Efficient And Accurate Optic Disk Detection And Segmentation In Digital Fundus Images

Retinal fundus images is widely used for screening and diagnosis of variousfundus diseases for the early prevention and treatment such as diabetic retinopathy,glaucoma, cataract and other blinding eye disease. Therefore early detection andtreatment can effectively inhibit the development of the disease. In order to realize theextensive screening and reduce the professional ophthalmologist’s work burden,automatic disease diagnosis system based on computer vision have widely attractedthe attention of researchers in recent years. The detection and segmentation of themain physiological structure of fundus image play an important supporting role forthe late diagnosis of disease. The optic disc(OD) is the main anatomical structure offundus images, accurate localization and segmentation of OD is helpful for thediagnosis of fundus diseases. For example: many OD segmentation algorithms needan initial seed point; the rough constant distance between the OD and the macula (themost sensitive area of vision),can be used to help estimate the location of the latter;since the OD can be easily confounded with bright lesions(exudate lesions), thedetection of its location is important to remove it from a set of candidate lesions.Due to variable illumination, uneven contrast and lesions interference, accurateOD localization and segmentation is not an easy matter. Especially in diseased images,the OD appearance may be completely destroyed, so the detection is very difficult. Inview of the existing literatures mentioned problems that exist in the OD localizationand segmentation algorithm, this paper is dedicated to improve the robustness and theaccuracy of the algorithm. The main work of this paper is organized as follows.A fast automatic OD localization algorithm is presented. First candidate ODlocations are identified by gradient and intensity information. Then, by observation,the main vasculature are converged in OD region and spread along vertical direction.A template is used to locate the true OD center. The algorithm was tested on fivepublicly-available databases. The OD was successfully located in1514images out of1540images (98.3%). The averaged time for one STARE image is only6.3second.And the proposed method shows good robustness in both normal and diseasedimages.Making use of vessel distribution and directional characteristics of retinalvasculature comprehensively, a novel accurate and fast optic disc (OD) centerdetection method is proposed. A feature combining3vessel distribution characteristics, i.e. local vessel density, compactness and uniformity, is presented tolocate column coordinate of OD center. Then according to the global vessel directioncharacteristic, a General Hough Transformation (GHT) technology is introduced toidentify the row coordinate of OD center. Four public datasets have been used toevaluate the proposed method. The OD was successfully located in339images out of340images (99.7%). And average computation times for STARE images are about2.1-10.3s, which relate to image size. It is better than many previous methods in theaspect of accuracy and efficiency.Reliable and efficient OD segmentation is significant to automatic retinal imagediagnosis. This paper presents a fast automatic OD segmentation algorithm. First ODlocations are identified by projection technology. Then, the main blood vessels aroundthe OD vicinity are removed by Gabor filtering. And the vessel pixels are filled byinterpolation arithmetic with neighborhood information, moreover an edgepreservation filter is adopted to smooth OD region. Finally the OD boundary aredetected by using CV model and level sets algorithm. The algorithm was tested onMESSIDOR databases. The average segmentation accuracies is0.83. The proposedmethod shows a good performance.