The morphological changes in retina boundaries are important indicators of retinal diseases, and the subtle changes can be captured by images obtained by optical coherence tomography (OCT). The retinal layer boundary segmentation based on OCT images can assist in the clinical judgment of related diseases. In OCT images, due to the diverse morphological changes in retina boundaries, the key boundary-related information, such as contexts and saliency boundaries, is crucial to the judgment and segmentation of layer boundaries. However, existing segmentation methods lack the consideration of the above information, which results in incomplete and discontinuous boundaries. To solve the above problems, this study proposes a coarse-to-fine method for the segmentation of retinal layer boundary in OCT images based on the end-to-end deep neural networks and graph search (GS), which avoids the phenomenon of “faults” common in non-end-to-end methods. In coarse segmentation, the attention global residual network (AGR-Net), an end-to-end deep neural network, is proposed to extract the above key information in a more sufficient and effective way. Specifically, a global feature module (GFM) is designed to capture the global context information of OCT images by scanning from four directions of the images. After that, the channel attention module (CAM) and GFM are sequentially combined and embedded in the backbone network to realize saliency modeling of context information of the retina and its boundaries. This effort effectively solves the problem of wrong segmentation caused by retina deformation and insufficient information extraction in OCT images. In fine segmentation, a GS algorithm is adopted to remove isolated areas or holes from the coarse segmentation results obtained by AGR-Net. In this way, the boundary keeps a fixed topology, and it is continuous and smooth, which further optimizes the overall segmentation results and provides a more complete reference for medical clinical diagnosis. Finally, the performance of the proposed method is evaluated from different perspectives on two public datasets, and the method is compared with the latest methods. The comparative experiments show that the proposed method outperforms the existing methods in terms of segmentation accuracy and stability.