmentation and can segment retinal vessels better than many existing retinal vessels segmentation methods.