目的：对视网膜光学相干断层扫描图像中不同层和积液区域的分割。方法：提出一种基于深度学习的轻量级的神经网络，参考DRUNet体系、膨胀卷积和残差网络的架构，通过连接不同深度网络处得到的上采样输出，进行多尺度特征融合，使网络能够更好地识别出图像中的边界信息。结果：改进型DRUNet显著提升了视网膜分层的效果，准确率较U-Net提高了1.25%，同时能提前1~2次迭代达到传统U-Net的准确度。结论：本文采用的网络结构提高了对视网膜光学相干断层扫描图像的分割性能，同时降低了网络参数，具有强大的应用潜力。

Objective: To achieve the segmentation of different layers and fluid areas on the optical coherence tomography (OCT) image of the retina. Methods: A lightweight neural network based on deep learning was proposed. The network structure adopted in this study was designed based on the architecture of dilated-residual U-Net. By connecting the upsampling output obtained at different depth networks, multi-scale feature fusion was performed to enable the system to accurately identify the boundaries on the OCT image. Results: Compared with U-Net, this algorithm could achieve the same accuracy with 1–2 epochs less, and the accuracy was also improved by 1.25%. Conclusion: The proposed network improves the segmentation performance of retinal OCT images, and reduces the number of parameters, which demonstrates the network has great application potential.