New Optic Disc Localization Method for Retinal Images

Florin Rotaru, Silviu Ioan Bejinariu, Cristina Diana Niţă, Ramona Luca Institute of Computer Science, Romanian Academy, Iasi Branch, Romania

Abstract - The paper proposes an improved optic disc localisation method in color retinal images. First, the optic disc area in retinal images of any dimensions is identified. Then the method iteratively extracts the optic disc edges and obtains a circular optic disc boundary approximation by a Hough transform. However this is a first step of a retinal image analysis project which will be completed later with other tasks. The final goal is to detect in early stages signs of ophthalmic pathologies by successive analysis of ophthalmoscopy images.

I. INTRODUCTION Since the early detection of ophthalmic pathologies is crucial

for a reasonable treatment and the traditional ways do not cope with this complex task, automatic analysis is required to assist the specialists. During the last decade many automatic methods were proposed. Due to the large variations between individuals and uneven quality and diversity of the acquired retinal images, the problem is not yet fully solved.

However, some important results were achieved. Some of them are focused only on optic disc detection. The optic disc area measurement is essential for diagnosis of different aspects caused by diabetic retinopathy. Also, optic disc segmentation can be useful for glaucoma detection. This eye condition can be diagnosed by identifying in time the changes in the optic disc area. Another important argument for optic disc recognition is that once the disc is located, the fovea localization, otherwise difficult to compute, becomes an easier task due the relatively constant distance between the fovea and optic disc.

One of the first approaches to locate optic disc area was proposed in [8]. The method combines two algorithms: a pyramidal decomposition using Haar wavelet transform and an optic disc contour detection based on Haussdorf distance. White areas as exudates which might disturb the right optic disc area detection are eliminated during pyramid synthesis. Finally, the low resolution level contains only the useful information. On the lowest resolution image ten optic disc candidates are retained. First the highest intensity values compared to the mean intensity over a search region are considered. Then the selected regions are smoothed and the brightest pixel is retained as possible optic disc centre candidate. In the second step contiguous regions are aggregated into a single zone. Applying several times a Canny edge detection a binary edge map and a noisy edge map are obtained. From the second one a threshold value is computed and subsequently used to binarize the aggregated regions. On two level regions, optic disc contour detection was performed using Hausdorf distance.

An automatic optic disc detection based on majority voting for a set of optic disc detectors, is proposed in [3]. There were implemented five previously proposed methods to detect optic disc centre: pyramidal decomposition [8], edge detection [8], entropy filter [12], fuzzy model [5] and Hough transform [10]. A circular template of radius R = 45, specific for the size of processed images, is fit on each pixel in the image to count the outputs of these algorithms that fall within the radius. The circle with the maximum number of optic disc detector outputs in its radius is the chosen area to refine the optic disc detection. An improved version of the voting method was proposed in [2].

Considering more complex analysis of retinal images, there are two distinctive approaches: bottom-up and bottom-down.

First one locates the optic disc and then starting from that area track the retinal vessels and does the required measurements. The top-down approaches track the retinal vessels and get the optic disc as the root of the vessels tree. There are arguments pro and contra for one or another approach. The optic disc direct segmentation is quite difficult when there are exudates in the disc neighborhood. Also, bright areas in the very vicinity of the optic disc might distort its round shape. On the other hand vessels tracking might be a very challenging task if there are discontinuities in the vessel structure. We mention two top-down approaches, [5] and [7]. While in [5] the retinal vessels convergence point is detected by employing a voting-type algorithm named fuzzy convergence, in [7] first there are identified the four main vessels in the image. Then the four branches are modeled by two parabolas whose common vertex is identified as optic disc centre.

A very known now bottom-up technique was proposed in [4]. The optic disc area localization is performed by employing a principal component analysis method. This implies a previous training step. Then a modified active shape model is proposed to identify the disc boundary.

Another important work is the one presented in [1]. Part of this methodology was also implemented in our system to process retinal images. There were implemented three methods for a rough identification of optic disc area: the maximum difference between the maximum and minimum grey levels in working window, the maximum variance method and frequency low pass filter method. The optic disc area is located using a voting procedure. The green channel of the RGB input image is used. The voting procedure establishes the estimated disc centre in the following way: 1) if all three candidates are close to their centre this one is proposed as an approximate disc centre; 2) if only two from three candidates are close to the centre the

978-1-4673-6143-9/13/$31.00 ©2013 IEEE

average point of these two is chosen; 3) if all candidates are far apart from their centre the candidate proposed by the second method, the most reliable considered by the authors, is chosen.

Then a 400x400 window is centered on the estimated disc centre, and extracted from original green and red channels. A morphological filter is employed from [6] to erase the vessels in the new window. A Prewitt edge detector is then applied and by the same Otsu technique the image is binarised. The result is cleaned by morphological erosion and finally a Hough transform is applied to get the final optic disc boundary. The boundary with the best fitting from the two channels is chosen. The authors report for 1200 retinal images a score of 100% for approximated localization and a score of 86% for final optic disc localization.

II. NEW OPTIC DISC RECOGNITION METHOD In order to exactly locate the optic disc first we started

following a similar methodology as the one proposed in [1]. Tests were done on RGB retinal images of 720x576 and 2592x1728 resolutions, some of them of patients strongly affected by eye disorder. From the three methods of the voting procedure presented in [1] good optic disc area localization we obtained only with the Low-Pass Filter Method, the third method of the voting procedure in [1]. Our implementation it is a common one: to smooth out the little white patches which can perturb the right disc localization the green channel of the input image is transformed in frequency domain. As in [1] on the image of the magnitude of the FFT transform a Gaussian low-pass filter was applied:

( )⎟⎟⎠

⎞⎜⎜⎝

⎛−= 2

0

2

2,exp),(

DvuDvuH (1)

where ( )vuD , is the Euclidean distance from point ( )vu, to

the origin of frequency domain and 0D is the cutoff frequency, of 25 Hz. The filtered result is transformed back to the spatial domain and the brightest pixel of the result image is chosen as an optic disc area centre candidate.

Good results were obtained also with another approach derived from the Maximum Difference Method, proposed in [1] and developed and used in [11]. For 720x576 images we kept the same voting procedure as in [11] using the Low-Pass Filter Method and the modified Maximum Difference Method. For high resolution images we relied only on Low-Pass Filter Method to approximately locate the optic disc area.

The results obtained with these procedures, for both resolutions, are illustrated by figure 1.

Once the preliminary optic disc centre was established, as in [1] and [11], the further work was done on a square window centered on the optic centre. However this time the searching window side is a fraction of image height. The tests were done on green channel of 86 retinal images of 720x576 size and 40 images of 2592x1728 resolutions. Following the same technique employed from [6] in the established window the blood vessels were eliminated. Next we shall describe shortly this method.

a)

b)

c)

d)

Figure 1. Results of detecting approximate optic centre position by two voting procedures for low resolution image, figures a) and b). Point marked

with little cross is provided by the first method and the one indicated by large cross is computed by second voting algorithm. When the two points are far

apart, as in the b) image, the centre computed by the first method is chosen. A result using Low-Pass Filter Method for high resolution image is depicted in

figure d). Part of original high resolution image is illustrated in figure c).

The opening ( )IBγ of an image I by the structuring element B is:

( ) ( )( )II BBB εδγ = (2)

where the erosion ( )IεB is defined as:

( )[ ]( ) ( )by,axIminy,xIB)b,a(

B ++=∈

ε (3)

and the dilation ( )IBδ is:

( )[ ]( ) ( )by,axIminy,xIB)b,a(

B ++=∈

δ (4)

As in [1], a line was considered the structuring element, but in our case with 1 pixel width and the length being a fraction of image height. For each of 12 different orientation of the line an opening of the original selected window was performed. The clean image, without vessels, is:

( )( )IIiBiB γ

12,...1min=

= (5)

This means that each pixel of the result image has the minimum value from the set of 12 values of the same coordinate pixel in each of the 12 openings of I .

Results of the vessels erasing operation are illustrated by figure 2, for high resolution images.

Figure 2. The result of vessels erasing for image 1.d.

The next operation on image BI is a Canny filtering followed by binarization to obtain the disc edges, in order to perform a final circle fitting. Let us note with CannyI the binary image containing the Canny edges.

Due the great variability of pathology and image primary sources a fixed threshold for Canny filter is not desirable. We propose a new iterative approach, different from the one implemented in [11]:

1. Compute a binarization threshold using Otsu method, [9],

on image BI , without performing the binarization.

2. Choose a value close to Otsu threshold as a primary threshold for Canny filtering.

3. Perform Canny filtering (pixels representing edges are “white”, the others are “black”).

4. Process CannyI : Erase the white pixels having more than

three black neighbors in an 11x11 (high resolution) or 6x6 (low resolution) neighborhood of current pixel but in the original image (input for Canny filtering) – these are supposed to be border pixels of the big retina circle (see figure 3).

5. If there are not enough white pixels (less than a predefined threshold) adapt the threshold for Canny filtering and resume process from step 3.

6. Compute minr and maxr , the minimum and maximum values of

circles radius, as fractions of the original image width.

7. For an interval [ ]maxmin r,r of circle radius compute a circle

fitting by Hough transform applied on the whole window. 8. Choose the centre radius with the best fitting score and best

distribution of fitting points. 9. If the fitting score is not desirable or there are few points to

perform the fitting, decrease the Canny threshold by a certain amount (constant in our implementation) and perform Canny filtering and process the binary image result as in step 4. Not more than a predefined number of iterations resume the process from step 7.

10.Even the fitting score and the distribution fitting points are acceptable run at least one more time all the steps 7-9 with a new Canny decreased threshold.

11.If the detected circles have comparable fitting scores and fitting point distributions, choose the circle with the longest radius.

A result of image CannyI processing is illustrated by figure 3, where the image 3.a is the output of Canny filter and the image 3.b is the output of cleaning procedure described in step 3 of the global algorithm.

a)

b)

Figure 3. Result of processing the binary image output of Canny filter – step 4. a) primary Canny output; b)final result.

Results of the iterative approach, steps 7-9, only for high resolution images, are illustrated by figure 4, where there are depicted: a) the original image from figure 1c; b) the Canny edge extraction results for the starting threshold; c) Canny edges for adapted threshold; d) final circle.

a)

b)

c)

d)

Figure 4. a) Part of the input RGB high resolution image, the same as in figure 1c; b) the Canny edge extraction results for the starting threshold; c)

Canny edges for adapted threshold; d) final circle.

While Canny filtering was performed using the OpenCV function for Hough transform, we implemented our own method in order to get more control on the distribution of the fitting points. This way some configurations can be rejected even there are generated by an acceptable number of fitting points if the points are not equally distributed around circle centre.

a)

b)

c)

d)

e)

f)

g)

h)

Figure 5. On the left column (a, c, e, g): original high resolution retinal images. On the right (b, d, f, h): the final optic disc localization results

illustrated

III. RESULTS AND CONCLUSIONS Tests have been done on 86 RGB retinal images of 720x576

resolution and 40 images of 2592x1728 resolution. The rough optic disc localization has been successful on whole the image set. The final circle fitting failed on two low resolution images strongly affected. Due the iterative way to establish the Canny threshold and circle radium the proposed method is suitable to process retinal images of any resolutions, different accuracies and a great variability of pathology. The papers mentioned in section I mainly mention a specific resolution image.

Figure 5 illustrates some final circle localization results for high resolution images.

We shall go on testing the method using other retinal image databases, of various quality and resolution, provided by our collaborators from University of Medicine and Pharmacy „Gr. Popa” Iasi.

The reported work is a first stage of a larger project that will be completed later on with other tasks as fovea detection and measurement of retinal vessels. The final goal is to detect in

early stages signs of ophthalmic pathologies as diabetic retinopathy or glaucoma.

The optic disk localization procedure was implemented and tested in an image processing framework developed by authors. It is implemented as a Windows application, in C++ using Microsoft Visual Studio. For image manipulation and some processing functions, the OpenCV library is used.

ACKNOWLEDGMENTS The work was done as part of research cooperation with

University of Medicine and Pharmacy „Gr.Popa” Iasi to analyze retinal images for early prevention of ophthalmic diseases.

REFERENCES [1] A. Aquino, M.E. Gegundez-Arias, and D. Marin, “Detecting the optic

disc boundary in digital fundus images using morphological, edge detection, and feature extraction techniques”, IEEE Transactions on Medical Imaging, 2010.

[2] B. Harangi and A. Hajdu, “Improving the accuracy of optic disc detection by finding maximal weighted clique of multiple candidates of individual detectors”, IEEE 9th International Symposium on Biomedical Imaging (ISBI2012), Barcelona, Spain, 2012, 602-605.

[3] B. Harangi, R.J. Qureshi, A. Csutak, T. Peto, and A. Hajdu: “Automatic detection of the optic disc using majority voting in a collection of optic disc detectors”, IEEE 7th International Symposium on Biomedical Imaging (ISBI 2010), Rotterdam, The Netherlands, 2010, 1329-1332.

[4] H.Li and O.Chutatape, “Automated feature extraction in color retinal images by a model based approach”, IEEE Transactions on Biomedical Engineering, Vol.51, No.2, February 2004.

[5] A. Hoover and M. Goldbaum, “Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels,” IEEE Trans. Med. Imag., vol. 22, no. 8, pp. 951–958, Aug. 2003.

[6] C. Heneghan, J. Flynn, M. O’Keefe, and M. Cahill, “Characterization of changes in blood vessel width and tortuosity in retinopathy of prematurity using image analysis,” Med. Image Anal., vol. 6, pp. 407–429, 2002.

[7] M. Foracchia, E. Grisan, and A. Ruggeri, “Detection of optic disc in retinal images by means of a geometrical model of vessel structure,” IEEE Trans. Med. Imag., vol. 23, no. 10, pp. 1189–1195, Oct. 2004.

[8] M. Lalonde, M. Beaulieu, and L. Gagnon, “Fast and robust optic disk detection using pyramidal decomposition and Hausdorff-based template matching”, IEEE Trans. Medical Imaging, Vol. 20, pp. 1193-1200, Nov. 2001.

[9] N. Otsu, "A threshold selection method from gray-level histograms," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 9, No. 1, 1979, pp. 62-66.

[10] S. Ravishankar, A. Jain, and A. Mittal, “Automated feature extraction for early detection of diabetic retinopathy in fundus images”, CVPR - IEEE Conference on Computer Vision and Pattern Recognition, pp. 210-217, 2009.

[11] F.Rotaru, S.Bejinariu, C.D.Niţă, and M.Costin, "Optic disc localization in retinal images", 5th IEEE International Workshop on Soft Computing Applications, IEEE-Sofa, 23-25 August, 2012, Szeged, Hungary, Soft Computing Applications - Advances in Intelligent Systems and Computing Volume 195, 2013, Springer Verlag.

[12] A. Sopharak, K. Thet Nwe, Y. Aye Moe, M. N. Dailey, and B. Uyyanonvara, “Automatic exudate detection with a naive Bayes classifier”, International Conference on Embedded Systems and Intelligent Technology, Grand Mercure Fortune Hotel, Bangkok, Thailand, pp. 139–142, 2008.