Title:HybridNetSeg: A Compact Hybrid Network for Retinal Vessel Segmentation

Abstract:A large number of retinal vessel analysis methods based on image segmentation have emerged in recent years. However, existing methods depend on cumbersome backbones, such as VGG16 and ResNet-50, benefiting from their powerful feature extraction capabilities but suffering from high computational costs. In this paper, we propose a novel neural network (HybridNetSeg) dedicated to solving this drawback while further improving overall performance. Considering deformable convolution can extract complex and variable structural information, and larger kernel in mixed depthwise convolution makes contribution to higher accuracy. We have integrated these two modules and propose a Hybrid Convolution Block (HCB) using the idea of heuristic learning. Inspired by the U-Net, we use HCB to replace a part of the common convolution of the U-Net encoder, drastically reducing the parameter count to 0.71M while accelerating the inference process. Not only that, we also propose a multi-scale mixed loss mechanism. Extensive experiments on three major benchmark datasets demonstrate the effectiveness of our proposed method