Ganesan Karthikeyan, Acharya Rajendra U, Chua Chua Kuang, Laude Augustinus
Department of Electronic and Computer Engineering (ECE), Ngee Ann Polytechnic, Singapore
Department of Electronic and Computer Engineering (ECE), Ngee Ann Polytechnic, Singapore Department of Biomedical Engineering, Faculty of Engineering, University of Malaya, Kuala Lumpur, Malaysia.
Proc Inst Mech Eng H. 2014 Sep;228(9):962-70. doi: 10.1177/0954411914550847. Epub 2014 Sep 17.
Identification of retinal landmarks is an important step in the extraction of anomalies in retinal fundus images. In the current study, we propose a technique to identify and localize the position of macula and hence the fovea avascular zone, in colour fundus images. The proposed method, based on varying blur scales in images, is independent of the location of other anatomical landmarks present in the fundus images. Experimental results have been provided using the open database MESSIDOR by validating our segmented regions using the dice coefficient, with ground truth segmentation provided by a human expert. Apart from testing the images on the entire MESSIDOR database, the proposed technique was also validated using 50 normal and 50 diabetic retinopathy chosen digital fundus images from the same database. A maximum overlap accuracy of 89.6%-93.8% and locational accuracy of 94.7%-98.9% was obtained for identification and localization of the fovea.
识别视网膜地标是提取眼底图像异常的重要步骤。在当前研究中，我们提出了一种技术，用于在彩色眼底图像中识别并定位黄斑的位置，进而定位无血管区。所提出的方法基于图像中不同的模糊尺度，与眼底图像中存在的其他解剖地标位置无关。通过使用开放数据库MESSIDOR提供了实验结果，使用骰子系数验证我们分割的区域，并由人类专家提供地面真值分割。除了在整个MESSIDOR数据库上测试图像外，还使用从同一数据库中选择的50张正常和50张糖尿病视网膜病变数字眼底图像对所提出的技术进行了验证。对于中央凹的识别和定位，获得了89.6%-93.8%的最大重叠准确率和94.7%-98.9%的定位准确率。
