Simultaneous Artery-vein Segmentation in Fundus Images Using an Improved Fully Convolutional Network

The segmentation and classification of retinal artery and vein form an important metric in the computer-aided diagnosis of various systemic and vascular diseases. Most of the existing arteriovenous classification methods heavily depend on the accurate segmentation of the vascular structure. In this paper, we propose an improved fully convolutional neural network that can simultaneously segment and classify vessels into arteries and veins. The deformable convolution and residual path are integrated into the U-shape architecture to extract context information and eliminate the semantic gap while combining low-level features with high-level ones. The proposed network achieves an accuracy of 95.6% for vessel classification and 95.89% for arteriovenous classification, which are comparable with those achieved with the state-of-the-art methods on the DRIVE dataset.