Improving Topology Consistency of Retinal Vessel Segmentation via a Double U-Net with Asymmetric Convolution

Retinal vessel segmentation (RVS) plays a significant role in the diagnosis of ocular diseases, like diabetic retinopathy and glaucoma disease. However, many works have neglected keeping the topology consistency of the vascular segmentation, which is more crucial for the clinical diagnosis system. In this paper, we propose a double U-shape network to tackle this problem. The first U-Net architecture is DAS-UNet. With the help of the dense connectivity and the parallel atrous convolution (PAC) block, DAS-UNet can exploit various receptive fields to segment retinal vessel accurately. Through salient computing block (SCB), it can focus more on responsive regions and suppress uncorrelated regions. In addition, we add an auxiliary U-Net which adopts asymmetric convolutions to strengthen the kernel skeleton and correct the connectivity incoherence of retinal vessels. By exploiting the weighted Binary Cross Entropy loss (BCE loss), the double U-shape network can segment retinal vessels more accurately and improve the topological consistency of the segmented vessels. We tested the proposed network for accurate RVS task on DRIVE benchmark, which achieved the SOTA performance with a better segmentation results in terms of topology.