23. Retinal Vessels Segmentation Based on Multi-scale Hybrid Convolutional Network

Retinal fundus image can reveal the information on the early symptoms of diabetes, hypertension, hyperlipidemia and other diseases. Accurate segmentation of retinal vessels can assist the detection and diagnosis of the related diseases. Due to the intricate characteristic information of retinal vessels images, traditional segmentation methods lead to inaccurate segmentation for small vessels and pathological segmentation errors. In this paper, we propose a new multi-scale hybrid convolution U-Net. Firstly, we take the hybrid convolution module by combining dilated convolution and standard convolution as the core structure for feature extraction to obtain more abundant semantic feature information, while expanding the receptive field. Then, we add a multi-scale fusion module to the network encoding and decoding connection part, which fuses the feature information of different layers to reduce the loss of information and enhance the representation ability of the network. We evaluate the performance of the proposed method on two public retinal datasets (DRIVE and CHASE_DB1). The results of quantitative and qualitative experiments show that the proposed model can improve good accuracy in retinal vessels segmentation.