An efficient multistage segmentation method for accurate hard exudates and lesion detection in digital retinal images

Abstract

Digital retinal images are commonly used for hard exudates and lesion detection. An efficient segmentation method is needed to detect and discern the lesions from the retinal area. In this paper, a hybrid method is presented for digital retinal image processing for diagnosis and screening purposes. The goal of this research is to suggest a supervised/semi supervised approach for exudates detection in fundus images and it is also to investigate a technique to find the optimum structure. The image is first transformed into fuzzy domain after an initialization. A cellular learning automata model is used to detect any abnormality on the image which is related to a lesion. The automaton is created with an extra term as the rule updating term to increase the flexibility and capability of the cellular automata. The selection and updating of rule are implemented automatically We also performed allocating the score and penalty value for the cells toward the process of segmentation Three main statistical criteria are introduced as the sensitivity, specificity and accuracy. A number of 50 retinal images with visually detection hard exudates and lesions are the experimental dataset for evaluation and validation of the method. For STARE retina image dataset, for a neighborhood of 5 × 5, score of ϑ = 0.01, penalty of ξ = 0.01, ratio of state overall variation in three sequential cycles in cellular automata $\bar{η} = 0.5$ , updating additive value σ = 0.02 & rule selection threshold value ρ = 0.8 the mean value of statistical criteria averaged over all dataset can reach 99% which is an outstanding assessment result for the proposed method.

Keywords

Digital retinal images hard exudates and lesions detection fuzzy theory cellular learning automata statistical evaluation

1 Introduction

Two main interesting issues for computer aided diagnosis in ophthalmology are as the automatic detection of blood vessels and detection of the optic disc. The shortcoming of detecting and counting lesions in the human retina like micro aneurysms and exudates are considered as the consuming and high possible to human error therefore a lot of works have been done on order to solve the problem of lesion detection automatically [1]. The most prevalent reason of visual deterioration and the modern overwhelming blindness among workers of industrialized countries is Diabetic retinopathy (DR) which is a visual complication of diabetes [2, 3]. Due to clinical importance and epidemical problem, the researchers had a lot of efforts to improve the methods for diagnosis and medical medication by developing and implementing algorithms to perform retinal image analysis [4], fundus image enhancement [5] and monitoring [6]. Designing automatic image analysis algorithms had special significance to detect hard exudates (HEs) (7). HEs represent the best distinguished markers for the being of retinal oedema which is the main problem of visual loss in non-proliferation forms of DR [2]. Moreover, during early stages of DR, HEs are one of the widespread lesions during initial stages of DR [2].

Many techniques have been presented for HE detection in fundus images considering a variety of methods [7], the methods includes the usage of image contrast and brightness analysis [8 –15], Bayesian classifier [16, 17], neural network [18, 19]. In order to segment a candidate bright exudates area, system was developed by Zhang et al. [20] that used local contrast enhancement, Fuzzy C-mean clustering in LUV color space, but the major problem with FCMC is to estimate the number of clusters to use. In order to detect contours typical of exudates, a new mathematical reconstruction was applied by Walter et al. [21]. The results showed the productivity values of 92.4% and sensitivity value of 92.8% with a set of 15 abnormal retinal images. But this method couldn’t distinguish between exudates from cotton wool spots. Gardner et al. [22] used the back propagation neural network in order to perform the segmentation of exudates. If we compare the result of Gardener et al. [22] method with similar expert ophthalmologist, they achieved a sensitivity of 88.4% and specificity of 83.5%. In order to detect the exudates, but it had a drawback which it couldn’t get good result on low quality images.

Many of past techniques have used retinal features which were quite obvious on acquired imagery. To acquire high quality and apparent visible retinal features is time consuming and very difficult and troublesome for patient. Also sometimes there are a number of exudates with different sizes apart or stuck which requires a model based method for detection. Therefore, in this paper the cellular learning automata in fuzzy domain is used for automated and accurate detection of exudates on retinal images. The final objective is to propose a method and system to automatically detecting of exudates, to provide decision support and to reduce workloads for expert ophthalmologists’ by a supervised method which can also act as automation system.

Cellular automata (CA) consist of a regular grid of cells, each of which can be in only one of a finite number of possible states. The previous states of a region of surrounding neighborhood of cells will determine the state of a cell. The state of a cell will be updated in discrete time steps synchronously. The rule which identically contained in each cell is basically a finite state machine It normally forms a rule table with an input for every possible neighborhood configuration of states. Cellular automata are discrete dynamical systems. These are very helpful for implementing and analyzing the phenomena ordering, turbulence, chaos, symmetry breaking, etc. and a lot of applications have been designed in modeling systems in some domains such as biology sociology, Physics Though only a few simple rules are attributed to each cell, one of the main usage of cellular automata is that the combination of matrix of cells with their local interaction leads to more sophisticated emergent global behavior. That is, although each cell has an extremely limited view of the system (just its immediate neighbors), localized information is propagated at each time step, enabling more global characteristics of the overall CA system [27].

In this paper, fuzzified cellular learning automata is introduced for segmentation of digital retinal images. A retinal image dataset consisting 50 retinal images with hard exudates are used to assess the proposed method. The paper is organized as follows: Section 2 presents the dataset and the methodology. Results are given in Section 3 and in Section 4 the results are discussed. Finally, some conclusions will be shown in Section 5.

2 Materials and method

2.1 Image database

Subset of STARE project’s dataset was used in order to obtain main dataset [23‖26]. The subset consists of 81 retinal images for assessing the proposed exudates detection method. The captured images were taken using a TopCon TRV-50 fundus camera, FOV(field of View) = 35° and then it was digitized at 605×700, 24-bits pixel The dataset is used for both the evaluation of the proposed method and also comparison to some other methods. The dataset consists of a total of 50 color retinal images which are taken without pupil dilation with a KOWA-7 non-mydriatic retinal camera with a 45° FOV. The image size is 768×576 pixels at 24 bits per pixel in RGB format. All retinal images analyses are performed on a Core2 Duo 2.2 GHz Laptop using MATLAB 7.6.0 for all implementations.

2.2 Transformation of retinal images into fuzzy domain

Let supposed to have an image of size M × N with L gray levels g ranging from 0 to L - 1. The image γ can be viewed as an array of fuzzy singletons [19]. Each element of the array is the membership value μ_{γ
ⁱ} (g_xy) of the gray level g_xy, corresponding to the (x, y) th pixel, regarding to an image property such as brightness, edginess, homogeneity, etc. Using the fuzzy sets notation image γ can be represented as: $\begin{matrix} γ^{i} & = & {\frac{μ_{γ^{i}} (g_{xy})}{g_{xy}} | i = 0, 1, \dots, M - 1, \\ j = 0, 1, \dots, N - 1} \end{matrix}$ (1)

To enhance the gray levels across the edges and to make better the contrast of image, gray levels of the pixels are normalized to unit interval [0, 1] corresponding to the whole gray level range [0, 255]. The primary motion matrix is obtained by calculation of Ω function over $\overset{⌣}{δ} \times \overset{⌣}{δ}$ neighborhood windows around all pixels as $Ω^{i} (x, y) = {\begin{matrix} \nabla^{i, i - 1} \times \nabla^{i, ϑ - 1} & if \frac{1}{\overset{ˇ}{δ}} \sum | γ^{i} (x - 1 : x + 1, y - 1 : y + 1) - γ^{i - 1} (x - 1 : x + 1, y - 1 : y + 1) | > T_{1} \\ 0 & otherwise, \end{matrix}$ (2) where $\nabla^{i, i - 1} = \frac{1}{{\overset{ˇ}{δ}}^{2}} \sum | γ^{i} (x - 1 : x + 1, y - 1 : y + 1) - γ^{i - 1} (x - 1 : x + 1, y - 1 : y + 1) |$ (3) and $\nabla^{i, ϑ^{i}} = \frac{1}{{\overset{ˇ}{δ}}^{2}} \sum | γ^{i} (x - 1 : x + 1, y - 1 : y + 1) - \frac{1}{i - 1} \sum_{j = 1 : i - 1} γ^{j} (x - 1 : x + 1, y - 1 : y + 1) |$ (4)

Matrix Ωⁱ is obtained as the primary motion detected in ith image of the image set Γ. Parameter T₁ is the threshold value for motion detection in corresponding $\overset{ˇ}{δ} \times \overset{ˇ}{δ}$ neighborhood windows in image sequence ith and (i + 1) ^th. Coefficient ∇^i,i-1 means the first order absolute gradient of the sequent images γⁱ and γⁱ - 1 and ∇^i,ϑⁱ depicts the first order gradient of the image relative to the average of the previous images γ¹ : γ^i-1. Multiplication of the two terms in Equation (2) guarantees both of these statements: (a) difference with the previous image implies motion and (b) slight differences with the average of images represents vibration and undesired motions (e.g. shadows, camera shaking or local illumination variations).

The matrix Ωⁱ is then undergone normalization in the [minimum, maximum] range as $\begin{matrix} Ω^{i} \to Ω_{-}^{i} \\ = (Ω^{i} - min (Ω^{i})) \times \frac{L - 1}{max (Ω^{i}) - min (Ω^{i})} \end{matrix}$ (5)

Therefore matrix Ωⁱ is normalized to the range [0, 255]. Equation (5) represents also a defuzzification formula though some elements of the matrix are out of the unit interval. Now, Cellular Learning Automata is used to detect the object of interest in the primary motion matrix Ωⁱ which has been converted to gray scale image format.

2.3 Retinal image fuzzy automaton

A retinal image fuzzy automaton in the Fuzzy Cellular Learning Automata is created as $A = (States; Neighborhood; Rules;$ (6) $\begin{matrix} Parameters; Update) \\ S : States as {S_{i}^{t}} in the range S_{i}^{t} \in [min, max], \\ S_{i}^{t} : ith State at time t where : \\ S_{i}^{t} : {μ_{i}^{t} \in [0, 1] | \exists a one - to \\ - one function F (.) ∴ \forall μ_{i}^{t} \to \exists F (μ_{i}^{t}) \\ \in [min (S), max (S)]} \\ and \\ S & = & {S_{:}^{t} | t = 1 : t} \\ N : Neighborhood as δ \times δ Extended Moore \\ Neighborhoods : \\ N^{δ} (x, y) : \\ {(x - [\frac{δ}{2}] : x + [\frac{δ}{2}], y - [\frac{δ}{2}] : y + [\frac{δ}{2}])} \\ where δ \in {2 k + 1 | k \in ℕ}, \\ R : Rules as \\ R : {p_{i} \to q_{i} | p_{i} & q_{i} are logical terms \\ and i = 1 : N_{R}} \\ which are once defined by the user accorsing \\ to a set of training samples . \\ P & = & (ϑ, ξ, T) : Paramerters as ϑ : score, \\ ξ : penalty and T : similarity threshold \\ where T = {ρ^{i}}_{i = 1 : N_{R} .} \\ U : Updating functions for learning rules \\ U : Ξ (R \to \hat{R}) \end{matrix}$

States in the automaton corresponds to decision variables of the problem. The rules are set as transition functions: $\begin{matrix} R : ψ (S^{t} (x, y) \to S^{t + 1} (\hat{x}, \hat{y})) \\ = {\begin{matrix} S^{t + 1} (\hat{x}, \hat{y}) = S^{t} (x, y) + ϑ & if \exists λ^{i} \in Λ \to λ^{i} ≎ Ω_{-}^{i} (N^{δ} (x, y)) \\ S^{t + 1} (\hat{x}, \hat{y}) = S^{t} (x, y) - ξ & otherwise \end{matrix} \end{matrix}$ (7) where ψ (.) corresponds to the function of transition, S^t (x, y) is the current state and $S^{t + 1} (\hat{x}, \hat{y})$ is the next state, ϑ and ξ are the score and penalty values, respectively, Λ = {λⁱ} _{i=1:N_Λ} is the set of rules and the term $λ^{i} ≎ Ω_{-}^{i} (.)$ implies the conditional rule applying. The condition under which the appropriate rule is selected and then applied can be logically expressed as: $λ^{i} ≎ Ω_{-}^{i} = {\begin{matrix} 1 & if \frac{1}{δ^{2}} \times \sum \frac{| λ^{i} - Ω_{-}^{i} |}{L - 1} > ρ^{i} \\ 0 & otherwise \end{matrix}$ (8)

Table 1

The abbreviations which are used in formula (20), (21) and (22)

Abbreviation	Stands for	Description
TP	True positive	Pixels correctly detected as object
FP	False positive	Pixels incorrectly marked as object
TN	True negative	Pixels correctly marked as background
FN	False negative	Pixels incorrectly labeled as background

Where ρⁱ is the accordance factor for the ith rule, state rules for this equation at above the conditional term can be considered as similar to process of searching within a look up table. Threshold ρⁱ selection directly specifies the sensitivity to the rules selection and to the state changes. Very low ρⁱ values lead to significant score assignments and it may result to enlarging the state values and also vice versa. Therefore, this parameter can be justified in each time interval of the cellular automata and this updating process leads to a learning scheme. Threshold value ρⁱ is updated in each cycle according to this formula: $U : ρ^{i} = {\begin{matrix} ρ^{i} + σ & if η > \bar{η} \\ ρ^{i} - σ & otherwise \end{matrix}$ (9) where σ is the updating additive value ranged between 0.001 and 0.1. Parameter η is the ratio of state overall variation in three sequent cycles in cellular automata and $\bar{η}$ is a threshold value for η. For simplicity $\bar{η}$ can be initially set as 0.6 and η is randomly selected in the unit interval [0, 1]. Equation (3) makes the threshold value ρⁱ larger if the rule selection sensitivity and then the score assignment raises significantly since larger ρⁱ values limit the score assignments (and vice versa). The evolution cycle of the Cellular Learning Automata goes forward according to the learning rule. Score and penalty parameters can also be updated in each cycle but they are preferred to be set as low as possible initially and the convergence to be reached in a longer time. Applying appropriate learning functions to the rules will guarantee the convergence and goal tracing capability of the model.

2.4 Segmentation of retinal images

For lesion detection in retina images, a binary image comprising two segments of healthy area of retina as background and the detected lesion(s) as foreground is desired. Therefore, since after the cycle t = τ, the resulting image is in the fuzzy domain and its pixels are in unit interval, a non-sensitive threshold ς ranged between 0.1 and 0.9 is selected to produce is binary image I_b as $\begin{matrix} for all (\hat{x}, \hat{y}) \in S^{t} (.) \\ if S^{t} (\hat{x}, \hat{y}) ⩾ ς \to I_{b} (\hat{x}, \hat{y}) = 1 \\ esle I_{b} (\hat{x}, \hat{y}) = 0 \\ endif \\ endfor \end{matrix}$

2.5 Segmentation quantitative criteria

To report the performance of the classifier, sensitivity, specificity and accuracy as three common criteria are reported and calculated as follow: $Sensitivity (%) = \frac{TP}{TP + FN} \times 100$ (10) $Specificity (%) = \frac{TN}{TN + FP} \times 100$ (11) $\begin{matrix} Accuracy (%) & = & \frac{TP + TN}{TP + FP + FN + TP} \times 100 \end{matrix}$ (12) where the abbreviation and their meaning are summarized in Table 1. The sensitivity results describe the percentage of correctly detected object’s pixels. Specificity result demonstrates the percentage of incorrect assignments of the pixels and the accuracy shows the validity of the assignments. The higher accuracy percentage reveals the higher veracity of results.

3 Simulation results

A number of 50 retinal images with hard exudates and lesion detected visually are considered as the dataset to evaluate the segmentation method which was proposed. Implementations of the proposed method were done in MATLAB software suing image processing toolbox. All codes were developed in MATLAB M-files while a set of the images are copied in the same directory as the M-file. The program consists of four main parts as the denoising step, initialization part, segmentation method and image binarization and assessment part.

In our experiments, we first evaluate our methods to find the optimum parameters values and the most suitable structure. Then, the results of the retina exudates detection for the proposed method are compared to the results of some other methods in recent literature. All detections are executed for lesion based images with no limitations in the number and regions of the lesions. If a lesion overlaps at least in part with the ground truth then it be called true positive (TP), if in the automatic segmentation there can be not found any corresponding lesion then it be called a False Negative (FN), If here be an exudate in automatic segmentation but there is no corresponding lesion has been manually segmented and True Negative is the true assignment to the non-lesion regions. In the first part of the experiments, the effects of score and penalty values for two iteration values are studied. All runs have been done for two neighborhood sizes of 3×3 and 5×5. Samples of the original retinal images with hard exudates are shown in Fig. 1. Figure 2 shows the fuzzifed images for which all pixels’ grayscale are normalizedto (0,255).

Fig. 1

Original Retinal Images with hard exudates.

Fig. 2

Fuzzified Retinal Images normalized in (0,255) grayscale.

Segmented images of retinal images are shown in Fig. 3. The images in Fig. 3 correspond to the implementation of the proposed method with a marginally optimum set of parameters while their numerical results are given in Table 2. The results of statistical analysis on score and penalty assignments are given in Table 2 while the proposed method. It is shown that the optimum score and penalty values stands about 0.01 and 0.02 while the product of score/penalty values and the number of iterations is important. It is also shown that for our dataset images the neighborhood size of 5×5 which is an extended Moore neighborhood leads to better results compared to simple Moore neighborhood. Table 3 presents the results of a brief comparison between the Fuzzy cellular automata and fuzzy cellular learning automata. For this part of the experiment, the rules are updated for two values of $\bar{η}$ & σ. Parameter $\bar{η}$ affects the rate of score assignment and σ determines the threshold for rule selection. Optimum values for $\bar{η}$ & σ are found as $\bar{η} = 0.5$ & σ = 0.02. Different similarity thresholds for rule selection from 0.5 to 0.95 are used to figure out the detected exudates and it has been shown that low threshold values result in enlarged detected regions and vice versa. It happens because low threshold values cause higher rates of rule selection and applying rules relates to the score assignment and exudates configuration. Similarity threshold value ρ ∼ 0.8 - 0.85 equally for all rules has been found optimum.

Fig. 3

Segmented Retinal Images by the FCLA method.

Table 2

Statistical parameters for exudates detection process by the proposed method (τ = 50) averaged over 50 images

		FCLA Method
	δ × δ	ϑ = 0.005	ϑ = 0.01	ϑ = 0.02	ϑ = 0.05
		ξ = 0.005	ξ = 0.01	ξ = 0.02	ξ = 0.05
Sensitivity	3×3	∼88%	∼92%	∼96%	∼92%
	5×5	∼91%	∼96%	∼97%	∼93%
Specificity	3×3	∼96%	∼97%	∼99%	∼98%
	5×5	∼98%	∼99%	∼99%	∼98%
Accuracy	3×3	∼89%	∼94%	∼96%	∼95%
	5×5	∼92%	∼95%	∼97%	∼92%
Average	3×3	∼91%	∼94.3%	∼97%	∼95%
	5×5	∼93.7%	∼96.7%	∼97.7%	∼94.3%

Table 3

Statistical parameters for exudates detection process by the proposed FCLA method for different rule updating parameters when N^δ = 5 ×5, ϑ = 0.01, ξ = 0.01 & τ = 30

		Updating rules with default ρ = 0.8
	Fixed	$\bar{η} = 0.5 &$	$\bar{η} = 0.6 &$	$\bar{η} = 0.4 &$	$\bar{η} = 0.6 &$	$\bar{η} = 0.4 &$
	rules	σ = 0.02	σ = 0.01	σ = 0.01	σ = 0.05	σ = 0.05
Sensitivity	∼83%	∼88%	∼88%	∼84%	∼90%	∼83%
Specificity	∼92%	∼97%	∼91%	∼93%	∼87%	∼96%
Accuracy	∼84%	∼89%	∼83%	∼87%	∼81%	∼81%
Average	∼86.3%	∼91.3%	∼87.3%	∼88%	∼86%	∼86.7%

The final state of the FCLA iterative method gives out a retina image with segmented exudates. Based on the number of iterations, defined rules and the values of score and penalty the final state maybe an image in fuzzy domain its pixels ranged in the unit interval. Hence, a threshold parameter for image binarization ς is proposed to assign the pixels of the images either 0 or 1 according to their fuzzy value. If appropriate rules are defined and suitable values for ϑ, ξ & τ are set, the sensitivity to parameter ς is too low. Then the values in the range [0.3, 0.9] will lead to slightly different binary images. The differences between the statistical parameters for three different values of ς are demonstrated in Table 4.

Table 4

Statistical parameters for exudates detection process by the proposed FCLA method for different threshold values of image binarization when N^δ = 5 ×5, ϑ = 0.01, ξ = 0.01, $\bar{η} = 0.5$ , σ = 0.02 & ρ = 0.8

	τ = 30			τ = 60			τ = 90
	ς = 0.3	ς = 0.6	ς = 0.9	ς = 0.3	ς = 0.6	ς = 0.9	ς = 0.3	ς = 0.6	ς = 0.9
Sensitivity	∼85%	∼88%	∼87%	∼87%	∼91%	∼91%	∼95%	∼99%	∼97%
Specificity	∼96%	∼97%	∼94%	∼98%	∼99%	∼99%	∼99%	∼99.9%	∼99.6%
Accuracy	∼87%	∼91%	∼89%	∼89%	∼95%	∼94%	∼97%	∼99%	∼98%
Average	∼89.3%	∼92%	∼90%	∼91.3%	∼95%	∼94.7%	∼97%	∼99.3%	∼98.2%

A comparison to some other methods is performed in terms of the statistical parameters. To evaluate the efficiency of the proposed FCLA method, the results of the six methods of references [], K-Nearest Neighbor method are listed in Table 5. The results for the proposed method are attained with the optimum parameters found in the previous parts of the experiment. These parameters are N^δ = 5 ×5, ϑ = 0.01, ξ = 0.01, τ = 50, $\bar{η} = 0.5$ , σ = 0.02, ρ = 0.8 & ς = 0.5. It is seen that the proposed method with its optimum parameters can detect the exudates in retina images accurately with high sensitivity and specificity.

Table 5

Statistical parameters for exudates detection process by the proposed FCLA method compared to some other recent methods. For FCLA method N^δ = 5 ×5, ϑ = 0.01, ξ = 0.01, τ = 50, $\bar{η} = 0.5$ , σ = 0.02, ρ = 0.8 & τ = 0.5

	Method	NI/NIEx	Sensitivity	Specificity	Accuracy
Zhang et al. [19]	SVM	213/213	88%	84%	–
Wang et al. [15]	SVM	154/54	100%	71%	–
Osareh [20]	FCMC	300/300	96%	94.6%	–
Garcia et al. [21]	MLP	50/25	84%	62.7%	–
Wisaeng et al. [22]	SVM	120/98	94%	89.52%	–
KNN	KNN	50/20	93%	92%	93%
CLA [28]	CLA	50/20	96.3%	98.7%	96.1%
Proposed method 1	Hybrid	50/20	98.3%	98.1%	97.3%
Proposed method 2	Hybrid	50/20	99.4%	99.6%	99.1%

4 Conclusion

In this research, a hybrid method was presented for digital retinal image processing for diagnosis and screening purposes. A supervised/semi-supervised method for detecting of exudate in fundus images was presented and the optimum structure was found in a comprehensive analysis. The images were first transformed into fuzzy domain. A cellular learning automata model was used to detect any abnormality on the image which is related to a lesion. STARE retina image dataset, for a neighborhood of 5 × 5, score of ϑ = 0.01, penalty of ξ = 0.01, ratio of state overall variation in three sequent cycles in cellular automata $\bar{η} = 0.5$ , updating additive value σ = 0.02 & rule selection threshold value ρ = 0.8 the mean value of statistical criteria averaged over all dataset can reach 99% which is an outstanding assessment result for the proposed method. With considering the advantages and proficiencies of the proposed method, one may be interested in a more rapid algorithm for real time applications where the processing should be accomplished in a short fraction of time. Hence, the authors suggest more study and analysis of the computation time for the proposed method in future works.

References

Walter

and Klein

J.C.

, Segmentation of Color Fundus Images of the Human Retina: Detection of the Optic Disc and the Vascular Tree Using Morphological Techniques LNCS 2199, 2001, pp. 282–287.

García

, Sánchez

C.I.

, Poza

, López

M.I.

and Hornero

, Detection of hard exudates in retinal images using a radial basis function classifier, Annals of Biomedical Engineering 37(7) (2009), 1448–1463.

Singer

D.E.

, Nathan

D.M.

, Fogel

H.A.

and Schachat

A.P.

, Screening for diabetic retinopathy, Ann Intern Med 116(8) (1992), 660–671.

Sánchez

C.I.

, Hornero

, López

M.I.

, Aboy

, Poza

and Abásolo

, A novel automatic image processing algorithm for detection of hard exudates based on retinal image analysis, Medical Engineering & Physics 30(3) (2008), 350–357.

Cree

M.J.

, Olson

J.A.

, McHardy

K.C.

, Sharp

P.F.

and Forrester

J.V.

, The preprocessing of retinal images for the detection of fluorescein leakage, Phys Med Biol 44(1) (1999), 293–308.

Pinz

, Prantl

and Datlinger

, Mapping the human retina, IEEE Trans Med Imaging 17(4) (1998), 606–619.

Patton

, Aslam

T.M.

, MacGillivray

, Deary

I.J.

, Dhillon

and Eikelboom

R.H.

, et al., Retinal image analysis: Concepts, applications and potential, Prog Retin Eye Res 25(1) (2006), 99–127.

Eslami

and Radha

, The contourlet transform for image denoising using cycle spinning, Proceeding of Asilomar Conference on Signals, Systems, and Computers 2003, pp. 1982–1986.

Akita

and Kuga

, A computer method of understanding ocular fundus images, Pattern Recognition 15(6) (1982), 431–443.

10.

Kochner

, Schulman

, Obermaier

, Zahlmann

, Mann

and Englmeier

K.H.

, A image processing system for analyzing color fundus photographs with regard to diabetic retinopathy, Klinische Monatsblatter für Augenheilkunde 211 (1997), 11.

11.

and Chutatape

, Automated feature extraction in color retinal images by a model based approach, IEEE Trans Biomed Eng 51(2) (2004), 246–254.

12.

Philips

, Forrester

and Sharp

, Automated detection and quantification of retinal exudates, Graefe’s Archive for Clinical and Experimental Ophthalmology 231(2) (1993), 90–94.

13.

Walter

, Klein

J.C.

, Massin

and Erginay

, A contribution of image processing to the diagnosis of diabetic retinopathy—detection of exudates in color fundus images of the human retina, IEEE Trans Med Imaging 21(10) (2002), 1236–1243.

14.

Ward

N.P.

, Tomlinson

and Taylor

C.J.

, Image analysis of fundus photographs—the detection and measurement of exudates associated with diabetic retinopathy, Ophthalmology 96 (1989), 80–86.

15.

Zahlmann

, Kochner

, Ugi

, Schulhmann

, Liesenfeld

and Wegner

, et al., Hybrid fuzzy image processing for situation assessment: A knowledge-based system for early detection of diabetic retinopathy, IEEE Eng Med Biol Mag 19(1) (2000), 76–83.

16.

Goh

K.G.

, Hsu

, Lee

M.L.

, Wang

ADRIS: An automatic diabetic retinal image screening system, In: Cios

K.J.

, editor. Medical data mining and knowledge discovery, Springer-Verlag, New York 2000, pp. 181–210.

17.

Wang

, Hsu

, Goh

, Lee

M.L.

, An effective approach to detect lesions in color retinal images, Proceedings of the IEEE Conference on Computer Vision and Pattern Recogniton, 2000, 181–186.

18.

Gardner

, Leverton

, Young

, Lusty

, Dunstan

and Owens

, Automatic detection of diabetic retinopathy using an artificial neural network: A screening tool, Br J Ophthalmol 80(11) (1996), 940–944.

19.

Osareh

, Automated identification of diabetic retinal exudates and the optic disc. Ph.D. thesis. Bristol; 2004.

20.

Zhang

and Chutatape

, Top-down and bottom-up strategies in lesion detection of background diabetic retinopathy, Proceedings of the 20-25 Jun. 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition IEEE Computer Society 2 (2005), 422–428.

21.

Walter

, Klevin

J.C.

and Massin

, A contribution of image processing to the diagnosis of diabetic retinopathy detection of exudate in color fundus images of the human retina, IEEE Trans Med Imaging (2002), 1236–1243.

22.

Gardner

G.G.

, Keating

, Williamson

T.H.

and Elliott

A.T.

, Automatic detection of diabetic retinopathy using an artificial neural network: A screening tool, Br J Ophthalmol 80 (1996), 940–944.

23.

STARE Project Website Clemson University, Clemson, SC [online]. Available: http://www.ces.clemson.edu/~ahoover/stare.

24.

Osareh

, Mirmehdi

, Thomas

and Markham

, Classification and localization of diabetic related eye disease, In Proceeding of the 7th European Conference on Computer Vision 2002, pp. 502–516.

25.

Garcia

, Hormero

, Sanchez

, Lopez

and Diez

, Feature extraction and selection for automatic detection of hard exudate in retinal images, IEEE Conference for Engineering in Medicine and Biology Society, 2007, pp. 4969–4972.

26.

Wisaeng

, Hiransakolwong

and Pothiruk

, Automatic detection of retinal exudates using a support vector machine, Applied Medical Informatics 32(1) (2013), 33–42.

27.

Rosin

P.L.

, Training cellular automata for image processing, IEEE Trans, Image Process 15 (2006), 2076–2087.

28.

Chahkandi Nejada

, Azadbakht

, Adenihvand

, Mohammadi

and Mirzamohammad

, Fuzzy cellular learning automata for lesion detection in retina images, Journal of Intelligent & Fuzzy Systems 27 (2014), 2297–2303.