Efficient Robust Optic Disc Detection In Retinal Fundus Image

In recent years, retinal fundus images are widely used in diagnosis of glaucoma, cataracts and other retinal diseases, as well as diabetes, hypertension, coronary heart disease and other diseases which have serious impacts in visual. Because of the large number of the patient and the scarcity of ophthalmologist, automatic disease diagnosis system for retinal fundus based on computer vision have widely attracted the attention of researchers. The optic disc(OD) is one of the main physiological structure of the retinal fundus images, and accurate positioning of OD in fundus image has the following important role for visual inspection system:(1) Since the distance between the OD and the macula can be calculated, the OD positioning can help to locate the macula;(2) OD is the convergence zone of retinal vascular, and the main vessel from the disc extends to the other regions of the entire retina fundus image, so the OD positioning help vessels segmentation;(3) Accurate positioning of the OD can distinguish itself from yellow lesions(exudation lesions) that maybe easily confused with OD, which helps to improve the accuracy and robustness of retinal lesion detection.The changes caused by the imaging conditions and image quality problems, and a variety of diseases and other factors that interfere or damage the appearance of the OD, making the OD positioning not an easy problem to solve. There doesn't exist an algorithm that be generally applicable to large-scale test. Existing algorithms still have some deficiencies in terms of robustness or efficiency. For the above problems, this paper commits to study new algorithm to improve the robustness and efficiency of the OD positioning algorithm. The main work of this paper is organized as follows.In order to test the error circumstances of algorithms that use the blood vessel characteristics and visual characteristics when images present different image quality and lesion type, we select three typical methods to do experimental analysis: a OD localization method based on a linear operator, a method based on vessel distribution and directional characteristics and a method based on projection. These three methods are tested and analysis the algorithm robustness in a large-scale of lesions image sets that we collected. Experimental results show that the method that use the appearance characteristics of the OD or only use vessel characteristics are more sensitive to the appearance quality of OD and the vessel quality. When the quality deteriorates, the accuracy will decrease in a greater degree. In addition, as the number of lesions types increased, the localization accuracy of these three methods will decline.A fast OD localization algorithm is presented. Firstly candidate OD regions are identified by Lo G blob detection. Then we multipurpose use the appearance characteristics and vessel characteristics of candidate region to identify the true OD position. Since this method can be performed on the fundus image after size reduction, it is possible to obtain real-time detection results. And because this method has a detecting capability of multi-scale, it can be adapted to different sizes of image data sets. The method is tested on 1540 images of five public data sets, and the positioning accuracy reached 97.9%. The average detection time of the five data sets is about 0.2s, which is far faster than most of the current detection algorithm.A robust OD detection based on multi-features and two-stage decision is presented. First, the method uses the global vascular distribution and directional characteristics and the local appearance properties to find several candidate OD points. Then we introduce the local feature descriptor HOG features to describe the details of the candidate OD points, and through the SVM model to distinguish optic disc and non-optic disc area. Finally non-maximum suppression rule based on correlation measure is used to determine the final position of OD. Experimental results show that this algorithm has good robustness, and it shows good detection accuracy in both normal and diseased fundus image. The detection accuracy reached 97.9% in four public image set.