Automated segmentation of retinal blood vessels plays an important role in the diagnosis of diseases such as diabetes and hypertension. Existing algorithms have insufficient ability to segment blood vessels into small blood vessels and lesions. In this paper, a retinal vessel segmentation method based on an improved holistically nested edge detection (HED) network is proposed to solve the segmentation problem. In the proposed method, firstly a residual deformable convolution block is used instead of the ordinary convolution block to enhance the ability of the model to capture the shape and size of the blood vessel; Subsequently, the original pooling layer is replaced by a dilated convolution layer to preserve the spatial locations of blood vessels; finally, an HED network framework with a short connection structure at the bottom is used for feature extraction and fusion of pre-trained networks, in which the model can better fuse the high-level structural information of the blood vessels and low-level details of the blood vessels in the retinal image extracted by the backbone network. By verifying the digital retinal images for vessel extraction (DRIVE) and the structured analysis of the retina (STARE) datasets, the sensitivities are 81.75% and 80.68%, the specificities are 97.67% and 98.38%, the accuracies are 95.44% and 96.56%, and the area under curve (AUC) of receiver operating (ROC) are 98.33% and 98.12%, respectively. The proposed method achieves comprehensive segmentation performance, which is superior to that of other advanced methods.