CAMU-Net:基于Attention U-Net的视网膜血管分割改进模型  

CAMU-Net:an improved model for retinal vessel segmentation based on Attention U-Net

作　　者：唐云飞 但志平[1,2] 洪郑天 陈永麟 程沛霖 成果 刘芳婷 TANG Yunfei;DAN Zhiping;HONG Zhengtian;CHEN Yongin;CHENG Peiin;CHENG Guo;LIU Fangting(Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,College of Computer and Information Technology,China Three Gorges University,Yichang 443002,China;Yichang Key Laboratory of Intelligent Medicine,College of Computer and Information Technology,China Three Gorges University,Yichang 430002,China;Tongji Hospital,Tongji Medical College,Huazhong University of Science and Technology,Wuhan 430030,China;Tsinghua Shenzhen International Graduate School,Shenzhen 518055,China;Hong'an County People's Hospital,Huanggang 438400,China)

机构地区：[1]三峡大学计算机与信息学院水电工程智能视觉监测湖北省重点实验室,湖北宜昌443002 [2]三峡大学计算机与信息学院智慧医疗宜昌市重点实验室,湖北宜昌430002 [3]华中科技大学同济医学院附属同济医院,湖北武汉430030 [4]清华大学深圳国际研究生院,广东深圳518055 [5]红安县人民医院,湖北黄冈438400

出　　处：《中国医学物理学杂志》2024年第8期960-968,共9页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(U1401252)。

摘　　要：提出一种改进的U-Net模型(CAMU-Net),以达到精准分割视网膜血管的目的。CAMU-Net模型通过添加残差增强模块来提取区域特征中的重要信息,增强模型对区域特征的了解;通过添加特征细化模块来促进特征的提取,提高新模型的全局特征收集能力;通过添加通道注意力机制模块来捕捉图像特征,精确分割结果;通过引入多尺度特征融合结构来提升模型感知目标边界等细节的能力。在DRIVE数据集上进行消融实验,得出各模块的实际效果,验证各模块对于本模型视网膜血管分割各方面提升的作用;在DRIVE和STARE数据集上和其他主流网络模型进行对比分析,结果表明CAMU-Net模型优于其他模型。An improved U-Net model(channel attention module U-Net,CAMU-Net)is proposed to achieve precise segmentation of retinal vessels.CAMU-Net model enhances its understanding of regional features by employing residual enhancement convolution to extract important information from the regions,improves the global feature acquisition capability by introducing feature refinement module to promote feature extraction,realizes precise segmentation by adding channel attention module to capture image features accurately,and enhances its capability to perceive target boundaries and details through a multi-scale feature fusion structure.The ablation study on the DRIVE dataset validates the role of each module in retinal vessel segmentation.The comparison with other mainstream network models on DRIVE and STARE datasets verify that CAMU-Net model is superior to other models.

关 键 词：视网膜血管 图像分割 深度学习 CAMU-Net 注意力机制 

分 类 号：R318[医药卫生—生物医学工程]