Dense-Atrous U-Net with salient computing for Accurate Retina Vessel Segmentation

Retina vessel segmentation is a fundamental step in the diagnosis of diabetic retinopathy. U-Net is one of the most popular methods for medical image segmentation. However, the consecutive pooling and convolution operations lead to the loss of spatial information. In this paper, we propose a densely connected U-Net with parallel atrous convolution block and salient computing block, which is named DAS-UNet. There are several advantages of DAS-UNet for vessel segmentation. First, the dense connectivity strengthen feature propagation. Second, the parallel atrous convolution block (PAC Block) employs different receptive fields to generate more abstract features for accurate segmentation. Third, we can highlight responsive region and suppress irrelevant regions through salient computing block (SCB), which allows us to obtain a clear vessel segmentation map than the U-Net. Comprehensive experiments on DRIVE benchmark indicates that DAS-UNet achieves a state-of-the-art performance with a significantly clear result.