计算机科学与探索 ›› 2021, Vol. 15 ›› Issue (11): 2063-2076.DOI: 10.3778/j.issn.1673-9418.2103099

• 综述·探索 • 上一篇    下一篇

深度学习在视网膜血管分割上的研究进展

李兰兰,张孝辉,牛得草,胡益煌,赵铁松,王大彪   

  1. 1. 福州大学 物理与信息工程学院 福建省媒体信息智能处理与无线传输重点实验室,福州 350116
    2. 福州大学 机械工程及自动化学院,福州 350116
    3. 广东省第二人民医院,广州 510317
  • 出版日期:2021-11-01 发布日期:2021-11-09

Research Progress of Deep Learning in Retinal Vessel Segmentation

LI Lanlan, ZHANG Xiaohui, NIU Decao, HU Yihuang, ZHAO Tiesong, WANG Dabiao   

  1. 1. Fujian Key Lab for Intelligent Processing and Wireless Transmission of Media Information, College of Physics and Information Engineering, Fuzhou University, Fuzhou 350116, China
    2. College of Mechanical Engineering and Automation, Fuzhou University, Fuzhou 350116, China
    3. Guangdong Second Provincial General Hospital, Guangzhou 510317, China
  • Online:2021-11-01 Published:2021-11-09

摘要:

视网膜血管分割得到的视网膜特征可以用于辅助糖尿病视网膜病变等眼病的诊断。近年来基于深度学习的血管自动分割算法以自动提取图像特征、精度高、速度快的这些优点吸引了大量研究。对近年基于深度学习的视网膜血管分割研究进行回顾,包括常见的眼底图像数据库、常用的数据增强、图像预处理、图像切片的操作。从网络架构的角度将近期的深度学习血管分割算法归类为级联结构神经网络、多路径神经网络、多尺度神经网络,并对网络进行介绍、对比、性能分析、复杂度分析、缺点分析。同时对于神经网络现实部署的研究也进行了介绍。结果表明,现有眼底图像数据库的数据量还较少,数据增强和图像预处理较多使用方法分别为水平竖直翻转和图像灰度化。从现有研究达到的性能上看,级联结构和多路径的神经网络较为适合视网膜血管的分割;从现有的复杂度来看,部分模型的推断时间可以达到毫秒级,计算消耗可以达到兆以下;从现有算法的缺点看,某个算法只能解决部分现有挑战。在移动设备硬件资源限制的情况下,轻量级的神经网络是一个值得探索的方向。

关键词: 深度学习, 视网膜血管分割, 神经网络

Abstract:

The retinal features obtained by retinal blood vessel segmentation can be used to assist the diagnosis of diabetic retinopathy and other ocular diseases. In recent years, the automatic segmentation algorithm of blood vessels based on deep learning has attracted a lot of research. The reason is that the method can automatically extract image features and has the advantages of high accuracy and fast speed. This paper reviews the research on retinal blood vessel segmentation based on deep learning in recent years. It first discusses the establishment of fundus image databases, commonly used data enhancement, image preprocessing, and image slicing operations. Then, recent deep learning algorithms are classified as cascaded neural network, multi-path neural network, multi-scale neural network in the perspective of network architecture, and these networks are carried out introduction, comparison, performance analysis, complexity analysis and disadvantage analysis. Besides, the introduction of the research on the actual deployment of neural networks is also given. The results show that the data amount in the existing fundus image database is still limited, and the most commonly used methods of data enhancement and image preprocessing are respectively horizontal and vertical flipping and image gray-scaling. Observed from the performance achieved by existing research, cascaded and multi-path neural networks are more suitable for retinal vessel segmentation. Observed from the existing complexity, the inference time of many models can reach the millisecond level, and the computational consumption can reach below million. Observed from the shortcomings of existing algorithms, an algorithm can only solve part of the existing challenges. In the case of mobile device hardware resource constraints, light-weight neural network is a direction worthy of exploration.

Key words: deep learning, retinal vessel segmentation, neural networks