Retinal Vessel Segmentation Algorithm based on Multi-scale U-net Fusion

Abstract

An improved U-net retinal vessel segmentation algorithm was proposed to solve the problems of low accuracy of retinal vessel segmentation, easy breakage and under-segmentation of small vessels in single U-net retinal vessel segmentation algorithm. The improvement points are mainly in the cutting scale of the input image, and the multi-scale method is used, that is, the original retinal vessel image is cut by a variety of different scales. In each scale cutting input to the corresponding input scale U - net network segmentation, the multi-scale prediction results; finally, the prediction results of the scale were fused, and the fusion method could be the maximum value method or the average value method. Experiments show that compared with the basic U-net segmentation method, the proposed method has a certain improvement in the accuracy of retinal blood vessel segmentation, and the segmentation effect of small blood vessels is improved, obtaining a segmentation accuracy of 96.6% on the DRIVE dataset.