A method for the detection and classification of diabetic retinopathy using structural predictors of bright lesions
A method for the detection and classification of diabetic retinopathy using structural predictors of bright lesions
复制标题
DOI：
10.1016/j.jocs.2017.01.002
复制
发表时间：
2017-03-01
影响因子：
3.3
通讯作者：
Fernandes, Steven Lawrence
中科院分区：
文献类型：
--
作者：
Amin, Javeria;Sharif, Muhammad;Fernandes, Steven Lawrence
世界各地的糖尿病负担以及糖尿病视网膜病变的后果可导致患者永久失明。通过自动方法检测眼底图像中的渗出物是在糖尿病视网膜病变筛查中具有许多应用的重要任务。意识到这一点的重要性，本文提出了一个系统自动分类渗出液和非渗出区域的视网膜图像。提出的技术是基于预处理的候选病灶提取，特征提取和分类。在图像预处理中，对灰度图像进行了Gabor滤波，使其更有利于病灶的增强。候选病灶的分割是基于数学形态学。使用统计和几何特征的组合为每个候选病变选择特征集。所提出的方法是通过公开访问的数据集的帮助下，如真阳性，假阳性和曲线下面积进行统计分析的性能参数进行评估。公开可用的数据集，如e-ophtha，HRIS，MESSIDOR，DIARETDBI，VDIS，DRIVE，HRF和一个本地数据集被用来测试建议的系统。结果表明，平均AUC为0.98，准确度高达98.58%，明显高于现有方法。(C)2017 Elsevier B. V.版权所有。
Diabetic burden around the world with a consequence of diabetic retinopathy can lead to permanent blindness in patients. Exudates detection in fundus images through an automated method is a vital task that has many applications in diabetic retinopathy screening. Realizing it important, a system being proposed in this paper automatically classifies exudates and non-exudates regions in retinal images. Presented technique is based on pre-processing for candidate lesion extraction, features extraction and classification. In pre-processing, Gabor filter is applied to the gray scale image which makes it useful for lesion enhancement. Segmentation of candidate lesion is based on mathematical morphology. A features set is selected for each candidate lesion using a combination of statistical and geometric features. Presented method is evaluated via publicly accessible datasets with the help of performance parameters such as true positive, false positive and area under curve for statistical analysis. Publicly available datasets such as e-ophtha, HRIS, MESSIDOR, DIARETDBI, VDIS, DRIVE, HRF and one local dataset are used to test the suggested system. The achieved results show an average AUC of 0.98 and accuracy as high as 98.58% which are substantially higher than the existing methods. (C) 2017 Elsevier B.V. All rights reserved.