眼底硬性渗出物自动检测系统的研究与实现

发布时间：2018-02-26 00:23

本文关键词： 糖尿病视网膜病变眼底图像硬性渗出物背景估计集成分类深度学习　出处：《哈尔滨工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着人类生产生活方式、饮食习惯的变化以及生活水平的提高,糖尿病已经成为世界范围内广泛影响人们身体健康的慢性疾病。糖尿病患者由于缺乏胰岛素或细胞对胰岛素抵抗作用异常而产生高血糖症状。持续的高血糖打乱了机体正常的代谢活动,造成代谢功能紊乱,从而诱发多种并发症。其中,糖尿病视网膜病变作为糖尿病最严重的并发症之一,已成为导致成年人视力受损甚至失明的主要原因之一。因此,对糖尿病视网膜病变的早期诊断与及时对症治疗对降低患者的失明风险具有重要意义。目前,对糖尿病视网膜病变的筛查主要通过眼科医师人工检查患者的眼底图像来实现,这种检查方式效率较低并且很大程度上依赖于眼科医师的临床经验。因此,研究眼底图像病变自动检测技术,实现对眼底病变客观、高效、准确的检测具有积极意义和实用价值。硬性渗出物的出现是糖尿病视网膜病变的早期症状,所以可以通过对硬性渗出物的检测实现对糖尿病视网膜病变的早期检测与诊断。本文对眼底硬性渗出物自动检测技术的研究主要包括两个方面:基于传统计算机视觉技术的检测方法和基于深度学习的检测方法。其中基于传统计算机视觉技术的检测方法可以划分为硬性渗出物粗分割阶段和硬性渗出物细分类阶段。在粗分割阶段,提出利用形态学方法进行背景估计,并结合形态学重建和阈值化技术初步分割得到含有硬性渗出物的区域。在细分类阶段,引入了区域块特征和区域结构特征等描述子,提取了恰当的区域特征向量,利用集成分类方法构建的装袋决策树对粗分割阶段得到的区域进行进一步分类,得到最终的硬性渗出物检测结果。基于深度学习的方法分别研究了传统的分类网络和生成对抗网络,并对生成对抗网络进行了一定程度的改进,使其更适于进行分类,最终实现了使用生成对抗网络对眼底图像的硬性渗出物进行像素级的检测。本文在公开的眼底图像数据集DIARETDB1上对提出的自动检测算法进行了测试,得到了敏感性98.08%,特异性94.39%,阳性预测值94.07%,准确率95.77%的结果。与其他的相关研究工作进行比较后可以发现本文提出的方法具有一定的优越性和潜在的应用价值。
[Abstract]:With the development of human production and life style, changes in eating habits and the improvement of living standards, diabetes has become a world wide range of chronic diseases affecting people's health. Diabetic patients due to the lack of effect and abnormal symptoms of hyperglycemia on insulin or insulin resistance cells. Sustained high blood glucose disrupted the normal metabolic activities of the human body, causing metabolic function disorder, which cause a variety of complications. The diabetic retinopathy is one of the most serious complications of diabetes, has become one of the major causes of adult visual impairment and even blindness. Therefore, the early diagnosis of diabetic retinopathy and timely symptomatic treatment is of great significance to reduce the patient's risk of blindness. At present, the screening of diabetic retinopathy mainly through the fundus image examination in patients with artificial ophthalmologists to achieve this. Check the clinical experience of low efficiency and rely heavily on the ophthalmologist. Therefore, the automatic detection technology on pathological fundus images, to achieve the objective of retinopathy, high efficiency, has the positive significance and practical value for accurate detection. Hard exudation is the early symptoms of diabetic retinopathy, so by early detection and diagnosis of diabetic retinopathy detection of hard exudates. The fundus hard exudates automatic detection technology research mainly includes two aspects: the traditional detection methods of computer vision technology and detection method based on deep learning. Based on the traditional detection method based on computer vision technology can be divided into hard exudates the coarse segmentation stage and exudate subdivision stage. In the stage of coarse segmentation, using morphological method to estimate the background Meter, and combined with morphological reconstruction and threshold segmentation technology initially contain hard exudates area. In the classification stage, introduced the region characteristics and regional structural characteristics and regional feature vector descriptor extraction appropriate, further classification decision tree is constructed using the bag integrated classification method to get the coarse segmentation stage area, get the final detection results of hard exudates. Method of deep learning classification based on network and traditional network against generation were studied, and the formation of combat network was improved to a certain extent, so it is more suitable for the classification, finally realizes the detection against network generated using hard exudates of fundus image pixel the level of DIARETDB1. The test is made on automatic detection algorithm proposed in fundus images of public data set, the sensitivity was 98.08%, specificity Sex 94.39%, positive predictive value 94.07% and accuracy rate 95.77%. Compared with other related research work, we can find that the method proposed in this paper has certain advantages and potential application value.

【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R587.2;R774.1;TP391.41

【参考文献】