This study proposes an improved HRNet based algorithm to solve the common problems of microvascular detail loss and lesion information misjudgment in the existing retinal vascular segmentation algorithms. In the pre-processing stage, the contrast between the blood vessels and the background is improved by contrast-limited adaptive histogram equalization and adaptive Gamma correction. During coding, HRNet original convolution is replaced by deformable convolution to improve the adaptability of convolution to complex vascular morphological structures. Concerning multi-scale feature aggregation, spatial pyramid pooling and multi-scale convolution are introduced to expand the receptive field and enhance the attention to the local features of the target. Consequently, vascular artifacts and subtle information loss can be improved. Simulation on the DRIVE database shows that the accuracy, sensitivity, and specificity of the proposed algorithm are 95.79%, 80.33%, and 98.12%, respectively.