糖尿病视网膜病变眼底图像分类方法

1)深圳大学医学部,广东深圳 518060; 2)广东省医学信息检测与超声成像重点实验室,广东深圳 518060; 3)医学超声关键技术国家地方联合工程实验室,广东深圳 518060; 4)中山大学附属第一医院眼科,广东广州 510275

图像处理; 眼底图像; 糖尿病视网膜病变; 计算机辅助诊断; 自动检测; 图像分类

Classification methods for diabetic retinopathy fromretinal images
Liang Ping1, Xiong Biao1,2,3, Feng Juanjuan4, Liao Ruiduan4, Wang Tianfu1,2,3, and Liu Weixiang1,2,3

Liang Ping1, Xiong Biao1,2,3, Feng Juanjuan4, Liao Ruiduan4, Wang Tianfu1,2,3, and Liu Weixiang1,2,31)Health Science Center, Shenzhen University, Shenzhen 518060, Guangdong Province, P.R.China2)Guangdong Key Laboratory for Biomedical Measurements and Ultrasound Imaging, Shenzhen 518060, Guangdong Province, P.R.China3)National-Regional Key Technology Engineering Laboratory for Medical Ultrasound, Shenzhen 518060, Guangdong Province, P.R.China4)Ophthalmology Department of First Affiliated Hospital of Sun Yat-sen University, Guangzhou 510275, Guangdong Province, P.R.China

image processing; fundus images; diabetic retinopathy(DR); computer-aided diagnosis; automatic detection; image classification

DOI: 10.3724/SP.J.1249.2017.03290

备注

评述糖尿病视网膜病变(diabetic retinopathy,DR)眼底图像自动分类方法的研究进展.介绍基于局部病灶的分类方法和基于全局图像的分类方法.其中,基于局部病灶的分类方法主要是渗出物、出血点和微血管瘤病灶的检测,根据检测出的病灶类型、数量和位置等信息进行DR分类; 基于全局图像的分类方法是对图像全局特征信息进行分类.分析了常用数据集、各类方法优缺点和分类性能,指出尽管DR眼底图像自动分类已经有大量研究,但实现一个通用的DR自动分类系统在数据数量与质量、分类方法和系统性能等方面还有一定挑战.

This paper reviews the existing automatic classification methods of diabetic retinopathy(DR). There are two kinds of methods for DR fundus image classification. One is based on local lesions, and the other is based on global image information. The former mainly detects some specific lesions, such as exudation, hemorrhage and microaneurysm, and then performs image classification according to the type, location and number of these lesions. The latter classifies fundus images using global image features. Besides, this paper summarizes commonly used public datasets, advantages and disadvantages of some classification algorithms and their performances. Although many research works have been focused on developing algorithms for automatically classifying DR fundus images, there are still many challenges to develop a universal computer-aided diagnosis system for automatic DR classification. The challenges include acquiring lots of high-quality DR fundus images, designing robust algorithms and improving the total performance of the system.

·