Abstract: To solve the problems of difficult diagnosis of diabetic retinopathy and inconsistent evaluation criteria in different places, a retinal microvascular segmentation algorithm was proposed based on attention mechanism and dense convolution, that is, image segmentation technology was used to assist the diagnosis, which not only reduced the workload but also ensured the accuracy. Using LadderNet as the basis network, and to highlight the microvascular information more, attention mechanism was added to make the characteristic information of microvascular more complete and more accurate. Dense convolution was used to enhance the transmission of feature information, to reduce the number of parameters, and to further improve the performance of image segmentation. The algorithm proposed has better segmentation performance and can better complete the task of retinal microvascular segmentation.