Two-Stage Retinal Vessel Segmentation Based on Improved U-Net

Aiming at the problems of current retinal image segmentation methods, such as blurring of fine vessels pixels and loss of details at the edge of thicker retinal vessels, we designed a two-stage segmentation method based on the combination of improved U-Net and Mini-U-Net network in this paper. Firstly, we added a small-size Mini U-Net to the full-size U-Net network to perform secondary training on the blurred vessel pixels in the retinal image to improve the segmentation effect of blurred blood vessel pixels; secondly, the original convolutional layer in the encoding and decoding processes of the two networks is changed to a residual convolution module to preserve the original feature information more completely; finally, an attention mechanism is introduced at the jump connection of the two networks to improve fine vessels segmentation precision. The precision ratio of this method on the DRIVE and STARE public fundus image datasets were 0. 8331 and 0. 8563, the recall rates were 0. 8396 and 0.8639, the F1-Score were 0.8351 and 0.8609, and the accuracy rate were 0.9698 and 0. 9787, respectively. The total segmentation results of the proposed method are better than those of other methods.