Документ взят из кэша поисковой машины. Адрес оригинального документа : http://imaging.cs.msu.su/pub/2011.Graphicon.Chernomorets_Krylov_Nasonov_Semashko.Retina.en.pdf Дата изменения: Mon Aug 22 14:18:34 2011 Дата индексирования: Sat Apr 9 23:46:58 2016 Кодировка:

Automated processing of retinal images
A.A. Chernomorets1, A.S. Krylov1, A.V. Nasonov1, A.S. Semashko1, V.V. Sergeev1, V.S. Akopyan2, A.S. Rodin2, N.S. Semenova2 2 Faculty of Computational Mathematics and Cybernetics Faculty of Fundamental Medicine Lomonosov Moscow State University, Russia {chernomorets, kryl}@cs.msu.ru
where I ( x, y ) -- value of the pixel with coordinates ( x, y) ,

1

Abstract
The problem of automated processing of retinal images for semi automatic diagnosis of retinal diseases is considered in this paper. The main aspects of retinal image processing are discussed. Methods for non-uniform illumination correction, blood vessel detection and macula and optic disc segmentation, finding dark and light spots in the macula area are suggested. Keywords: retinal image processing, eye fundus, telemedicine.

0 -- small coefficient.
The pixel is classified as the background pixel if D( x, y) DT . We use DT 0, 7 . 2. The second step consists of luminosity correction as

I C ( x, y ) 128

I ( x, y) , C ( x, y)

1. INTRODUCTION
Automated analysis of retinal images using software programs makes it possible to perform mass diagnosis (screening) without involving a medical officer. This is very important in terms of time saving and health budget economy and gives a possibility to make the criteria of detection of the most dangerous retinal diseases more objective and to detect retinal diseases at early stages. The following major task groups can be marked out in the pro blem of automated processing of retinal images: 1. Preprocessing of retinal images. 2. Segmentation of retinal objects which include blood vessels, macula -- the central part of the retina and the optic disc (optic nerve head). 3. Detection and assessment of pathologies.

where C ( x, y) is the mean intensity value of background pixels in the neighborhood of the pixel at ( x, y) , 0 . Fig. 1 illustrates the proposed method of luminosity correction.

) original image b) result Fig 1: Example of correction of the non-uniform luminosity

2. PREPROCESSING OF RETINAL IMAGES
Image preprocessing consists of correction of non-uniform luminosity, color normalization and contrast enhancement. Methods of general image enhancement (sharpening of blurred images, noise suppression, etc.) can be added to this list. In this work we use a method of luminosity correction that is based on segmentation of background pixels and subsequent computation of luminosity function based only on the background image [1]. The advantage of this approach is that it does not produce ringing effect. The algorithm consists of two steps: 1. First we segment background pixels based on analysis of statistical values. For each pixel of the image we compute mean value ( x, y) and standard deviation ( x, y) within a window of a fixed size, then we compute Mahalanobis distance as follows:

3. OBJECT SEGMENTATION IN RETINAL IMAGES 3.1 Vasculature segmentation
There are several general approaches to automated vessel segmentation [2]: matched filtering; ridge detection; vessel tracking that allows connecting disjoint parts of the vessel; pixel-based classification; mathematical morphology. In the proposed method we apply morphological amoebas to the result of Gabor filtering of the image.

3.1.1 Gabor filtering
2-D Gabor filtering is a convolution with a kernel that can be formulated as

D ( x, y )

I ( x , y ) ( x, y ) , ( x, y )


2 x 2 2 y exp i 2x g x, y; , , , , exp 2 2 x x cos y sin where is the rotation operation. y x sin y cos



,

The structuring element for the pixel ( x0 , y0 ) consists of all pixels with the shortest path to the pixel ( x0 , y0 ) not exceeding the given threshold. The examples of amoebas' shapes are given in Fig. 3.

In

these notations, the parameter defines the wavelength of the sinusoidal factor, sets the orientation of the filter, is the phase offset, is the scale of the filter and specifies the ellipticity. Values of the parameters , and are chosen in order to detect parts of vessels of different directions and widths. We use 6 directions from 0° to 150° with a step of 30° and 4 values for parameter :
0 is the maximum 4 2 4 vessel width. Other parameters are set to: = 3, = , = 1.

1

0 , 0 , 0 an d 0 , where

1

3

The result is the maximum modulus of the filtering over all angles for multiple scales. Fig. 2 illustrates vessel enhancement by Gabor filtering.

Fig 3: Examples of shapes of morphological amoebas. Pixels that form the structuring element are marked in blue. The algorithm of segmentation using morphological amoebas consists of the following steps: 1. Via thresholding of the result of Gabor filtering the set M that consists only of vessel pixels is produced.
0

2. The set of vessel pixels is morphologically dilated using morphological amoebas. The set is then eroded using circular structuring element and is joined to the set from the previous step:

M

n 1

M n Erosion( AmoebasDilation(M n ))
n 1

3. The process terminates when M

Mn .

The result of the segmentation using morphological amoebas is shown in Fig. 4. Fig 2: Result of vessel enhancement using Gabor filtering.

3.1.2 Morphological amoebas
The segmentation method is applied to the result of the Gabor filtering. Thresholding does not produce satisfying results, yielding to either a lot of false positive responses or loss of significant amount of smaller vessels. Mathematical morphology allows to enhance thresholding methods. For segmentation of retinal vasculature we use morphological amoebas, described in [3,4], with modified amoeba distance fun ction. The general idea of morphological amoebas is that the shape of the structuring element is adapted for each pixel of the image. For each pair of adjacent pixels the distance based on pixels' intensities is computed. We use the following distance function: Result of Gabor filtering Result of segmentation

4: The result of the vessel segmentation using morphological amoebas.

d (( x0 , y0 ), ( x1 , y1 )) ( x1 x0 ) 2 ( y1 y0 ) 2 ( I ( x0 , y0 )) 2 ( I ( x1 , y1 )) 2 , 0.
The summand ( I ( x 0 , y 0 ) 2 I ( x1 , y1 )
2

3.2 Macula and optic disc segmentation
The localization of macula and optic disc can be performed via analysis of the vasculature. We find these regions in assumption that form of the major vessel arcades is close to circle, the center of this circle is the center of the macula and the circle crosses the optic disc. Let {P ( xk , yk )}, k 1, ..., K be a set of coordinates of major k vessel pixels. The resulting circle is the one that minimizes the dispersion of distances from pixels Pk to the circle boundary.

makes the distance

shorter in dark areas and longer in light areas. It restricts the amoeba expansion only inside the vessel. Parameter defines the cost of expansion. The summand
(x x )2 ( y y )2 1 0 1 0

does not allow the amoeba to grow indefinitely in case of a co nstant image.


The macula region is taken as a circle with 1/3 of the radius of obtained circle and the same center. A bright region on the circle boundary is taken as the initial approximation of the center of optic disc that is further used for detection of its boundary. Example of macula and optic disc localization is shown in Fig. 5.

ties of the candidate areas which include compactness, ellipticity, dispersion, average intensity in red and green channels. The first property is used to exclude large blood vessels. The remaining small vessels are segmented from dark spots using the SVM (Support Vector Machines) method. An example of dark spot detection is shown in Fig. 7.

Fig 5: Result of macula and optic disc localization. The macula region is highlighted in the center of the image.

Fig 7: An example of dark spot detection.

4.2 Finding optic disc boundary
The analysis of optic disc boundary is important for glaucoma diagnosis. To find its boundary the method based on active contours was developed [5]. At the preprocessing step the blood vessels are removed from the image using morphological closing and opening in CIE Lab color space. [6]. Next the blood vessel map is constructed from the difference of the obtained image and the source image. Then the edge map is calculated as the modulus of the intensity gradient. Using the Hough transform, the initial contour best approximating the optic disc boundary is taken. The external force field Gradient Vector Flow (GVF) [7] which is used in active contours method is constructed from the boundary map.

4. DETECTION AND ASSESSMENT OF PATHOLOGIES 4.1 Finding spots in the macula area
Appearing of dark and light spots in the retina can be an indication of the retinopathy.

4.1.1 Light spot detection
We perform the detection of lesions in the macular area. The green channel of the preprocessed image is used for the light spot analysis. Morphological closing and opening is performed to exclude blood vessels from the image. The size of the structuring element is chosen in accordance to image size. Then we calculate dispersion values in every pixel of the image, apply thresholding and find coordinates of light spot candidates. Next to distinguish between light spots and outliers we fill every light spot candidate area with the value taken from surrounding background pixels, then subtract this image from the original one and apply the threshold to find light spots. An example of light spot detection is shown in Fig. 6.

Fig 8: An example of morphological preprocessing. Left: the source image; middle: the result of blood vessel removal; right: the blood vessel map. At the next step, the preliminary result is calculated. This result is usually inaccurate and the deviation of the obtained boundary from the true optic disc boundary is usually observed at intersections of the optic disc boundary with blood vessels. At the last step, the correction of the obtained boundary is performed. We construct the ellipsis approximating the obtained boundary. Every point of this ellipsis is assigned a weight depen ding on the edge map and the blood vessel map. The external force attracting the contour to this ellipsis is added. Another force (the pressure force) directed along the external contour normal is ad ded to the points of intersection of the contour with blood vessels.

Fid 6: An example of light spot detection in retinal images.

4.1.2 Dark spot detection
Like for the light spot detection, the green channel of the cessed image is used for dark spot detection. We perform filtering, then apply thresholding to find dark spot can Next dark spots are separated from outliers. We calculate prepromedian didates. proper-


About authors
A.A. Chernomorets ­ PhD student of Faculty of Computational Mathematics and Cybernetics (CMC), Lomonosov Moscow State University (MSU). A.S. Krylov ­ professor, head of Laboratory of Mathematical Methods of Image Processing, CMC MSU. Fig 9: shown shown results Examples of optic disc detection. True contours are by white color. The results of the proposed method are by green color. Black dashed lines are the preliminary (without applying active contours). A.V. Nasonov ­ member of scientific staff of CMC MSU A.S. Semashko ­ PhD student of CMC MSU. V.V. Sergeev ­ student of CMC MSU. V.S. Akopyan ­ professor, head of Ophthalmology department, Faculty of Fundamental Medicine (FFM), MSU. A.S. Rodin ­ assistant of Ophthalmology department, FFM MSU. N.S. Semenova ­ assistant of Ophthalmology department, FFM MSU.

The following performance values were obtained on DRIVE and MESSIDOR bases: average overlap ­ 0.89, sensitivity ­ 94.1%, specificity ­ 98.4%. This confirms the effectiveness and the reliability of the proposed method.

5. CONCLUSION
The proposed in the paper algorithms for non-uniform illumination correction, blood vessel detection, macula and optic disc segmentation in retinal images, finding light and dark spots in the macula area compose the basis of the telemedicine system of retinal image analysis developed by the authors group. The target of the future work is creating the integrated software system based on the optimization and coordination of the work of the proposed algorithms. The work was supported by the state contract 8/3-617-10 with the Moscow government.

6. REFERENCES
[1] G.D.Joshi, J.Sivaswamy. Colour Retinal Image Enhancement based on Domain Knowledge // 6th Indian Conf. on Computer Vision, Graphics and Image Processing, 2008, pp. 591­598. [2] Winder R.J., Morrow P.J., McRitchie I.N., Bailie J.R., Hart P.M.. Algorithms for digital image processing in diabetic retin opathy. Computerized Medical Imaging and Graphics, vol. 33, pp. 608­622, 2009. [3] M.Welk, M.Breub, O.Vogel. Differential Equations for Mo rphological Amoebas // Lecture Notes in Computer Science, Vol. 5720/2009, 2009, pp. 104­114. [4] Romain Lerallut, Etienne DecenciХre, Fernand Meyer. Image filtering using morphological amoebas// Image and Vision Computing, Vol. 25, 2007,pp. 395­404. [5] A. S. Semashko, A. S. Krylov, A. S. Rodin. Using Blood Vessels Location Information in Optic Disk Segmentation // 16th International Conference on Image Analysis and Processing (ICIAP'2011), accepted for publication. [6] A. Osareh, M. Mirmehdi, B. Thomas, R. Markham. Colour morphology and snakes for optic disc localisation. // 6th MIUA Conference, 2002, pp. 21--24. [7] C. Xu, J. Prince. Snakes, shapes, and gradient vector flow. // IEEE Trans. Image Processing 7, 1998, pp. 359 --369.