基于深度学习的视网膜血管分割方法综述  

A Survey of Retinal Vessel Segmentation Algorithms Based on Deep Learning

作　　者：吕佳[1,2] 王泽宇 LÜ Jia;WANG Zeyu(National Center for Applied Mathematics in Chongqing,Chongqing Normal University;College of Computer and Information Sciences,Chongqing Normal University)

机构地区：[1]重庆师范大学重庆国家应用数学中心 [2]重庆师范大学计算机与信息科学学院,重庆401331

出　　处：《重庆师范大学学报（自然科学版）》2024年第4期110-125,共16页Journal of Chongqing Normal University:Natural Science

基　　金：重庆市教育委员会科研项目重点项目(No.KJZD-K202200511);重庆市科技局技术预见与制度创新项目(No.CSTB2022TFII-OFX0044)。

摘　　要：通过检索2016年至2024年来基于深度学习的视网膜血管分割方法相关文献,总结基于深度学习的视网膜血管分割方法并梳理现有方法存在的问题,为视网膜血管分割任务的进一步研究提供参考。首先,介绍了视网膜血管分割的背景。其次,从相关文献中总结了视网膜血管分割方法的数据集及数据预处理、评价指标、分割方法等。在数据集及数据预处理部分,介绍了视网膜图像数据集,总结了常用的特征增强技术、数据增强技术;在评价指标部分,介绍了视网膜血管分割方法常用的评价指标;在分割方法部分,将现有方法归纳为以下3类:基于提高特征提取能力的方法、基于血管特性的方法及基于实际应用的方法,其中基于提高特征提取能力的方法可分为感受野受限的方法、感受野扩展的方法以及特征细化的方法,基于血管特性的方法常利用多标签、损失函数等来改善粗细血管差异或类别不平衡的问题,基于实际应用的方法则更关注网络的轻量化及泛化能力。最后,分析了现有视网膜血管分割方法存在的问题,提出未来可能的研究方向。在未来一段时间,基于深度学习的视网膜血管分割方法仍将是医学图像处理研究的重点和热点之一,本文对于了解视网膜血管分割任务的现状和解决实际应用问题均具有一定价值。Deep learning-based retinal vessel segmentation methods were summarized,while the problems of existing methods were sorted out to provide reference for further research on retinal vessel segmentation tasks.Literature related to deep learning-based retinal vessel segmentation methods from 2016 to 2022 was searched,summarized and generalized.First,it introduces the background of retinal vessel segmentation.Then,it summarizes the dataset and data preprocessing,evaluation indexes,and segmentation methods in the related literature.In the dataset and data preprocessing section,commonly used retinal image datasets are introduced,and commonly used feature enhancement techniques and data enhancement techniques are summarized.In the evaluation indexes section,commonly used evaluation indexes of retinal vessel segmentation methods are presented.In the section of segmentation methods,it categorizes these methods into those that consider network feature extraction ability,those that consider vessel characteristics,and those that consider practical applications.Among them,the methods considering the feature extraction ability of the network can be classified into methods with restricted receptive fields,methods with extended receptive fields,and methods with feature refinement,the methods considering vascular characteristics often use multi-labels,loss functions,etc.to improve the problem of the difference between thick and thin vessels or the imbalance of categories;the methods considering practical applications pay more attention to the light-weighting and generalization ability of the network.Finally,it organizes and summarizes the problems of retinal vessel segmentation methods and possible future research directions.Retinal vessel segmentation methods based on deep learning will remain one of the focuses and hotspots of medical image processing research for some time in the future,and it is valuable for understanding the current status of the retinal blood vessel segmentation task and solving practical application problem

关 键 词：医学图像处理 视网膜血管分割 深度学习 U-Net 感受野 特征细化 轻量化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]