Abstract:The segmentation accuracy of retinal vessels has an important impact on the early diagnosis of ophthalmic diseases and diabetes. Facing the problem of poor segmentation performance of existing methods in microvessel and lesion regions, the paper proposes a segmentation model MRAU-net that maximizes the extraction of vascular information. The model introduces a multiscale feature extraction residual module and a multilevel residual null convolution layer at the encoding site, which is used to extend the perceptual field, learn multilevel image features and improve the utilization of vascular information by the model; the downsampling and short connection sites are incorporated into the CBAM attention mechanism and multi-channel attention module are incorporated to increase the recognition of blood vessels by the model and reduce the possibility of mis-segmentation. In this paper, experiments are conducted based on two publicly available datasets, DRIVE and STARE, to verify the segmentation capability of the improved model. The conclusions show that the accuracy (Acc) is 0.9652 and 0.9715 on both data, and the sensitivity (Se) is 0.8205 and 0.8256, respectively, which are more advantageous than other algorithms in terms of segmentation performance.