Academic Open Internet Journal

ISSN 1311-4360

www.acadjournal.com

Volume 17, 2006

 

 

 

Genetic Algorithm Based Location of Optic Disc in Retinal Images

 

D.Abraham Chandy 1, V.Vijaya Kumari 2

1 Senior Lecturer, 2 PG Scholar, ECE Department, Karunya Institute of Technology & Sciences Coimbatore, India 641 114

abrahamchandy@karunya.edu

 

ABSTRACT

This paper presents a gradient-based genetic algorithm for automatic location of optic disc in retinal images. The location of the optic disc is of critical importance in retinal image analysis and is required as a prerequisite stage of exudates detection, blood vessel tracking and segmentation. In the preprocessing stage, pixels with the highest 2% gray levels, simple clustering mechanism and optic disc diameter are considered to determine the candidate region and optic disc center. Genetic algorithm explores the combinatory space of possible contours (solutions) by means of crossover and mutation, followed by the evaluation of fitness and the selection of a new set of contours. The cumulative local gradient is used as a fitness function to find the fittest contour. The algorithm has been tested and the results indicate that this technique can automatically provide more acceptable location of the optic disc.

Keywords: Genetic algorithm, Optic disc, Gradient operator, Diabetic retinopathy.

 

1. Introduction

Diabetic Retinopathy remains to be the commonest cause of legal blindness in the working age population. Fortunately, more than 90 % of visual loss resulting from diabetic retinopathy can be prevented with prompt treatment if the retinopathy is detected early enough. Screening of diabetic retinopathy is thought to be an effective means of reducing preventable blindness. Color retinal photography is a very important screening approach in clinics. The development of automatic retinal image processing system facilitates the capability of ophthalmologist for the mass screening of diabetic retinopathy. The normal features of the retinal image include the optic disc, fovea and blood vessels. The optic disc is the brightest part in the normal retinal image that can be seen as a pale, well-defined round or vertically slightly oval disc. The optic disc is the entrance region of blood vessels and optic nerves to the retina and often serves as references of other features.

In [1], the optic disc is localized exploiting its high gray level variation. This approach has been shown to work well, if there are no or only few pathologies like exudates that also appear very bright and are also well contrasted. In [2], contours are detected by means of the Hough transform, i.e., the gradient of the image is calculated, and the best fitting circle is determined. This approach is quite time consuming and relies on conditions about the optic disc. In this paper, the threshold setting for the contour vertices during crossover and mutation makes the technique to work well at abnormal conditions. Also, the computational time is very much reduced. Section 2.1 deals with the approach to identify the optic disk center. Section 2.2 explains the idea of evolutionary algorithms and their application to optic disk boundary extraction. The results and conclusions are discussed in section 3 and 4 respectively.

 

 

2. Location of Optic disc

2.1 Identification of candidate regions and optic disc center

The pixels with the highest 2% gray levels in the red component are selected to localize the optic disc [3]. The nearby pixels are assembled into clusters. The clusters having centroids within a specified distance are combined to one cluster. If the number of the pixels in a cluster is less than 100 after combination, the cluster is abandoned. Finally all the remaining clusters are candidate regions. The centroid of the largest cluster is the optic disc center [4]. The possibility to miss the optic disc decreases when not only the cluster with the largest area but also all the clusters with individual area greater than 100 pixels are considered.

 

2.2 Genetic Algorithm

Genetic algorithms (GA) are adaptive methods, which may be used to solve search and optimization problems. They are based on the genetic processes of biological organisms. GAs work with a population of individuals, each representing a possible solution to a given problem. Each individual is assigned a fitness score according to how good a solution to the problem it is. The highly fit individuals are given opportunities to reproduce by cross breeding with other individuals in the population. This produces new individuals as offspring, which share some feature taken from each parent. The least fit members of the population are less likely to get selected for reproduction and so die out. A whole new population of possible solutions is thus produced by selecting the best individuals from the current generation and mating them to produce a new set of individuals. This new generation contains a higher proportion of the characteristic possessed by the good members of the previous generation. In this way over many generations good characteristics are spread throughout the population. By favouring the mating of the more fit individuals, the most promising areas of the search space are explored. If the GA has been designed well the population will converge to an optimal solution to the problem [5].

 

2.21 Initial Population

In any case, any possible contour that may count as an object boundary should be in the nearby of the outer limit of that given object. Therefore it is sensible to start the genetic algorithm by generating an initial population of solutions that lay close to the target, that is, a random set of contours that are in the nearby of the sought for contour. As the diameter of optic disc is in the range of 65~100 pixels [3] in the retinal image of 512 x 512 pixels, the initial contours are generated within this limit from the optic disc center.

 

2.22 Fitness and Selection

The fitness evaluation of a contour must quantify how close it is from the optimum. The selection is to consider the fittest (lowest cost) contour, where the cost is associated with the cumulative local intensity differences. The idea of optimal contour is a set of pixels that are both linearly connected and of very similar intensity [6]. Thus the local cost that is accumulated at each vertex of the contour diminishes when the local gradient in the image is high. The standard function to compute the local cost k (x,y) in the pixel (x,y) is

                                             (1)

The fitness f (Ci) of the ith contour is calculated by cumulating the local cost of each of its

vertices as follows

                                                                                (2)

The selection criterion is based on the elimination of those contours whose fitness is below a given relation to the maximal fitness of the present generation.

 

2.23 Crossover and Mutation

The highest fittest contour makes cross over with the neighbor highest cost contour. That is half of the vertices are swapped and the new contours are generated. While making cross over, if any of the vertices has highest variation in intensity, then its distance from the center is calculated. If the distance exceeds the threshold value (maximum diameter), then that vertex is replaced with the set value (minimum diameter) by mutation operator. The cost is again calculated for the vertices and the cumulative costs for all contours are obtained. The iterations are repeated by generating new population. The number of iterations is very much reduced due to the size of the contour, closeness to the target boundary and the application of mutation operator.

 

3. Results and Discussions

This algorithm of locating the optic disk has tested thirty five color retinal images. The experimental result in Fig.1 shows that this algorithm is successful in the tested images and can give accurate location of optic disc. Compared with the location of optic disc as the centroid of the largest cluster of brightest pixels [2], this algorithm achieves more accurate result. Among the test images, only one case fails in locating the optic disc because the exudates area is much larger than the optic disc. Selecting pixels of the highest 2% of the gray levels in the red component for candidate region is found to be optimal in most of the test images. Some test images works well for 1% the highest gray

         

              

Fig.1 (a) Original image               (b) Candidate area                    (c) Optic disc center

 

level also. Candidate region are clusters with more than hundred pixels. Initialization of population has been found important in convergence of the contour. For better location, the amount of simultaneous contours and generations need to be closer to the boundary. The number of contours considered, as initial solution space is twenty with twenty-eight vertices. Increasing the number of vertex increases the accuracy at the expense of more

              

 

 Fig.1 (d) Initial contours          (e) generated offsprings                 (f) Detected boundary

 

computation. During offspring generation the mutation operator effectively checks the position of the vertices not to exceed the optic disk boundary. Thus the number of iterations for the final contour convergence is very much reduced to less than hundred, which is usually of several thousand. The local cost that is accumulated at each vertex of the contour found to diminish when the contour approaches the optic disk boundary.

 

4. Conclusion

Image processing of color retinal images has the potential to play a major role in diagnosis of diabetic retinopathy. Gradient-based genetic algorithm has been presented for locating the optic disc. Substantial experiments have been performed, which show the accuracy and speed of this algorithm. The results are encouraging and more robust especially in the presence of light lesions. Clinical evaluation will be undertaken in order to be able to integrate the presented algorithm in a tool for diagnosis of diabetic retinopathy. In future, experiment will be carried out on more test images and images with exudates of size larger than optic disc. Such experiments could contribute to further improvements on this algorithm.

 

References

[1] C. Sinthanayothin, J. F. Boyce, H. L. Cook and T. H.Williamson, “Automated location of the optic disk, fovea, and retinal blood vessels from digital color fundus images”, British Journal of Ophthalmology, Vol. 83, No. 8, 1999, pp. 902-910. Journal of Cognitive Neuroscience, Vol. 3, No. 1, 1991, pp. 70-86.

[2] S. Tamura, Y. Okamoto and K. Yanashima, “Zero crossing interval correction in tracking eye-fundus blood vessels”, Pattern Recognition, Vol. 21, No. 3, 1988, pp.227-233.

[3] H. Li and O. Chutatape, “Fundus image features extraction,” Proceedings of the 22nd

Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Vol. 4, 2000, pp.3071 -3073.

[4] H. Li, O. Chutatape, “Automatic Location of Optic Disc in Retinal images”, IEEE ICIP, 2001, pp. 837-840.

[5] Whitley, D., “A Genetic Algorithm Tutorial “, Technical Report CS-93-103, Colorado State University, 1993.

[6] Rom´an Katz and Claudio Delrieux,” Boundary Extraction through Gradient-Based Evolutionary Algorithm”, JCS&T Vol. 3 No. 1 April 2003.

 

Technical College - Bourgas,

All rights reserved, © March, 2000