基于多尺度U型网络的图像分割方法

Image Segmentation Algorithm Based on Multi-scale U-shaped Network

眼底血管图像分割对糖尿病、心脏病和高血压等疾病诊断具有重要价值。眼底图像特征复杂度较高,当前分割算法存在分割不精细、细小血管特征提取难、细节丢失多等问题。为了解决这些问题,设计了多尺度特征U型网络框架,该网络在U-Net的基础上,在跳跃连接结构上引入金字塔池化整合多个尺度的特征,在编解码结构中用Leaky ReLU函数使负值特征能够得到激活。实验结果表明,该网络改善了细小血管的分割效果,分割精确度、敏感性、特异性分别达到了96.21%,74.02%和97.88%,对辅助医生的诊断具有明显的意义。

Retinal vessel image segmentation is of great value in the diagnosis of diabetes,heart disease,hypertension and other diseases.Considering the high feature complexity of retinal images,the current segmentation algorithms have some problems such as imprecise segmentation,difficulty in extracting small blood vessel features,and loss of details.To solve these problems,a U-shaped network framework with multi-scale features is designed.Based on U-Net,pyramid pooling is introduced into the jump connection structure to integrate multi-scale features,and Leaky ReLU function is used in the encoder-decoder structure to activate negative features.The experimental results show that the network improves the segmentation effect of small blood vessels,and the segmentation accuracy,sensitivity and specificity reach 96.21%,74.02%and 97.88%respectively,which if of obvious significance for assisting doctors in diagnosis.