Med Biol Eng Comput DOI 10.1007/s11517-015-1330-7
ORIGINAL ARTICLE
A new approach to optic disc detection in human retinal images using the firefly algorithm Javad Rahebi1 · Fırat Hardalaç1
Received: 27 August 2014 / Accepted: 8 June 2015 © International Federation for Medical and Biological Engineering 2015
Abstract There are various methods and algorithms to detect the optic discs in retinal images. In recent years, much attention has been given to the utilization of the intelligent algorithms. In this paper, we present a new automated method of optic disc detection in human retinal images using the firefly algorithm. The firefly intelligent algorithm is an emerging intelligent algorithm that was inspired by the social behavior of fireflies. The population in this algorithm includes the fireflies, each of which has a specific rate of lighting or fitness. In this method, the insects are compared two by two, and the less attractive insects can be observed to move toward the more attractive insects. Finally, one of the insects is selected as the most attractive, and this insect presents the optimum response to the problem in question. Here, we used the light intensity of the pixels of the retinal image pixels instead of firefly lightings. The movement of these insects due to local fluctuations produces different light intensity values in the images. Because the optic disc is the brightest area in the retinal images, all of the insects move toward brightest area and thus specify the location of the optic disc in the image. The results of implementation show that proposed algorithm could acquire an accuracy rate of 100 % in DRIVE dataset, 95 % in STARE dataset, and 94.38 % in DiaRetDB1 dataset. The results of implementation reveal high capability and accuracy of proposed algorithm in the detection of the optic disc from retinal images. Also, recorded required time for the detection of the optic disc in these images is 2.13 s for DRIVE dataset, 2.81 s for STARE dataset, and 3.52 s
* Javad Rahebi [email protected] 1
Department of Electrical and Electronics Engineering, Gazi University, Ankara, Turkey
for DiaRetDB1 dataset accordingly. These time values are average value. Keywords Firefly algorithm · Optic disc detection · Retinal images
1 Introduction Along with the advancements in the technology, significant progress has been made in number and effectiveness of computerized techniques that are being used in medical studies. Automatic image processing and analysis are used in the fields of medical diagnosis and treatment and are the most promising computerized understanding and visualization techniques. These techniques provide high-resolution retinal fundus images, and most of these images can be clinically useful. In this respect, retinal fundus images are sufficient for clinical use, and many features that can be used in diagnosis and treatment can be obtained by these techniques. The automatic determination of the conditions of diabetic retinopathy (DR), age-related macular degeneration, glaucoma, etc., can be counted among the most important benefits that the advancements in image processing have provided to the field of ophthalmology. Diseases such as DR are difficult to observe and diagnose because they initially do no cause any vision disorder. Therefore, the identification of such diseases requires regular eye examinations. Due to resource concerns, the regular examination of large numbers of people is no possible with traditional methods. However, these examinations can be performed with automatic systems. Additionally, the diagnosis and selection of treatment for such diseases is slow when traditional methods are employed because physicians cannot quickly examine large numbers of people or retinal
13
Med Biol Eng Comput
Fig. 1 a Original retinal image and its properties, b illustration of RIM and CUP location
images. Therefore, development of automatic systems is quite important [3, 8, 16, 19, 27, 29]. When retinal images are controlled automatically by the system, they can be examined and diagnosed by professional physicians when any suspicion of disease is encountered by the automatic control. The ability to effectively and successfully detect the optic disc is a priority for retinal images, and the physical properties of the optic disc that constitute the input for detection programs, such as brightness and width, should be carefully analyzed. Currently, the brightness of the optic disc and its combinations with the blood vessels can be shown to be the most important properties [2]. The shape of the optic disc is akin to an ellipse, and the sizes of optic discs are 1.8 ± 0.2 mm in width and 1.9 ± 0.2 mm in length [24]. These sizes have been reported to average 1.5 mm [13]. In the optic disc region in which the nerve fibers are located, a relatively brighter region is located in the center of the optic disc called the CUP (Fig. 1). The edge of the optic disc (RIM) forms a narrow band that surrounds the optic disc (Fig. 1). However, there can be large differences between fundus images. The optic disc can occasionally seem much brighter than its surroundings, but occasionally the disc can appear as a hollow ring. Sometimes, bright regions outside of the rim region are present due to disease. Moreover, the blood vessels can partly alter the image of the optic disc [15]. The algorithm proposed in [20] is based on ant colony optimization (ACO) and uses an anisotropic diffusion process that aims to smooth the retinal blood vessels. The ACO algorithm can detect the optic disc edges. Indeed, “the ACO algorithm could be efficient in extracting other features in the image such as the major blood vessels and macula. The diffusion process applied in the methodology
13
described here is an important way to segment only the optic disc and, consequently, to distinguish it from the other features also detected when only the ACO algorithm was used. Then, the localization of the optic disc by a point was a simple task. In addition, the ACO algorithm implemented in a parallel way should be exploited to reduce the computational load of the proposed method.” The methods used to determine and segment the optic disc vary from between morphological methods in which the brightness and geometric characteristics are employed with statistical methods, such as Bayesian methods, and the standard methods, such as Hough transformations and template matching. Based on these approaches, studies of the determination of the optic disc can be grouped in different manners. While some studies related to this issue have taken into account the numbers of blood vessels on the optic disc [2], some methods also include the assessment of the contours of the disc region [5, 12]. While some methods employ template-matching methods [13], other methods employ active contour or snake methods [18]. There are also approaches that employ machine learning [14], multilevel threshold, and shape specification methods [23, 25]. This classification is based on a general point of view, and there are studies that have employed different methods in hybrid manners [1, 22]. Some examples of the studies that have identified and segmented the optic disc are given below. In the study carried out by Fleming et al. [6], the optic disc center was determined by using a half ellipse. This study reported an optic disc detection success rate of 98.4 % but provided no success rate for the detection of the optic disc contours. In contrast, Carmona et al. employed a genetic algorithm to detect the optic disc and identify its contours. Moreover, for the pictures with less than 5 pixel of uncertainty,
Med Biol Eng Comput
a success ratio of 96 % was reported for the identification of the contours of the optic disc; these results were based on tests performed on 110 images. Lupascu et al. [16] identified the geometric location and contours of the optic disc. This study reported optic disc detection success rate of 95 % and an optic disc contour extraction success rate of 70 %. In this study, tissue identification and regressionbased methods were used to attempt to identify the most appropriate circle that confined the optic disc. Chaichana et al. [4] employed the Sobel filter to reveal the sides in the image and applied the CHT to specify the most circular structure among the specified sides. This study successfully detected the optic disc location in 39 of 40 retinal images from the DRIVE retina dataset (97.5 %). The authors of [31] detected the optic disc location by taking the brightest regions in the retinal images and the circular images they had identified with CHT as separate candidates and then comparing the candidates. With this method, these authors achieved a success rate of 85 % with retinal images taken from the Ophthalmology Department of Black Sea Technical University, Faculty of Medicine. Morales et al. [17] first transformed the retinal images to gray scale with principal component analysis (PCA) and then detected the optic disc location and segmented the sides using the stochastic watershed algorithm (segmentation). With this method, these authors achieved an 86.89 % optic disc location detection and side segmentation success rate with 110 retinal images in the DRINOS retina dataset. The main goal of this paper was implementing firefly algorithm to detecting the optic disc in retinal fundus images. The highest rates of lighting in the retinal images are related to the optic disc. Therefore, all of the insects can be move to this location, and the location of the optic disc in the image can be identified. Therefore, we used firefly algorithm for optic disc detection. In Sect. 2, firefly algorithm, image preprocessing, and related flowchart for applying the algorithm on retinal images are discussed. In Sect. 3, simulation results in DRIVE, STARE, and DiaRetDB1 datasets are evaluated thoroughly. In Sect. 4, related discussions on the results of the present paper and other papers are represented. Finally, conclusion is found in Sect. 5.
2 Materials and methods
colonies, and particles. The firefly intelligent algorithm is a new intelligent algorithm that was inspired by the firefly social behavior. This algorithm was initially introduced at Cambridge University by Yang in 2008 [30]. The population in this algorithm includes fireflies, each of which has a specific rate of lighting or fitness. Similar to other intelligent optimization methods, this method begins with an initial population of insects. 2.2 The algorithm structure For simplicity, the following three rules are considered in this algorithm: 1. All fireflies are unisex so that one firefly is attracted to other fireflies regardless of their sex. 2. The attractiveness is related to the rate of lighting; if there is no light, each firefly moves randomly. 3. The lighting of firefly is either virtual or produced by a criterion function. For example, in terms of a maximization problem, the rate of lighting is in accordance with the value of the criterion function. 2.3 Attraction There are two problems in the firefly algorithm: 1. The difference in light intensity 2. Formulating the rate of attraction For simplicity, we always assume that the attractiveness of a firefly is determined by its lighting, which in turn depends on the criterion function. For an environment with constant light absorption coefficient γ, we can calculate I as follows:
I = I0 e−γ r
(1)
where I0 is the source light intensity, and γ is the light absorption coefficient.As a firefly’s attractiveness is proportional to the light intensity seen by adjacent fireflies, the attractiveness β of a firefly is defined by:
β = β0 e−γ r
2
(2)
where β0 is attraction at r = 0. 2.4 Distance and movement
2.1 Firefly algorithm Generally, the intelligent algorithms that are inspired by nature are recognized as some of the most powerful methods for solving optimization problems. These intelligent algorithms include algorithms based on ant colonies, bee
The distances between each pair of fireflies, i and j, which are stated as xi and xj, are calculated as follows: d (3) (xi,k − xj,k )2 rij = xi − xj = k=1
13
Med Biol Eng Comput
Table 1 Parameter values of the proposed algorithm Parameters
Value
t, Number of iterations
15
K, Number of fireflies
100
γ, Absorption coefficient β0, Attractiveness at r = 0
0.06 0.9
α, Controlling the step size
0.52
delta, Randomness reduction (similar to an annealing schedule)
0.77
where xi,k is the K-th component in the distance coordinate xi from the i-th firefly. In the two-dimensional case, the equation is as follows: rij = (xi − xj )2 + (yi − yj )2 (4) The movement of the firefly i toward the stronger firefly from j is defined as follows: 1 −γ rij2 xi = xi + β0 e (xj − xi ) + α r − (5) 2 The second term of this equation indicates the rate of attraction, and the third term is the rate of randomization determined by parameter α, and r is a random number drawn from a uniform distribution between zero and one. In simulations, the values of α, β0, and γ were selected between 0 and 1. These values depend on the results of simulation and are acquired empirically which is presented in Table 1. Moreover, the randomization term can be replaced with a normal distribution function. 2.5 Preprocessing phase Usually, retinal images do not have high quality, and there is some noise regarding the process of imaging. Therefore, there is a need for preprocessing these images. The objectives in this phase are image enhancement, noise
elimination, smoothing, etc. The most commonly used filter for the elimination of random noise is the median filter. The filter functions by replacing a desired pixel’s value with the median of the adjacent pixels. The median filter creates less blurriness in an image compared to linear smoothing filters of the same size [7]. Unlike the neighborhood averaging filter (i.e., the mean filter), the median filter performs this task without blurring the sharp edges on the border of the retina. Basically, the median filter replaces the value of a pixel f (x, y) with the median of all of the pixels in the neighborhood of that pixel as shown in Eq. (6).
fmed (x, y) = median{f (s, t)}, (s, t) ∈ Wxy
(6)
where W represents the neighborhood centered on location (x, y) in the image. In this paper, median filter with 15 × 15 size is used. Each output pixel contains the median value in the 15 × 15 neighborhood around the corresponding pixel in the input image. Figure 2b–d shows the light intensities of a retinal image before and after the application of a median filter. 2.6 Applying the firefly algorithm to identify the optic disc The primary positions of fireflies are applied randomly on the image. In the present study, 100 fireflies were used. The fireflies were put randomly on the place of pixels. Each of the fireflies which showed the highest rate and value regarding pixel intensity showed that they have been placed in a better position in contrast to the other ones. In the proposed firefly algorithm, when a particle is compared with another particle (here, particle is equivalent to pixel), if the first particle is in a better situation than the second particle, the particles move to new locations that are determined by the location of the particle in the better situation and achieve the total optimum particle, which is determined
Fig. 2 a Main retinal image (the image of the gray level), b the resulting image after applying the filter, c 3D illustration of the main image’s light intensity, d 3D illustration of the image’s light intensity after applying the median filter
13
Med Biol Eng Comput
according to the locations of the two particles. In this algorithm regarding the distance of each particle from the particles being compared, Eq. (7) is used to determine the distance of each particle from the total optimum particle. (7) ri,Gbest = (xi − XGbest )2 + (yi − YGbest )2 Furthermore, to determine the new location of the particle, Eq. (8) is used.
2
xi = xi + B0 e−γ ri,Gbest xj − xi 1 +α r− 2
2
+ B0 e−γ ri,Gbest (xGbest − xi )
Initialization Process Applying the Median filter on the image
Initialization Firefly algorithm parameters
Randomly dispatch insects on the image
Evaluating each insect's light intensity, that is obtained according to the location of each insect in the pixel
(8)
Indeed, from Eq. (8), the resultant between the compared particle and the total optimum particle is obtained. In Eq. (8), the Gbest index is the pixel index that has most probability value to be selecting than the other pixels, and XGbest is the location of the best particle. By changing the trend of particle movements, the group movement of particles is enhanced, and the particles can more accurately search the area. Search among fireflies is the manner in which each firefly is compared with each other firefly. To identify the maximum point, if a firefly exhibits a lower lighting rate than the compared firefly, the said firefly moves toward the lighter firefly, and if there is a particle with better lighting in the next iteration of the algorithm, the particles again move toward the particle with the better lighting. The deficiency of this method is that the movement of the particles is determined by the rate of firefly’s lighting, which is the optimum location, and the total optimum has no effect on the algorithm search. Due to this deficiency, the entire area of the problem is not optimally searched, and a greater number of iterations are required to achieve the optimum point. This deficiency is resolved during the fireflies’ movements by improving the act of search in which the other fireflies participate. However, to improve this algorithm, it can also be shown that β0, which indicates the rate of light source absorption as used in Eq. (5), presents the same value for both the local optimum and the total optimum. In retinal images, existing high intensified radiant places which are not considered as optic disc disturb firefly algorithm in detecting the optic disc. In this study, we changed the values of β0 according to the transmitted light from the fireflies and different situations. Also, the corresponding values of each firefly are the difference between transmitted light (pixel intensity value) in previous iteration and the highest intensity rate of each firefly. In this manner, the value of β0 for each firefly in every step is equal to the value of each firefly divided by the average value of that firefly. With this method, we can obtain better values for the functions.
Comparing the light intensity of each insect
The movement of all insects towards the pixel with the lightest intensity
Updating the situation of insects in image pixels Updating the best insects Updating the randomization value of equation
Whether N iteration takes place?
No
Yes All insects are attracted to the optic disc Fig. 3 A summary of the implementation of the proposed firefly algorithm approach for identifying the optic disc in retina
2.7 Summary of optic disc detection in retinal images with the firefly algorithm A summary of the optic disc detection in the retinal images based on the firefly algorithm is presented in Fig. 3. 2.8 Dataset The DRIVE dataset contains 40 colorful retinal images. There are some symptoms of deficiency in seven of these images. These images are taken with a Canon CR5
13
nonmydriatic 3CCD camera at an angle of 40° from which the image of the eye is a circle with the diameter of 540 pixels. The size of the images in this dataset is 584 × 565, and the images are in JPEG format. The STARE dataset includes 20 colorful retinal images. Some symptoms of deficiency are present in 10 of these images. These images were taken using the Topcon TRV_50 fundus camera at an angle of 35°. The size of the images in this dataset is 605 × 700, and the images are in TIFF format. The DiaRetDB1 dataset consists of 89 color fundus images of which 84 contain at least mild nonproliferative signs of diabetic retinopathy, and five are considered normal and do not contain any signs of diabetic retinopathy according to all of the experts who participated in the evaluation of these images. The images were captured using the same 50° field-of-view digital fundus camera with different imaging settings. The size of the images in this dataset is 1152 × 1500, and the images are in PNG format. The proposed method for analyzing the dataset images was implemented using the MATLAB 2013a programming language and run on a PC with an Intel Duo CPU, 2.00 GHz RAM machine.
Med Biol Eng Comput
algorithm for the DRIVE, STARE, and DiaRetDB1 datasets are shown in Table 2. Notably, the elapsed times obtained with the proposed method were achieved without any alterations to the sizes of the input images. To achieve a performance criterion for this paper, the criterion for the correct localization of the optic disc was defined based on the evaluations of the probable optic disc centers compared to the manually selected centers. If the difference between the center of the optic disc, which is specified manually, and acquired position with firefly algorithm is less than the optic disc radius, then it can be said that we have detected the place of the optic disc accurately. The proposed optic disc localization method achieved a success rate of 95 % in the STARE dataset (i.e., the optic disc was correctly identified in 19 of the 20 images). The average distance between the probable optic disc center and the hand-selected center was 20.93. Moreover, the optic disc was successfully identified in all of the 40 DRIVE dataset images with the proposed method. The average distance between the estimated optic disc center and the hand-selected center was 9.5. The distances between the true locations and the estimated locations are illustrated in Fig. 7. Moreover, the success rates of this method are compared with those of other localization methods in Table 3.
3 Results Experiments were conducted to demonstrate the performance of the proposed approach. We tested the proposed method on three publicly available datasets, the DRIVE dataset [26], the STARE dataset [9], and the DiaRetDB1 dataset [11]. The results are shown in Figs. 4, 5, and 6. In the first step, the proposed method parameters were experimentally determined, and the results are shown in Table 1. To evaluate and quantify the performance of the proposed method, the optic disc centers and diameters were determined manually for all images. The estimated optic disc center was deemed acceptable if it was located within the circular optic disc area, i.e., if the distance between the estimated location and the hand-segmented location was smaller than the half of the hand determined diameter. Figures 4, 5, and 6 show the results of this process as applied to the DRIVE, DiaRetDB1, and STARE dataset images. The optic disc location estimations are represented by crosses (×). We used quantitative results for all the images of the datasets, and this is shown in Fig. 7, where each point represents the distance from the true location and the estimated location. From these figures, it can be observed that the proposed method could be an effective tool for roughly detecting the optic disc in retinal images. The times required to find the optic disc using the firefly
13
4 Discussion In this paper, the firefly algorithm was successfully applied to the detection of the optic discs in retinal images. From Figs. 4, 5, and 6, it can be seen that our method overcame the main problems that characterize retinal images obtained from screening programs. The algorithm was applied to pathological images and, in most cases, registered those pathologies and produced good results. Essentially, this algorithm can be affected by factors determined during the acquisition of the image that make optical disc detection more difficult, such as poor image contrast, the existence of high contrast lesions, and brighter effects. Exudate lesions, artifacts, and retinal sheen can be very bright and might lead to false detection of the optic disc (OD). We used median filters to resolve this problem. In the initialization process, the insects are dispatched randomly over the image. If this random dispatch does not place any insect inside the true OD region, the location is not detected. However, across all of the simulations (i.e., in all of the tests), we did not encounter this problem, as at least one insect was always placed inside the optic disc. The initial emission intensities of the insects were the intensities of the pixels that they had been placed on and did not vary with image size.
Med Biol Eng Comput
Fig. 4 Results of applying the proposed method to 20 DRIVE dataset images. The estimated localization of the optic disc is marked with a cross (×)
Fig. 5 Results of applying the proposed method to 10 selected DiaRetDB1 dataset images. The estimated localization of the optic disc is marked with a cross (×)
13
Med Biol Eng Comput
Fig. 6 Results of applying the proposed method to five selected STARE dataset images. The estimated localization of the optic disc is marked with a cross (×)
Fig. 7 Quantitative results for all the images of the datasets used, where each point represents the distance from the true location and the estimated location Table 2 Computational timing comparison Dataset
Number of test images
Optic disc detection time (s)
DRIVE STARE
40 20
2.13 2.81
DiaRetDB1
89
3.52
Table 3 Success rates of optic disc localization methods
13
With the proposed approach, detecting the optical disc is easy. Although detecting the optic disc completely is useful in situations such as glaucoma diagnosis, this was not the reason of this study. The main purpose of the present study was to develop and implement a natural inspired algorithm in detecting the place of optic disc in human retinal images. In this application, the proposed method achieved success rates of 94.38, 95, and 100 % in the DiaRetDB1, STARE, and DRIVE datasets, respectively. The results achieved by the application of the proposed method to the STARE and DiaRetDB1 datasets are as good as or better than those described in previous works [10, 20, 21, 24, 28]. Due to the importance of using real images, we intend to develop an authoritative system and clinical applications. In this regard, our use of the latter two datasets can be considered an advantage of the present work compared to other methods that can be found in the literature. Our proposed approach can be considered an essential step toward the development of computer-aided diagnostic (CAD) systems based on the application of the firefly algorithm to normal screening programs that seek to diagnose diabetic retinopathy. Indeed, based on the obtained experimental results, the firefly algorithm efficiently located the optic disc. Additionally, the parallel implementation of the firefly algorithm should be exploited to decrease the computational load observed in this study.
Localization methods
DRIVE dataset (%)
STARE dataset (%)
DiaRetDB1 (%)
Sinthanayothin et al. [24] Walter et al. [28] Rashid Qureshi et al. [21] Hung-Kuei Hsiao et al. [10] Carla Pereira et al. [20]
60 80 100 100 100
50 75 – 90 –
– – 94.02 – 93.25
Proposed method
100
95
94.38
Med Biol Eng Comput
5 Conclusion In this paper, we used the firefly algorithm to identify the location of the optic disc in human retinal images. With this method, some insects are initially randomly applied to the retina image, and each pixel in which an insect is placed is considered as a firefly lighting. Then, the insects are compared two by two, and insects that are less attractive move toward more attractive insects. Finally, a single insect is selected as the most attractive, and this insect represents the optimum response to the problem in question. Two important issues involved in this method are changes in the light intensities of the image pixels and the formulation of the insect’s attractions. The light intensity is proportional to the insect’s fitness. Generally, attraction is a relative parameter and is estimated by the other insects. Moreover, attraction depends on the distance between the insects. Because the highest rates of lighting in the retina are related to the optic disc, all of the insects move to this location, and the location of the optic disc in the image can thus be identified.
References 1. Abdel-Haleim A, Abdel-Razik Y, Ghalwash AZ, Sabry AA, Abdel-Rahman G (2008) Optic disc detection from normalized digital fundus images by means of a vessels’ direction matched filter. IEEE Trans Med Imaging 27:11–18 2. Akita K, Kuga H (1982) A computer method of understanding ocular fundus images. Pattern Recognit 15:431–443 3. Cassel GH, Billig MD, Randall HG (2001) The eye book: a complete guide to eye disorders and health. Johns Hopkins University Press, Baltimore 4. Chaichana T, Yoowattana S, Sun Z, Tangjitkusolmun S, Sookpotharom S, Sangworasil M (2008) Edge detection of the optic disc in retinal images based on identification of a round shape. Communications and information technologies, international symposium, pp 670–674 5. Cox MJ, Wood ICJ (1991) Computer-assisted optic nerve head assessment. Ophthalmic Physiol Opt 11:27–35 6. Fleming AD, Goatman KA, Philip S, Olson JA, Sharp PF (2007) Automatic detection of retinal anatomy to assist diabetic retinopathy screening. Phys Med Biol 52:331–345 7. Gonzales R, Woods C, Eddins RE (2004) Digital image processing. Prentice-Hall, Inc 8. Hoover A, Goldbaum M (2003) Locating the optic nerve in retinal image using the fuzzy convergence of the blood vessels. IEEE Trans Med Imaging 22:951–958 9. Hoover A, Kouznetsova V, Goldbaum M (2000) Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Trans Med Imaging 19(3):203–210 10. Hsiao H-K, Liu C-C, Yu C-Y, Kuo S-W, Yu S-S (2012) A novel optic disc detection scheme on retinal images. Expert Syst Appl 39(12):10600–10606 11. Kauppi T, Kalesnykiene V, Kamarainen J-K, Lensu L, Sorri I, Raninen A, Voutilainen R, Uusitalo H, Kälviäinen H, Pietilä J (2006) DIARETDB1 diabetic retinopathy database and evaluation protocol, Technical report 12. Kong HJ, Kim SK, Seo JM, Park KH, Chung H, Park KS (2004) Three dimensional reconstruction of conventional stereo optic
disc image. Annual international conference of the IEEE EMBS, Vol 12. San Francisco, pp 29–32 13. Lalonde M, Beaulieu M, Gagnon L (2001) Fast and robust optic disc detection using pyramidal decomposition and Hausdorffbased template matching. IEEE Trans Med Imaging 20:1193–1200 14. Li H, Chutatape O (2001) Automatic location of optic disc in retinal images. IEEE ICIP, Thessaloniki, pp 837–840 15. Lowell J, Hunter A, Steel D, Basu A, Ryder R, Fletcher E, Kennedy L (2004) Optic nerve head segmentation. IEEE Trans Med Imaging 23:256–264 16. Lupascu CA, Tegolo D, Rosa LD (2008) Automated detection of optic disc location in retinal images. 21st IEEE international symposium on computer-based medical systems, Finland, pp 17–22 17. Morales S, Naranjo V, Perez D, Navea A, Alcaniz M (2012) Automatic detection of optic disc based on PCA and stochastic watershed. In: Signal processing conference (EUSIPCO), proceedings of the 20th European, Bucharest, pp 2605–2609 18. Osareh A, Mirmehdi M, Thomas B, Markham, R (2002) Colour morphology and snakes for optic disc localization. The 6th medical image understanding and analysis conference, Vol 1, pp 21–24 19. Patton N, Aslam TM, MacGillivray T, Deary IJ, Dhillon B, Eikelboom RH, Yogesan K, Constable IJ (2006) Retinal image analysis: concepts, applications and potential. Prog Retin Eye Res 25:99–127 20. Pereira C, Gonçalves L, Ferreira M (2013) Optic disc detection in color fundus images using ant colony optimization. Med Biol Eng Comput 51:295–303 21. Qureshi RJ, Kovacs L, Harangi B, Nagy B, Peto T, Hajdu A (2012) Combining algorithms for automatic detection of optic disc and macula in fundus images. Comput Vis Image Underst 116:138–145 22. Reza AW, Eswaran C, Hati S (2008) Automatic tracing of optic disc and exudates from color fundus images using fixed and variable thresholds. J Med Syst 33:73–80 23. Sekhar S, Al-Nuaimy W, Nandi AK (2008) Automated localization of retinal optic disc using Hough transform. The 5th IEEE international symposium on biomedical imaging: from nano to macro, Paris, pp 77–80 24. Sinthanayothin C, Boyce JF, Cook HL, Williamson TH (1999) Automated localization of the optic disc, fovea, and retinal blood vessels from digital color fundus images. Br J Ophthalmol 83:902–910 25. Sopharak A, Uyyanonvara B, Barman S, Williamson TH (2009) Automatic detection of diabetic retinopathy exudates from nondilated retinal images using mathematical morphology methods. Comput Med Imaging Graph 32:720–727 26. Staal J, Abàmoff MD, Niemeijer M, Viergever MA, van Ginneken B (2004) Ridgebased vessel segmentation in color images of the retina. IEEE Trans Med Imaging 23(4):501–509 27. Tobin KW, Chaum E, Govindasamy VP, Karnowski T (2007) Detection of anatomic structures in human retinal imagery. IEEE Trans Med Imaging 26:1729–1739 28. Walter T, Klein JC, Massin P, Erginary A (2002) A contribution of image processing to the diagnosis of diabetic retinopathydetection of exudates in color fundus images of human retina. IEEE Trans Med Imaging 21:1236–1243 29. Welfer D, Scharcanski J, Kitamura C, Pizzol MD, Ludwig L, Marinho D (2010) Segmentation of the optic disc in color eye fundus images using an adaptive morphological approach. Comput Biol Med 40:124–137 30. Yang XS (2008) Nature-inspired metaheuristic algorithms. Luniver Press, Frome 31. Yavuz Z, ˙Ikibas¸ C, S¸evik U, Köse C (2009) A method for automatic optic disc extraction in retinal fundus images. 5th International advanced technologies symposium, Karabuk, pp 93–98
13
ORIGINAL ARTICLE
A new approach to optic disc detection in human retinal images using the firefly algorithm Javad Rahebi1 · Fırat Hardalaç1
Received: 27 August 2014 / Accepted: 8 June 2015 © International Federation for Medical and Biological Engineering 2015
Abstract There are various methods and algorithms to detect the optic discs in retinal images. In recent years, much attention has been given to the utilization of the intelligent algorithms. In this paper, we present a new automated method of optic disc detection in human retinal images using the firefly algorithm. The firefly intelligent algorithm is an emerging intelligent algorithm that was inspired by the social behavior of fireflies. The population in this algorithm includes the fireflies, each of which has a specific rate of lighting or fitness. In this method, the insects are compared two by two, and the less attractive insects can be observed to move toward the more attractive insects. Finally, one of the insects is selected as the most attractive, and this insect presents the optimum response to the problem in question. Here, we used the light intensity of the pixels of the retinal image pixels instead of firefly lightings. The movement of these insects due to local fluctuations produces different light intensity values in the images. Because the optic disc is the brightest area in the retinal images, all of the insects move toward brightest area and thus specify the location of the optic disc in the image. The results of implementation show that proposed algorithm could acquire an accuracy rate of 100 % in DRIVE dataset, 95 % in STARE dataset, and 94.38 % in DiaRetDB1 dataset. The results of implementation reveal high capability and accuracy of proposed algorithm in the detection of the optic disc from retinal images. Also, recorded required time for the detection of the optic disc in these images is 2.13 s for DRIVE dataset, 2.81 s for STARE dataset, and 3.52 s
* Javad Rahebi [email protected] 1
Department of Electrical and Electronics Engineering, Gazi University, Ankara, Turkey
for DiaRetDB1 dataset accordingly. These time values are average value. Keywords Firefly algorithm · Optic disc detection · Retinal images
1 Introduction Along with the advancements in the technology, significant progress has been made in number and effectiveness of computerized techniques that are being used in medical studies. Automatic image processing and analysis are used in the fields of medical diagnosis and treatment and are the most promising computerized understanding and visualization techniques. These techniques provide high-resolution retinal fundus images, and most of these images can be clinically useful. In this respect, retinal fundus images are sufficient for clinical use, and many features that can be used in diagnosis and treatment can be obtained by these techniques. The automatic determination of the conditions of diabetic retinopathy (DR), age-related macular degeneration, glaucoma, etc., can be counted among the most important benefits that the advancements in image processing have provided to the field of ophthalmology. Diseases such as DR are difficult to observe and diagnose because they initially do no cause any vision disorder. Therefore, the identification of such diseases requires regular eye examinations. Due to resource concerns, the regular examination of large numbers of people is no possible with traditional methods. However, these examinations can be performed with automatic systems. Additionally, the diagnosis and selection of treatment for such diseases is slow when traditional methods are employed because physicians cannot quickly examine large numbers of people or retinal
13
Med Biol Eng Comput
Fig. 1 a Original retinal image and its properties, b illustration of RIM and CUP location
images. Therefore, development of automatic systems is quite important [3, 8, 16, 19, 27, 29]. When retinal images are controlled automatically by the system, they can be examined and diagnosed by professional physicians when any suspicion of disease is encountered by the automatic control. The ability to effectively and successfully detect the optic disc is a priority for retinal images, and the physical properties of the optic disc that constitute the input for detection programs, such as brightness and width, should be carefully analyzed. Currently, the brightness of the optic disc and its combinations with the blood vessels can be shown to be the most important properties [2]. The shape of the optic disc is akin to an ellipse, and the sizes of optic discs are 1.8 ± 0.2 mm in width and 1.9 ± 0.2 mm in length [24]. These sizes have been reported to average 1.5 mm [13]. In the optic disc region in which the nerve fibers are located, a relatively brighter region is located in the center of the optic disc called the CUP (Fig. 1). The edge of the optic disc (RIM) forms a narrow band that surrounds the optic disc (Fig. 1). However, there can be large differences between fundus images. The optic disc can occasionally seem much brighter than its surroundings, but occasionally the disc can appear as a hollow ring. Sometimes, bright regions outside of the rim region are present due to disease. Moreover, the blood vessels can partly alter the image of the optic disc [15]. The algorithm proposed in [20] is based on ant colony optimization (ACO) and uses an anisotropic diffusion process that aims to smooth the retinal blood vessels. The ACO algorithm can detect the optic disc edges. Indeed, “the ACO algorithm could be efficient in extracting other features in the image such as the major blood vessels and macula. The diffusion process applied in the methodology
13
described here is an important way to segment only the optic disc and, consequently, to distinguish it from the other features also detected when only the ACO algorithm was used. Then, the localization of the optic disc by a point was a simple task. In addition, the ACO algorithm implemented in a parallel way should be exploited to reduce the computational load of the proposed method.” The methods used to determine and segment the optic disc vary from between morphological methods in which the brightness and geometric characteristics are employed with statistical methods, such as Bayesian methods, and the standard methods, such as Hough transformations and template matching. Based on these approaches, studies of the determination of the optic disc can be grouped in different manners. While some studies related to this issue have taken into account the numbers of blood vessels on the optic disc [2], some methods also include the assessment of the contours of the disc region [5, 12]. While some methods employ template-matching methods [13], other methods employ active contour or snake methods [18]. There are also approaches that employ machine learning [14], multilevel threshold, and shape specification methods [23, 25]. This classification is based on a general point of view, and there are studies that have employed different methods in hybrid manners [1, 22]. Some examples of the studies that have identified and segmented the optic disc are given below. In the study carried out by Fleming et al. [6], the optic disc center was determined by using a half ellipse. This study reported an optic disc detection success rate of 98.4 % but provided no success rate for the detection of the optic disc contours. In contrast, Carmona et al. employed a genetic algorithm to detect the optic disc and identify its contours. Moreover, for the pictures with less than 5 pixel of uncertainty,
Med Biol Eng Comput
a success ratio of 96 % was reported for the identification of the contours of the optic disc; these results were based on tests performed on 110 images. Lupascu et al. [16] identified the geometric location and contours of the optic disc. This study reported optic disc detection success rate of 95 % and an optic disc contour extraction success rate of 70 %. In this study, tissue identification and regressionbased methods were used to attempt to identify the most appropriate circle that confined the optic disc. Chaichana et al. [4] employed the Sobel filter to reveal the sides in the image and applied the CHT to specify the most circular structure among the specified sides. This study successfully detected the optic disc location in 39 of 40 retinal images from the DRIVE retina dataset (97.5 %). The authors of [31] detected the optic disc location by taking the brightest regions in the retinal images and the circular images they had identified with CHT as separate candidates and then comparing the candidates. With this method, these authors achieved a success rate of 85 % with retinal images taken from the Ophthalmology Department of Black Sea Technical University, Faculty of Medicine. Morales et al. [17] first transformed the retinal images to gray scale with principal component analysis (PCA) and then detected the optic disc location and segmented the sides using the stochastic watershed algorithm (segmentation). With this method, these authors achieved an 86.89 % optic disc location detection and side segmentation success rate with 110 retinal images in the DRINOS retina dataset. The main goal of this paper was implementing firefly algorithm to detecting the optic disc in retinal fundus images. The highest rates of lighting in the retinal images are related to the optic disc. Therefore, all of the insects can be move to this location, and the location of the optic disc in the image can be identified. Therefore, we used firefly algorithm for optic disc detection. In Sect. 2, firefly algorithm, image preprocessing, and related flowchart for applying the algorithm on retinal images are discussed. In Sect. 3, simulation results in DRIVE, STARE, and DiaRetDB1 datasets are evaluated thoroughly. In Sect. 4, related discussions on the results of the present paper and other papers are represented. Finally, conclusion is found in Sect. 5.
2 Materials and methods
colonies, and particles. The firefly intelligent algorithm is a new intelligent algorithm that was inspired by the firefly social behavior. This algorithm was initially introduced at Cambridge University by Yang in 2008 [30]. The population in this algorithm includes fireflies, each of which has a specific rate of lighting or fitness. Similar to other intelligent optimization methods, this method begins with an initial population of insects. 2.2 The algorithm structure For simplicity, the following three rules are considered in this algorithm: 1. All fireflies are unisex so that one firefly is attracted to other fireflies regardless of their sex. 2. The attractiveness is related to the rate of lighting; if there is no light, each firefly moves randomly. 3. The lighting of firefly is either virtual or produced by a criterion function. For example, in terms of a maximization problem, the rate of lighting is in accordance with the value of the criterion function. 2.3 Attraction There are two problems in the firefly algorithm: 1. The difference in light intensity 2. Formulating the rate of attraction For simplicity, we always assume that the attractiveness of a firefly is determined by its lighting, which in turn depends on the criterion function. For an environment with constant light absorption coefficient γ, we can calculate I as follows:
I = I0 e−γ r
(1)
where I0 is the source light intensity, and γ is the light absorption coefficient.As a firefly’s attractiveness is proportional to the light intensity seen by adjacent fireflies, the attractiveness β of a firefly is defined by:
β = β0 e−γ r
2
(2)
where β0 is attraction at r = 0. 2.4 Distance and movement
2.1 Firefly algorithm Generally, the intelligent algorithms that are inspired by nature are recognized as some of the most powerful methods for solving optimization problems. These intelligent algorithms include algorithms based on ant colonies, bee
The distances between each pair of fireflies, i and j, which are stated as xi and xj, are calculated as follows: d (3) (xi,k − xj,k )2 rij = xi − xj = k=1
13
Med Biol Eng Comput
Table 1 Parameter values of the proposed algorithm Parameters
Value
t, Number of iterations
15
K, Number of fireflies
100
γ, Absorption coefficient β0, Attractiveness at r = 0
0.06 0.9
α, Controlling the step size
0.52
delta, Randomness reduction (similar to an annealing schedule)
0.77
where xi,k is the K-th component in the distance coordinate xi from the i-th firefly. In the two-dimensional case, the equation is as follows: rij = (xi − xj )2 + (yi − yj )2 (4) The movement of the firefly i toward the stronger firefly from j is defined as follows: 1 −γ rij2 xi = xi + β0 e (xj − xi ) + α r − (5) 2 The second term of this equation indicates the rate of attraction, and the third term is the rate of randomization determined by parameter α, and r is a random number drawn from a uniform distribution between zero and one. In simulations, the values of α, β0, and γ were selected between 0 and 1. These values depend on the results of simulation and are acquired empirically which is presented in Table 1. Moreover, the randomization term can be replaced with a normal distribution function. 2.5 Preprocessing phase Usually, retinal images do not have high quality, and there is some noise regarding the process of imaging. Therefore, there is a need for preprocessing these images. The objectives in this phase are image enhancement, noise
elimination, smoothing, etc. The most commonly used filter for the elimination of random noise is the median filter. The filter functions by replacing a desired pixel’s value with the median of the adjacent pixels. The median filter creates less blurriness in an image compared to linear smoothing filters of the same size [7]. Unlike the neighborhood averaging filter (i.e., the mean filter), the median filter performs this task without blurring the sharp edges on the border of the retina. Basically, the median filter replaces the value of a pixel f (x, y) with the median of all of the pixels in the neighborhood of that pixel as shown in Eq. (6).
fmed (x, y) = median{f (s, t)}, (s, t) ∈ Wxy
(6)
where W represents the neighborhood centered on location (x, y) in the image. In this paper, median filter with 15 × 15 size is used. Each output pixel contains the median value in the 15 × 15 neighborhood around the corresponding pixel in the input image. Figure 2b–d shows the light intensities of a retinal image before and after the application of a median filter. 2.6 Applying the firefly algorithm to identify the optic disc The primary positions of fireflies are applied randomly on the image. In the present study, 100 fireflies were used. The fireflies were put randomly on the place of pixels. Each of the fireflies which showed the highest rate and value regarding pixel intensity showed that they have been placed in a better position in contrast to the other ones. In the proposed firefly algorithm, when a particle is compared with another particle (here, particle is equivalent to pixel), if the first particle is in a better situation than the second particle, the particles move to new locations that are determined by the location of the particle in the better situation and achieve the total optimum particle, which is determined
Fig. 2 a Main retinal image (the image of the gray level), b the resulting image after applying the filter, c 3D illustration of the main image’s light intensity, d 3D illustration of the image’s light intensity after applying the median filter
13
Med Biol Eng Comput
according to the locations of the two particles. In this algorithm regarding the distance of each particle from the particles being compared, Eq. (7) is used to determine the distance of each particle from the total optimum particle. (7) ri,Gbest = (xi − XGbest )2 + (yi − YGbest )2 Furthermore, to determine the new location of the particle, Eq. (8) is used.
2
xi = xi + B0 e−γ ri,Gbest xj − xi 1 +α r− 2
2
+ B0 e−γ ri,Gbest (xGbest − xi )
Initialization Process Applying the Median filter on the image
Initialization Firefly algorithm parameters
Randomly dispatch insects on the image
Evaluating each insect's light intensity, that is obtained according to the location of each insect in the pixel
(8)
Indeed, from Eq. (8), the resultant between the compared particle and the total optimum particle is obtained. In Eq. (8), the Gbest index is the pixel index that has most probability value to be selecting than the other pixels, and XGbest is the location of the best particle. By changing the trend of particle movements, the group movement of particles is enhanced, and the particles can more accurately search the area. Search among fireflies is the manner in which each firefly is compared with each other firefly. To identify the maximum point, if a firefly exhibits a lower lighting rate than the compared firefly, the said firefly moves toward the lighter firefly, and if there is a particle with better lighting in the next iteration of the algorithm, the particles again move toward the particle with the better lighting. The deficiency of this method is that the movement of the particles is determined by the rate of firefly’s lighting, which is the optimum location, and the total optimum has no effect on the algorithm search. Due to this deficiency, the entire area of the problem is not optimally searched, and a greater number of iterations are required to achieve the optimum point. This deficiency is resolved during the fireflies’ movements by improving the act of search in which the other fireflies participate. However, to improve this algorithm, it can also be shown that β0, which indicates the rate of light source absorption as used in Eq. (5), presents the same value for both the local optimum and the total optimum. In retinal images, existing high intensified radiant places which are not considered as optic disc disturb firefly algorithm in detecting the optic disc. In this study, we changed the values of β0 according to the transmitted light from the fireflies and different situations. Also, the corresponding values of each firefly are the difference between transmitted light (pixel intensity value) in previous iteration and the highest intensity rate of each firefly. In this manner, the value of β0 for each firefly in every step is equal to the value of each firefly divided by the average value of that firefly. With this method, we can obtain better values for the functions.
Comparing the light intensity of each insect
The movement of all insects towards the pixel with the lightest intensity
Updating the situation of insects in image pixels Updating the best insects Updating the randomization value of equation
Whether N iteration takes place?
No
Yes All insects are attracted to the optic disc Fig. 3 A summary of the implementation of the proposed firefly algorithm approach for identifying the optic disc in retina
2.7 Summary of optic disc detection in retinal images with the firefly algorithm A summary of the optic disc detection in the retinal images based on the firefly algorithm is presented in Fig. 3. 2.8 Dataset The DRIVE dataset contains 40 colorful retinal images. There are some symptoms of deficiency in seven of these images. These images are taken with a Canon CR5
13
nonmydriatic 3CCD camera at an angle of 40° from which the image of the eye is a circle with the diameter of 540 pixels. The size of the images in this dataset is 584 × 565, and the images are in JPEG format. The STARE dataset includes 20 colorful retinal images. Some symptoms of deficiency are present in 10 of these images. These images were taken using the Topcon TRV_50 fundus camera at an angle of 35°. The size of the images in this dataset is 605 × 700, and the images are in TIFF format. The DiaRetDB1 dataset consists of 89 color fundus images of which 84 contain at least mild nonproliferative signs of diabetic retinopathy, and five are considered normal and do not contain any signs of diabetic retinopathy according to all of the experts who participated in the evaluation of these images. The images were captured using the same 50° field-of-view digital fundus camera with different imaging settings. The size of the images in this dataset is 1152 × 1500, and the images are in PNG format. The proposed method for analyzing the dataset images was implemented using the MATLAB 2013a programming language and run on a PC with an Intel Duo CPU, 2.00 GHz RAM machine.
Med Biol Eng Comput
algorithm for the DRIVE, STARE, and DiaRetDB1 datasets are shown in Table 2. Notably, the elapsed times obtained with the proposed method were achieved without any alterations to the sizes of the input images. To achieve a performance criterion for this paper, the criterion for the correct localization of the optic disc was defined based on the evaluations of the probable optic disc centers compared to the manually selected centers. If the difference between the center of the optic disc, which is specified manually, and acquired position with firefly algorithm is less than the optic disc radius, then it can be said that we have detected the place of the optic disc accurately. The proposed optic disc localization method achieved a success rate of 95 % in the STARE dataset (i.e., the optic disc was correctly identified in 19 of the 20 images). The average distance between the probable optic disc center and the hand-selected center was 20.93. Moreover, the optic disc was successfully identified in all of the 40 DRIVE dataset images with the proposed method. The average distance between the estimated optic disc center and the hand-selected center was 9.5. The distances between the true locations and the estimated locations are illustrated in Fig. 7. Moreover, the success rates of this method are compared with those of other localization methods in Table 3.
3 Results Experiments were conducted to demonstrate the performance of the proposed approach. We tested the proposed method on three publicly available datasets, the DRIVE dataset [26], the STARE dataset [9], and the DiaRetDB1 dataset [11]. The results are shown in Figs. 4, 5, and 6. In the first step, the proposed method parameters were experimentally determined, and the results are shown in Table 1. To evaluate and quantify the performance of the proposed method, the optic disc centers and diameters were determined manually for all images. The estimated optic disc center was deemed acceptable if it was located within the circular optic disc area, i.e., if the distance between the estimated location and the hand-segmented location was smaller than the half of the hand determined diameter. Figures 4, 5, and 6 show the results of this process as applied to the DRIVE, DiaRetDB1, and STARE dataset images. The optic disc location estimations are represented by crosses (×). We used quantitative results for all the images of the datasets, and this is shown in Fig. 7, where each point represents the distance from the true location and the estimated location. From these figures, it can be observed that the proposed method could be an effective tool for roughly detecting the optic disc in retinal images. The times required to find the optic disc using the firefly
13
4 Discussion In this paper, the firefly algorithm was successfully applied to the detection of the optic discs in retinal images. From Figs. 4, 5, and 6, it can be seen that our method overcame the main problems that characterize retinal images obtained from screening programs. The algorithm was applied to pathological images and, in most cases, registered those pathologies and produced good results. Essentially, this algorithm can be affected by factors determined during the acquisition of the image that make optical disc detection more difficult, such as poor image contrast, the existence of high contrast lesions, and brighter effects. Exudate lesions, artifacts, and retinal sheen can be very bright and might lead to false detection of the optic disc (OD). We used median filters to resolve this problem. In the initialization process, the insects are dispatched randomly over the image. If this random dispatch does not place any insect inside the true OD region, the location is not detected. However, across all of the simulations (i.e., in all of the tests), we did not encounter this problem, as at least one insect was always placed inside the optic disc. The initial emission intensities of the insects were the intensities of the pixels that they had been placed on and did not vary with image size.
Med Biol Eng Comput
Fig. 4 Results of applying the proposed method to 20 DRIVE dataset images. The estimated localization of the optic disc is marked with a cross (×)
Fig. 5 Results of applying the proposed method to 10 selected DiaRetDB1 dataset images. The estimated localization of the optic disc is marked with a cross (×)
13
Med Biol Eng Comput
Fig. 6 Results of applying the proposed method to five selected STARE dataset images. The estimated localization of the optic disc is marked with a cross (×)
Fig. 7 Quantitative results for all the images of the datasets used, where each point represents the distance from the true location and the estimated location Table 2 Computational timing comparison Dataset
Number of test images
Optic disc detection time (s)
DRIVE STARE
40 20
2.13 2.81
DiaRetDB1
89
3.52
Table 3 Success rates of optic disc localization methods
13
With the proposed approach, detecting the optical disc is easy. Although detecting the optic disc completely is useful in situations such as glaucoma diagnosis, this was not the reason of this study. The main purpose of the present study was to develop and implement a natural inspired algorithm in detecting the place of optic disc in human retinal images. In this application, the proposed method achieved success rates of 94.38, 95, and 100 % in the DiaRetDB1, STARE, and DRIVE datasets, respectively. The results achieved by the application of the proposed method to the STARE and DiaRetDB1 datasets are as good as or better than those described in previous works [10, 20, 21, 24, 28]. Due to the importance of using real images, we intend to develop an authoritative system and clinical applications. In this regard, our use of the latter two datasets can be considered an advantage of the present work compared to other methods that can be found in the literature. Our proposed approach can be considered an essential step toward the development of computer-aided diagnostic (CAD) systems based on the application of the firefly algorithm to normal screening programs that seek to diagnose diabetic retinopathy. Indeed, based on the obtained experimental results, the firefly algorithm efficiently located the optic disc. Additionally, the parallel implementation of the firefly algorithm should be exploited to decrease the computational load observed in this study.
Localization methods
DRIVE dataset (%)
STARE dataset (%)
DiaRetDB1 (%)
Sinthanayothin et al. [24] Walter et al. [28] Rashid Qureshi et al. [21] Hung-Kuei Hsiao et al. [10] Carla Pereira et al. [20]
60 80 100 100 100
50 75 – 90 –
– – 94.02 – 93.25
Proposed method
100
95
94.38
Med Biol Eng Comput
5 Conclusion In this paper, we used the firefly algorithm to identify the location of the optic disc in human retinal images. With this method, some insects are initially randomly applied to the retina image, and each pixel in which an insect is placed is considered as a firefly lighting. Then, the insects are compared two by two, and insects that are less attractive move toward more attractive insects. Finally, a single insect is selected as the most attractive, and this insect represents the optimum response to the problem in question. Two important issues involved in this method are changes in the light intensities of the image pixels and the formulation of the insect’s attractions. The light intensity is proportional to the insect’s fitness. Generally, attraction is a relative parameter and is estimated by the other insects. Moreover, attraction depends on the distance between the insects. Because the highest rates of lighting in the retina are related to the optic disc, all of the insects move to this location, and the location of the optic disc in the image can thus be identified.
References 1. Abdel-Haleim A, Abdel-Razik Y, Ghalwash AZ, Sabry AA, Abdel-Rahman G (2008) Optic disc detection from normalized digital fundus images by means of a vessels’ direction matched filter. IEEE Trans Med Imaging 27:11–18 2. Akita K, Kuga H (1982) A computer method of understanding ocular fundus images. Pattern Recognit 15:431–443 3. Cassel GH, Billig MD, Randall HG (2001) The eye book: a complete guide to eye disorders and health. Johns Hopkins University Press, Baltimore 4. Chaichana T, Yoowattana S, Sun Z, Tangjitkusolmun S, Sookpotharom S, Sangworasil M (2008) Edge detection of the optic disc in retinal images based on identification of a round shape. Communications and information technologies, international symposium, pp 670–674 5. Cox MJ, Wood ICJ (1991) Computer-assisted optic nerve head assessment. Ophthalmic Physiol Opt 11:27–35 6. Fleming AD, Goatman KA, Philip S, Olson JA, Sharp PF (2007) Automatic detection of retinal anatomy to assist diabetic retinopathy screening. Phys Med Biol 52:331–345 7. Gonzales R, Woods C, Eddins RE (2004) Digital image processing. Prentice-Hall, Inc 8. Hoover A, Goldbaum M (2003) Locating the optic nerve in retinal image using the fuzzy convergence of the blood vessels. IEEE Trans Med Imaging 22:951–958 9. Hoover A, Kouznetsova V, Goldbaum M (2000) Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response. IEEE Trans Med Imaging 19(3):203–210 10. Hsiao H-K, Liu C-C, Yu C-Y, Kuo S-W, Yu S-S (2012) A novel optic disc detection scheme on retinal images. Expert Syst Appl 39(12):10600–10606 11. Kauppi T, Kalesnykiene V, Kamarainen J-K, Lensu L, Sorri I, Raninen A, Voutilainen R, Uusitalo H, Kälviäinen H, Pietilä J (2006) DIARETDB1 diabetic retinopathy database and evaluation protocol, Technical report 12. Kong HJ, Kim SK, Seo JM, Park KH, Chung H, Park KS (2004) Three dimensional reconstruction of conventional stereo optic
disc image. Annual international conference of the IEEE EMBS, Vol 12. San Francisco, pp 29–32 13. Lalonde M, Beaulieu M, Gagnon L (2001) Fast and robust optic disc detection using pyramidal decomposition and Hausdorffbased template matching. IEEE Trans Med Imaging 20:1193–1200 14. Li H, Chutatape O (2001) Automatic location of optic disc in retinal images. IEEE ICIP, Thessaloniki, pp 837–840 15. Lowell J, Hunter A, Steel D, Basu A, Ryder R, Fletcher E, Kennedy L (2004) Optic nerve head segmentation. IEEE Trans Med Imaging 23:256–264 16. Lupascu CA, Tegolo D, Rosa LD (2008) Automated detection of optic disc location in retinal images. 21st IEEE international symposium on computer-based medical systems, Finland, pp 17–22 17. Morales S, Naranjo V, Perez D, Navea A, Alcaniz M (2012) Automatic detection of optic disc based on PCA and stochastic watershed. In: Signal processing conference (EUSIPCO), proceedings of the 20th European, Bucharest, pp 2605–2609 18. Osareh A, Mirmehdi M, Thomas B, Markham, R (2002) Colour morphology and snakes for optic disc localization. The 6th medical image understanding and analysis conference, Vol 1, pp 21–24 19. Patton N, Aslam TM, MacGillivray T, Deary IJ, Dhillon B, Eikelboom RH, Yogesan K, Constable IJ (2006) Retinal image analysis: concepts, applications and potential. Prog Retin Eye Res 25:99–127 20. Pereira C, Gonçalves L, Ferreira M (2013) Optic disc detection in color fundus images using ant colony optimization. Med Biol Eng Comput 51:295–303 21. Qureshi RJ, Kovacs L, Harangi B, Nagy B, Peto T, Hajdu A (2012) Combining algorithms for automatic detection of optic disc and macula in fundus images. Comput Vis Image Underst 116:138–145 22. Reza AW, Eswaran C, Hati S (2008) Automatic tracing of optic disc and exudates from color fundus images using fixed and variable thresholds. J Med Syst 33:73–80 23. Sekhar S, Al-Nuaimy W, Nandi AK (2008) Automated localization of retinal optic disc using Hough transform. The 5th IEEE international symposium on biomedical imaging: from nano to macro, Paris, pp 77–80 24. Sinthanayothin C, Boyce JF, Cook HL, Williamson TH (1999) Automated localization of the optic disc, fovea, and retinal blood vessels from digital color fundus images. Br J Ophthalmol 83:902–910 25. Sopharak A, Uyyanonvara B, Barman S, Williamson TH (2009) Automatic detection of diabetic retinopathy exudates from nondilated retinal images using mathematical morphology methods. Comput Med Imaging Graph 32:720–727 26. Staal J, Abàmoff MD, Niemeijer M, Viergever MA, van Ginneken B (2004) Ridgebased vessel segmentation in color images of the retina. IEEE Trans Med Imaging 23(4):501–509 27. Tobin KW, Chaum E, Govindasamy VP, Karnowski T (2007) Detection of anatomic structures in human retinal imagery. IEEE Trans Med Imaging 26:1729–1739 28. Walter T, Klein JC, Massin P, Erginary A (2002) A contribution of image processing to the diagnosis of diabetic retinopathydetection of exudates in color fundus images of human retina. IEEE Trans Med Imaging 21:1236–1243 29. Welfer D, Scharcanski J, Kitamura C, Pizzol MD, Ludwig L, Marinho D (2010) Segmentation of the optic disc in color eye fundus images using an adaptive morphological approach. Comput Biol Med 40:124–137 30. Yang XS (2008) Nature-inspired metaheuristic algorithms. Luniver Press, Frome 31. Yavuz Z, ˙Ikibas¸ C, S¸evik U, Köse C (2009) A method for automatic optic disc extraction in retinal fundus images. 5th International advanced technologies symposium, Karabuk, pp 93–98
13
