With the increasing number of patients with chronic kidney disease (CDK), the workload of pathologists is heavy and the diagnostic efficiency is low. Therefore, the use of computer technology to assist the diagnosis of nephropathy becomes the trend of future development with a great application space. In recent years, UNet network has been widely used in medical image segmentation, and many improved algorithms have appeared. Most of existing methods focus on adding new modules or incorporating other design concepts on the basis of UNet, however, the structure of UNet network has not been fully analyzed. This paper points out two problems existing in UNet network: insufficient feature extraction in encoder interferes with the accuracy of image segmentation, and the fixed mode of skip connection in each layer leads to information redundancy in the network. To solve these problems, we improve the encoder and feature fusion method of UNet and named the new network as Relay-UNet. Experiment results show that under the condition of almost no change in consumption time, the Dice coefficient of Relay UNet on the glomerular dataset reaches 97.5%, and the training process is more stable than that of UNet.