文献类型：会议论文

英文题名：CM-Unet: A Convolutional Neural Network for Retinal Vessel Segmentation

作者：Xu, WenJing[1]; Chen, HongYu[1]; Wu, RongHua[1]; Tao, Chen[1]; Yu, Hui[2]; Liu, HongZhe[3]; Xu, Cheng[3]; Jian, MuWei[4]

第一作者：Xu, WenJing

机构：[1] Linyi University, School of Information Science and Technology, Linyi, China; [2] University of Portsmouth, School of Creative Technologies, Portsmouth, United Kingdom; [3] Beijing Union University, Beijing Key Laboratory of Information Service Engineering, Beijing, China; [4] Shandong University of Finance and Economics, School of Computer Science and Technology, Jinan, China

第一机构：Linyi University, School of Information Science and Technology, Linyi, China

通讯机构：[4]Shandong University of Finance and Economics, School of Computer Science and Technology, Jinan, China

会议论文集：Proceedings - 2023 IEEE SmartWorld, Ubiquitous Intelligence and Computing, Autonomous and Trusted Vehicles, Scalable Computing and Communications, Digital Twin, Privacy Computing and Data Security, Metaverse, SmartWorld/UIC/ATC/ScalCom/DigitalTwin/PCDS/Metaverse 2023

会议日期：August 28, 2023 - August 31, 2023

会议地点：Portsmouth, United kingdom

语种：英文

外文关键词：Attention mechanism; Dense connections; Retinal vessel segmentation; U-Net

摘要：Accurate segmentation of retinal vessels from retinal images is critical for the detection and diagnosis of many ocular diseases and associated cardiovascular diseases. However, the complex structure of retinal vessels, the low contrast between the background and the vessels, and the various noises generated by medical imaging systems make retinal vessel segmentation a great challenge. To address these challenges, we propose an architecture called CM-Unet for accurate segmentation of retinal blood vessels by improving the U-Net medical image segmentation model in this paper. Specifically, we achieve cross-layer cascading and reuse of features by constructing dense connections in the encoder part. In addition, we introduce global context blocks to filter noise in low-level semantic information and suppress irrelevant features to make the network more focused on the target feature region. Meanwhile, we use depthwise separable convolution instead of standard convolution to reduce the number of parameters. Experimental results of vascular segmentation on DRIVE and ARIA datasets show that the proposed method has significantly improved the segmentation accuracy demonstrating its effectiveness in vascular segmentation tasks. ? 2023 IEEE.