文章摘要

文章信息/Info

Title:: Research oncolor fundus image blood vessel segmentation based on Dense U-net

作者:: 张亮军1; 承垠林1; 马梦楠1; 马丽2; 周毅3△; 1.中山大学生物医学工程学院，广州 510006; 2.中山大学中山眼科中心，广州 510080； 3.中山大学中山医学院医学信息教研室, 广州 510080

Author(s):: ZHANG Liangjun1; CHENG Yinlin1; MA Mengnan1; MA Li2; ZHOU Yi3; 1. School of Biomedical Engineering, Sun Yat-Sen University, Guangzhou 510006,China;2.Zhongshan Ophthalmic Center, Sun Yat-Sen University, Guangzhou 510080;3. Department of Biomedical Engineering, Zhongshan School of Medicine, Sun Yat-Sen University, Guangzhou 510080

Keywords:: Retinal fundus image; U-net; Blood vessel segmentation; Neural network; Deep learning

摘要:: 眼底彩色照片图像血管分析可用于评估和监测各种眼科疾病，对患者提前干预治疗提供指导，减少致盲风险，具有十分重要的临床意义。目前眼底血管分割的算法及模型，对于细小血管如视网膜毛细血管的分割效果仍有待提高。本研究针对此问题提出Dense U-net网络架构，该模型在U-net网络中加入了Dense Block架构，可以提高细小血管的分割准确率，同时该算法模型在DRIVE（digital retinal images of vessel extration，DRIVE）公开数据集上进行了实验，相比现有的算法，本研究模型的准确率、灵敏度、特异性分别为0.9532、0.7977、0.9759，其中灵敏度的提高可以使得算法模型更准确地识别并分割出细小血管。

Abstract:: Retinal fundus image analysis can be used to assess and monitor various ophthalmic diseases, provide guidance for patients with advanced intervention, and reduce the risk of blindness, which has a great clinical significance. At present, the algorithm and model for retinal fundus segmentation for small blood vessels such as retinal capillaries remain to be improved. We proposed the Dense U-net network architecture for this problem. This model added the Dense Block framework to the U-net network, which can improve the segmentation precision of small blood vessels. At the same time, this algorithm model has tested on the DRIVE public dataset, compared with the existing algorithms, the proposed method achieved higher accuracy, sensitivity, and specificity of 0.9532, 0.7977, 0.9759. The improvement of accuracy allows the algorithm model to identify and segment small blood vessels more accurately.