Special Issue on CAD/Graphics 2017Classified optic disc localization algorithm based on verification model
Introduction
The optic disc (OD) is one of the major retinal structures, that usually appears as an approximately circular bright object in normal fundus images [1], and it is an important indicator for the detection of fovea and other retinal anatomical structures [2], [3]. The state of the OD in fundus images is a significant evidence for clinical diagnosis of many eye diseases, such as glaucoma, papilledema etc. In addition, the OD localization algorithm is often used before the segmentation to avoid the problem of local extreme value and reduce unnecessary computation [4], [5]. Therefore, the OD localization is the basis and prerequisite for automatic analysis and diagnosis of these diseases [5]. With the development of medical technology, the portable and household medical analysis appliances will be popular in the future. Hence, it puts forward a great demand for the real-time performance of the algorithm.
Many techniques have been proposed to detect the OD from retinal fundus images. In early studies, these techniques detected the OD based on image intensity. Stapor et al. [6] utilized mathematical morphology to detect the OD and its boundary based on geodesic reconstruction by dilation. Chrstek et al. [7] applied an averaging filter to the image, and located the OD roughly at the point of the highest average intensity. Sinthanayothin et al. [8] detected the OD by identifying the area with the highest average variation because of the contrast between blood vessels and background. In [9], candidates of OD center were determined by pyramid decomposition and Hausdorff distance was employed for circular template matching to edge image. These simple and low-complexity methods are very effective for the normal fundus images, but often fail on pathological images, where other regions of fundus may be characterized by round shape or elevated brightness.
Due to the limitations of these image intensity based methods, the vascular geometric structures are utilized to locate the OD center. Foracchia et al. [10] identified the position of the OD using a geometrical model on the vessels structure. In this method, the main vessels originating from the OD are geometrically modelled using two parabolas. It has maximum correlation coefficient when the center of the model coincides with the actual OD center. Youssif et al. [11] detected the OD by matching the directional filtering template and the extracted texture feature of blood vessel. These methods have a relatively high success rate in diseased images, but they are computationally very expensive because they require segmentation of the retinal vessels as an initial step of the localization process. In addition, Mahfouz and Fahmy [12] proposed a fast localization technique based on projections of image features which encode the x and y gradients of the OD. Xiong and Li [13] improved the projections feature localization, increased the candidate point with the edge information, and got a higher accuracy. Zhang and Zhao [14] proposed a fast OD detection method based on vessel distribution and direction characteristics, and they improved the efficiency of OD searching process by simplifying the 2-D search problem to two 1-D search problems. Usman et al. [15] got some candidate OD regions from the grayscale image after preprocessing, and detected OD from these candidate regions according to vessel density. Qureshi et al. [16] also obtained the candidate OD regions by simple thresholding, and evaluated the candidates using multi-attributes of the OD such as average optic disc size and vessels convergence. Wang et al. [17] extracted HOG features of OD area, used SVM to get the OD candidates, and utilized the correlation to determine the final OD location. Brightness characteristic is more effective for locating the OD in normal fundus images. But it can get a higher accuracy to make full use of vascular information. However, the time-consuming of the method based on vascular information cannot satisfy the needs of large-scale screening of retinopathy. How to balance the accuracy and efficiency of the OD localization algorithm is the main problem.
There are three important features of the OD in normal fundus images: (1) The OD is a bright circular object in fundus images. (2) Retinal blood vessels pass through the OD. (3) The OD is the entrance region of retinal blood vessels. The first two features are the local characteristics of the OD, while the last one is the global characteristic. Fig. 1 presents several categories of the fundus images. Fig. 1 (a) is a normal fundus image. In Fig. 1 (b), the OD is not complete. In Fig. 1 (c), the OD is dark and low contrast. In Fig. 1 (d and e), there are large areas of lesions in the image. Fig. 1 (f) is the papilledema, and the OD has no discernible boundary. In this paper, a classified OD localization algorithm is proposed, and the main contributions of this work are summarized as follow: (1) A verification model is presented according to the local features, which is mainly to check whether the region contains the real OD. (2) A practical framework is proposed to integrate the two classes methods based on image intensity and main blood vessels through the verification model, for the purpose that two different methods are used to process normal and abnormal fundus images, respectively.
Section snippets
Method
In this section, the classified OD localization algorithm based on verification model is proposed. We present a verification model to check the candidate region that is obtained by image intensity. If the verification is passed, the corresponding position of the region is determined as the OD center. But if the candidate region does not exist or the verification is failed, the OD is determined by parabolic fitting of the main blood vessels and OD relocation. The flowchart is shown in Fig. 2.
Results
The proposed method is evaluated on four public databases: STARE [1], DRIVE [18], DIARETDB0 [19] and DIARETDB1 [19]. STARE database contains 81 fundus images of size 605 × 700 pixels, DRIVE database consists of 40 fundus images of size 565 × 548 pixels, DIARETDB0 database includes 130 fundus images of size 1000 × 1152 pixels, and DIARETDB1 database includes 89 fundus images of size 1000 × 1152 pixels. These four databases have some abnormal fundus images, and their detail properties are
Discussion
In this paper, because of the low time complexity in the method based on intensity and the high accuracy in the method based on vascular structure, these two methods are integrated through the verification model. In [12], [13], [17], they firstly get a number of OD candidates, and then detect the OD from these candidates. [12] used the image gradient, [13] improved the method in [12] by combining the edge feature, and [17] utilized Multi-features which are combined with global vessel
Conclusion
In this paper, a classified OD localization algorithm based on verification model is proposed. This framework integrates two different localization methods based on image intensity and main blood vessels to deal with various fundus images. The normal and abnormal fundus images are classified by utilizing the verification model. The proposed method is well preformed on four public databases, and experiment results show that our approach can not only guarantee the time efficiency of the normal
Acknowledgments
This work is partly supported by the National Natural Science Foundation of China (Grant Nos. 61573380, 61702559 and 61562029) and the Fundamental Research Funds for the Central Universities of Central South University (Grant No. 2017zzts143).
References (21)
- et al.
Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels
IEEE Trans Med Imaging
(2003) - et al.
A novel method of macula fovea and optic disk automatic detection for retinal images
J Electron Inf Technol
(2014) - et al.
Fast detection of the optic disc and fovea in color fundus photographs
Med Image Anal
(2009) - et al.
Automated segmentation of optic disc and optic cup in fundus images for glaucoma diagnosis
Proceedings of the international symposium on computer-based medical systems
(2012) - et al.
Superpixel classification based optic disc and optic cup segmentation for glaucoma screening
IEEE Trans Med Imaging
(2013) - et al.
Segmentation of fundus eye images using methods of mathematical morphology for glaucoma diagnosis
(2004) - et al.
Optic disc segmentation in retinal images
Inform Aktuell
(2002) - et al.
Automated localisation of the optic disc, fovea, and retinal blood vessels from digital colour fundus images
Br J Ophthalmol
(1999) - et al.
Fast and robust optic disc detection using pyramidal decomposition and hausdorff-based template matching
IEEE Trans Med Imaging
(2001) - et al.
Detection of optic disc in retinal images by means of a geometrical model of vessel structure
IEEE Trans Med Imaging
(2004)
Cited by (27)
Diabetic retinopathy techniques in retinal images: A review
2019, Artificial Intelligence in MedicineCitation Excerpt :Accuracy of the method is mentioned from 92.90% to 100%, for different data sets. Beiji Zou et al. introduced a method for OD localization [107]. The green channel is selected and contrast enhancement is achieved by applying median filter.
Foreword to the Special Issue on CAD/Graphics 2017
2018, Computers and Graphics (Pergamon)Automated detection of optic disc contours in fundus images using decision tree classifier
2020, Biocybernetics and Biomedical EngineeringCitation Excerpt :The pixel with maximum confidence score is considered as OD. Mendonça et al. [22] and Zou et al. [23] localized OD by merging the information obtained from vascular network and intensity values from the red and green channel. The variability and distribution of vessels are obtained using the vascular direction entropy concept.
Robust intensity variation and inverse surface adaptive thresholding techniques for detection of optic disc and exudates in retinal fundus images
2019, Biocybernetics and Biomedical EngineeringCitation Excerpt :Zou et al. (2018) localized OD by integrating intensity information and vascular information. The verfication model is used for classification [17]. Kaur et al. (2018) have developed a diagnostic system for image enhancement, anatomical structure elimination, feature set extraction and grading the different levels of DR [18].