Elsevier

Medical Image Analysis

Volume 13, Issue 6, December 2009, Pages 859-870
Medical Image Analysis

Fast detection of the optic disc and fovea in color fundus photographs

https://doi.org/10.1016/j.media.2009.08.003Get rights and content

Abstract

A fully automated, fast method to detect the fovea and the optic disc in digital color photographs of the retina is presented. The method makes few assumptions about the location of both structures in the image. We define the problem of localizing structures in a retinal image as a regression problem. A kNN regressor is utilized to predict the distance in pixels in the image to the object of interest at any given location in the image based on a set of features measured at that location. The method combines cues measured directly in the image with cues derived from a segmentation of the retinal vasculature. A distance prediction is made for a limited number of image locations and the point with the lowest predicted distance to the optic disc is selected as the optic disc center. Based on this location the search area for the fovea is defined. The location with the lowest predicted distance to the fovea within the foveal search area is selected as the fovea location. The method is trained with 500 images for which the optic disc and fovea locations are known. An extensive evaluation was done on 500 images from a diabetic retinopathy screening program and 100 specially selected images containing gross abnormalities. The method found the optic disc in 99.4% and the fovea in 96.8% of regular screening images and for the images with abnormalities these numbers were 93.0% and 89.0% respectively.

Introduction

Detection of the optic disc and fovea location in retinal images is an important aspect of the automated detection of retinal disease in digital color photographs of the retina. Together with the vasculature, the optic disc and the fovea are the most important anatomical landmarks on the posterior pole of the retina. The optic disc is the area of the retina where the retinal vasculature enters and leaves the eye and it marks the exit point of the optic nerve. The appearance of the optic disc is different from the surrounding retinal tissue (see Fig. 1) and parts of the disc can potentially be classified as abnormal by disease detection algorithms, detecting and masking the disc is the main motivation for this work. Examples of abnormalities that can be confounded with the optic disc are exudates, cottonwool spots and drusen (Niemeijer et al., 2007). Additionally, the optic disc location is useful in the automated tracking of glaucoma (Abràmoff et al., 2007) and the detection of neovascularisations on the optic disc, a rare but serious abnormality. The fovea is responsible for sharp central vision and is located in the center of a darker area (see Fig. 1). Because of its important function in vision, the distance at which lesions are located from the fovea influences their clinical relevance (Treatment Diabetic Retinopathy Study Research Group and photocoagulation for diabetic retinopathy: ETDRS report 9, 1991).

This work is part of a larger project to develop an automated screening system for diabetic retinopathy. Diabetic retinopathy is a common complication of diabetes and the largest cause of blindness and vision loss in the working population of the western world (Klonoff and Schwartz, 2000). We have previously published the results of large scale evaluations of a comprehensive automated screening system on 10,000 exams (40,000 images) (Abràmoff et al., 2008) and 15,000 exams (60,000 images) (Niemeijer et al., 2009). Analysis of the results of these evaluations indicated that the location of the optic disc and especially location of the fovea may potentially be useful in improving the detection of subtle cases of diabetic retinopathy. A number of exams missed by the automated screening system contained only one or two lesions, however, these were located close to the fovea and were thus marked as “suspect” by the screening program ophthalmologists. It is our expectation that information about the location of lesions on the retina may help the system to detect these cases.

A substantial number of publications have dealt with the detection of the location of just the optic disc (see (Abdel-Razik Youssif et al., 2008), for an overview). The state-of-the-art optic disc localization methods use the orientation of the vasculature to detect the position of the optic disc (Abdel-Razik Youssif et al., 2008, Foracchia et al., 2004, Tobin et al., 2007, Niemeijer et al., 2007, Fleming et al., 2007). Detection of the fovea has received less attention, likely due to the fact the fovea is harder to detect and does not feature any high contrast structures. Sinthanayothin et al. (1999) used the increased pigmentation around the fovea to detect its location using a template matching technique and reported a performance of 80.4% on 100 images. Li and Chutatape (2004) used a similar cue to find the fovea and additionally used the location of the vascular arch (see Fig. 1) to constrain the search area. A detection performance of 100% for the fovea is reported on 89 images.

Recently, we have presented an automated method (Niemeijer et al., 2007) to detect the location of the optic disc and fovea. The method used an optimization method to fit a point distribution model (Cootes et al., 1995) to the fundus image. After fitting, the points of the model indicated the location of the normal anatomy. This method was able to find the fovea location in 94.4% and the optic disc in 98.4% of 500 images. On a separate set of 100 heavily pathological images the system showed a performance of 92.0% and 94.0% respectively. The method requires the vascular arch to be at least partially visible but works on images centered on the fovea as well as centered on the optic disc. Tobin et al. (2007) presented an automatic method for detection of the optic disc and the fovea. The method started by locating the optic disc and the vascular arch. Based on these two anatomical landmarks the location of the fovea was inferred. The authors reported 90.4% detection performance for the optic disc and 92.5% localization performance for the fovea in 345 images. The method requires that retinal images are approximately centered on the fovea and that the vascular arch is visible. A similar method for fovea centered images was presented by Fleming et al. (2007): after detection of the vascular arch the optic disc is found using a Hough transform and the fovea is detected by template matching where the template was derived from a set of training images. The optic disc was detected in 98.4% of cases and the fovea in 96.5% of cases in 1056 images.

The main issue with the previous methods that use the vascular arch is that they require it to be visible in the image and most methods are developed for fovea centered images only. The screening data in typical screening programs is acquired at different sites, using different cameras and operators. This leads to a substantial variability in the image quality as well as imperfect adherence to the imaging protocol. The screening protocol used by the particular screening program (Abràmoff and Suttorp-Schulten, 2005) that provides the images used in this work requires the acquisition of two images per eye, one centered on the optic disc and one centered on the fovea. Half of the screening images in the program that supplied the data used in this study are therefore optic disc centered. Methods that require the visibility of the complete vascular arch generally make strong assumptions about the way in which an image is acquired and are less well suited for application on the real world screening data. Screening programs can be structured in different ways. In case of an on-line system where the patient receives almost immediate feedback, rapid processing of the image data is key. In a program where the images are processed off-line, speed may be less of an issue but due to the scale of population screening programs (e.g. hundreds of thousands of patients) an increase in processing speed will have a large impact on the overall efficiency and throughput. Our own previously presented system (Niemeijer et al., 2007) uses a complex optimization procedure and will take approximately 10 min to find the anatomical landmarks. The method by Fleming et al. (2007) takes 2 min to process an image.Tobin et al. (2007) did not report the processing time of their algorithm. As these algorithms were not all benchmarked on the same computer system, the runtime should only be used as an indication and not to directly compare the different methods.

We propose to formulate the problem of finding a certain position in a retinal image as a regression problem. A kNN regressor is trained to estimate the distance to a certain location in a retinal image (i.e. the optic disc and the fovea) given a set of measurements obtained around a circular template placed at a certain location in the image. By finding the position in the image where the estimated distance is smallest, both the optic disc and fovea can be detected.

The main contribution of this work is a fast, robust method for the detection of the location of the optic disc and the fovea. The method makes few assumptions about the way in which the retinal image has been acquired (i.e. optic disc centered or fovea centered) but does require the larger part of the optic disc and the fovea to be in the image in order to successfully detect their location. The method integrates cues from both the local vasculature and the local image intensities. The method is trained using 500 images and is evaluated on a separate, large dataset of 600 images.

This paper is structured as follows. Section 2 describes the data used in this research. The method itself is described in Section 3. The results of the algorithm are shown in Section 4. The paper end with a discussion and conclusion in Section 5.

Section snippets

Materials

In this work 1100 digital color fundus photographs were used. They were sequentially selected from a diabetic retinopathy screening program (Abràmoff and Suttorp-Schulten, 2005). The images represent real world screening data, acquired at twenty different sites using different cameras at different resolutions. In all cases JPEG image compression was applied. Each of the images was acquired according to a screening imaging protocol. From each eye two digital color photographs were obtained, one

Pre-processing

Each of the images is pre-processed before the localization of the optic disc and the fovea commences. First the FOV is detected by calculating the gradient magnitude of the red plane of the image. Several different FOV masks are then fitted to this gradient magnitude image and the best fitting mask is chosen. The template-masks have been previously manually segmented, when a new camera is added to the screening program pool its FOV mask image is segmented and stored. This template based method

Results

The method was applied to the pre-processed green plane image of all 600 test images. To determine our evaluation criteria we have looked at the criteria used in the literature. Counting optic disc detections inside the diameter of the optic disc as a true detection is an accepted evaluation method used by several of the references provided in the introduction (e.g. Tobin et al., 2007, Niemeijer et al., 2007, Fleming et al., 2007). As far as the fovea is concerned, the border of the fovea

Discussion and conclusion

An automatic system for the detection of the location of the optic disc and the fovea has been presented. The system uses a kNN-regressor and a circular template to estimate the distance to a location in the image. It first finds the optic disc and then searches for the fovea based on the optic disc location. The performance of the system is good, in a heterogeneous set of regular screening images it successfully detected the location of the optic disc and the fovea in 99.4% and 93.4% of all

References (25)

  • T. Cootes et al.

    Active shape models – their training and application

    Computer Vision and Image Understanding

    (1995)
  • P. Pudil et al.

    Floating search methods in feature selection

    Pattern Recognition Letters

    (1994)
  • A.A.H. Abdel-Razik Youssif et al.

    Optic disc detection from normalized digital fundus images by means of a vessels’ direction matched filter

    IEEE Transactions on Medical Imaging

    (2008)
  • M.D. Abràmoff et al.

    Web-based screening for diabetic retinopathy in a primary care population: the eyecheck project

    Telemedicine Journal and E Health

    (2005)
  • Abràmoff, M.D., Alward, W.L.M., Greenlee, E.C., Shuba, L., Kim, C.Y., Fingert, J.H., Kwon, Y.H., 2007. Automated...
  • M.D. Abràmoff et al.

    Evaluation of a system for automatic detection of diabetic retinopathy from color fundus photographs in a large population of patients with diabetes

    Diabetes Care

    (2008)
  • M. Ahmed et al.

    A rotation invariant rule-based thinning algorithm for character recognition

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2002)
  • L. Devroye et al.

    A Probabilistic Theory of Pattern Recognition

    (1996)
  • R.O. Duda et al.

    Pattern Classification

    (2001)
  • Early Treatment Diabetic Retinopathy Study Research Group, 1991. Early photocoagulation for diabetic retinopathy: ETDRS...
  • Fleming, A.D., Goatman, K.A., Philip, S., Olson, J.A., Sharp, P.F., 2007. Automatic detection of retinal anatomy to...
  • M. Foracchia et al.

    Detection of optic disk in retinal images by means of a geometrical model of vessel structure

    IEEE Transactions on Medical Imaging

    (2004)
  • Cited by (0)

    1

    Meindert Niemeijer was supported by the Netherlands Organization for Scientific Research (NWO), he now is with the Image Sciences Institute, University Medical Center Utrecht, Utrecht, the Netherlands.

    2

    Michael Abramoff supported by the National Eye Institute (EY017066), the Netherlands Organization for Health Related Research (ZonMW), Research to Prevent Blindness, NY, NY and the Wellmark Foundation.

    View full text