IDENTIFICATION OF EXUDATES USING FUZZY MATHEMATICAL MORPHOLOGY

Diabetic Retinopathy is the damage to the retina caused by complication and the most common cause of blindness in Thailand. Retinal image is essential for expert ophthalmologists to diagnose diseases. Several of method can achieve good performance on retinal feature are clearly visible. Unfortunately, the color retinal image in Thailand are low-resolution images. The existing method cannot identified low-resolution image. Therefore, this study is part of a larger effort to develop a new method for identification of exudates in low-resolution retinal image. In this study a fuzzy mathematical morphology based on fuzzy logical operator and mathematical morphology method is presented. The color retinal image are segmented by using fuzzy logical operator following key preprocessing step, i.e., color normalization, contrast enhancement, noise removal and color space selection. Afterward, a segmentation using mathematical morphology method was applied in this step. This enables its difference in our methods compared to other approach and the methods can achieve good performance even on low-resolution retinal images. Respect to the experimental results, the results obtained with fuzzy mathematical morphology better than the ones obtained with the fuzzy logical operator only method.


INTRODUCTION
Diabetic retinopathy is a microvascular complication of diabetes, causing abnormalities in the retina. Exudates represent leak from surrounding capillaries and microaneurysms in diabetic retinopathy. The rapid increase of exudates pushes the limits of the exudates screening and semi-automatic retinal image analysis methods provide a potential solution. Unfortunately, the color retinal image in Thailand are low-resolution images. Thus, color retinal image segmentation in low-resolution images is difficult; take a long time and the expensive computational cost has become a bottleneck that limits clinical application. Serveral techniques have been employed to identification of exudates.
The thresholding and region growing technique are widely used. Kavitha and Devi (2005) proposed median filtering and morphology operation for vascular identification. Mul-tilevel threshoding is used to segment bright regions assumed to be the exudates or optic disc. They detec the optic disc as the converging point of vascular and then classifiy the other bright regions as exudates. Li and Chutatape (2003) used region growing and edge detection of exudates. Usher et al. (2004) identified the exudates region by using a region growing followed with edge detection techniques. The application failed to identified lowresulotion image and small exudates. Fuzzy logical operator has also been proposed as a possible solution to identification of exudates problem. Zhang and Chutatape (2005) segmented exudates by using fuzzy in Luv value. Osareh et al. (2001) used fuzzy c-mean clustering to detect digital retinal image and then used neural network train dataset and sparate exudates and non-exudates area by support vector machine. Gardner et al. (1996) used back propagation neural network to automated identification of diabetic retinopathy. Comparative the results by neural network method with of the expert ophthalmologist, the neural network achieved a sensitivity and specificity for the identification of diabetic retinopathy are 88.4 and 83.5%, respectively. The techniques did not work well on low-resolution quality image. All methodologies described have some limitation. The approaches based in identification of exudates by machine learning need a long trianing time and it becomes computationally too heavy. Morever, the simple method for identification of exudates does not enable to detect the exudates boundary.

JCS
From the above studies, identification of exudates can be broken down to two situations: Low-resolution image and multimodal image (Fig. 1). The existing algorithm can achieve good performance when these two situations are not combined together. For example, it is difficult to identification of exudates an images taken at different times and different fields of view. To solve this problem, we proposed a new method for identification of exudates using fuzzy mathematical morphology method to segment and identification of exudaes in low-resolusion and multimodal image.

Patients
The database consists of 2,084 retinal images, of which 968 at signs of exudates and 1,116 images do not contain any signs of exudates according to ophthalmologists participated in the annotation. The retinal images were taken by the local hospital with single degree 40 degree field of view. The image resolution is 700×500 pixels at 24 bits per pixel. All color retinal images are tested on a Core2 Duo 1.60 GHz using MATLAB for algorithms. The overall procedure identification of exudates demonstrated in Fig. 2.

Data Preparation
In source image, the retinal image is created by emitting a flash lamp produced light into patient's eye through complex retina camera optics and capturing the reflected light to the camera sensor. One of the main problem for identification of exudates is the wide variability in the color of retinal image from different patients and different time. Thus, the color of exudates in some region of an images may appear dimmer than the background color of other regions. To solving these problem, we put retinal image through four preprocessing steps before commencing the identification of exudates. In the first step is to normalize the color of the source images across the database. We used histogram matching applied to each individual RGB values to histogram matches that of another image in database. A common application of this is to match the images from two sensors with slightly different responses, or from a sensor whose response changes over time. The following is an example using a greyscale image. The first image is the reference image and the histogram is given. The second image is the one whose histogram will be matched to the first. The third shows the resulting matched image ( Fig. 3A-F). Fig. 3E, although this histogram matching is suitable for retinal image. However, the contrast decreases with the more distance of pixels from the center, especially in the periphery. The number of pixels in these areas may have negligible influence on the the computation of a global transformation whose shape does not necessarily guarantee the desired local enchancement (Giordano et al., 2006). The solution is to devise the center of the neighborhood region is then moved to an adjacent pixel location and the procedure is repeated. The second pre-processing step, we apply local contrast enchancement (Chang and Wu, 1998) to a transformation of the values inside small windows in the image in a way that all values are distributed around the mean and show all possible intensities. The techniques of local contrast enchancement are described below.

Considered from
Given each f in the initial image and a small M×M running window w, then the image is filtered to produce the new image pixel f(i, j) Equation (1): where the ψ is defined as Equation (2): The f max and f min are referred as the whole image maximum and minimum intensity values; while <f> w and σ w indicate the local window mean and standard deviation within each window provides considerable contrast enhancement, which are defined as Equation (3): where, (k, l) represents the location of each pixel within window w. The size of window M should be chosen to be large enough to contain a statistically representative distribution of the local variations of pixels. On the other hand, it must be small enough to not be influenced by the gradual variation of the contrast between the retinal image center and the periphery (Osareh et al., 2003). Here, the window size was empirically set to 69×69 for our processing, although the other values may be also appropriate. The exponential function (Equation 4) produces significant enhancement when the contrast is low, while it provides less enhancement if the contrast is already high. The examples of retinal images after the contrast enhancement (Osareh et al., 2009). While the contrast enhancement improves the contrast of exudates, it may also enhance the contrast of some non-exudates background pixels, so that these pixels can wrongly be identified as exudates. The third preprocessing step, we used median filtering to remove or reduce noise in an image. The benefit of median filtering operation is simultaneously reducing noise and preserving edges. The median filter considers each pixel in the image in turn and looks at its nearby neighbors to decide

JCS
whether or not it is representative of its surroundings. Instead of simply replacing the pixel value with the mean of neighboring pixel values, it replaces it with the median of those values. The median is calculated by first sorting all the pixel values from the surrounding neighborhood into numerical order and then replacing the pixel being considered with the middle pixel value. Figure 4 and 5 illustrates an example of median filtering and example calculation. Finally, Red, RGB vlaue of original retinal image is transformed to Luv color space.

Detection of Optic Disc
Identification of optic disc plays an important role in detect of exudates. However, the optic disc is similar intensity with exudates. Therefore, to prevent the optic disc from interfering with exudates, our proposed a new method to detection of optic disc before identification of exudates (Wisaeng et al., 2013).

Identification Using Fuzzy Logical Operator
The fuzzy logical operator is a data clustering algorithm in which each data point belongs to a cluster, to a degree specified by a membership grade. Cannon et al. (1986) originally introduced this technique. Fuzzy partitions a collection of n vectors, c j , j = 1,…,n, into c fuzzy groups and finds a cluster center in each group, such that a cost function based on a distance function is minimized. Because of the fuzzy partition, a given data point can belong to several groups with a degree of membership specified by a grade between 0 and 1, such that the constrains in Equation (5) are statisfied, (Gil et al., 1999). These values are collected in a membership matrix, U. However, imposing normalization stipulates that the summation of degrees of memberships for a data set must always be equal to unity (Homaeinezhad et al., 2011): Where: m = An exponential weighting function that controls the fuzziness of the membership function N = Number of features, 5 for our case C = Number of cluster, u ij is the degree of membership of x i in the cluster j x i = The ith of d-dimensional measured data c j = The d-dimension center of the cluster and| |·| = The Euclidean distance between the ith cluster and the jth data point The squared distance between ith and jth shown in Equation (6) is computed in A-norm via Equation (7). Fuzzy partitioning is carried out through an iterative optimization of the objective function shown above, with the update of membership u ij and the cluster centers c j by using Equation (9 and 10), respectively: This iteration will stop when Equation (11) is satisfied: Where: T = A termination criterion K = The iteration number This procedure converges to a local minimum or a saddle point of J. The input to the fuzzy logical operator is a set of feature. The number of cluster is required at the beginning and the algorithm is proposed by (Sopharak et al., 2009) and composed following step: Step 1: Initialize the fuzzy partition matrix U = (u ij )(U (0) ) by generating random numbers in the range 0-1 subject to Equation (12): Step 2: At k-step; calculate the centers vectors C (k) = (c j ) with (U (k) ) according to Equation (10).
Step 3: Update the fuzzy partition matrix U (k) , U (k+1) by the new computed u ij according to Equation (19).
Step 4: Compute the objective function according to Equation (6). If the difference between adjacent values of the objective function is less than Termination criterion (T) the stop the iteration; otherwise return to step 2.

Identification Using Mathematical Morphology Method
The resulting image from segmentation by fuzzy logical operator is assign as input to the segmentation by using mathematical morphology method. The mathematical morphology method by dilation was applied on the previous overlay image (The resulting segmentation from fuzzy logical operator). A simple structuring that can be used for repairing the gaps until the resulting of dilation the original image with structuring element have been bridging.
The dilation can be computed by using Equation (13). With A and B as sets in x 2 , the dilation of A by B, denoted A⊕B, is defined as: where, A⊕ B is the resultant picture of the dilation of A with B. This equation is baed on obtaining the reflection of B about its origin and shifting this reflection by x. The dilation of A by B then is the set of all displacements, x, such that B and A overlap by at least one element. Based on this interpretation, Equation (13) may be rewritten as Equation (14): Set B is commonly referred to as the structuring element in dilation, as well as in other morphological operations. In matlab, dilation can be implemented by a function, called imdilate. (Satisfying, MorphPic = imdilate (OriginPic, SE); MorphPic, is the Science Publications JCS morphological picture, OriginPic, the resultant picture and SE, is the structuring element.

Performance Measurement
To automatic pixel based evaluation methods performance, the pixel based evaluation protocal is to generate a sensitivity (SE), specificity (SP) and accuracy (AC) score values of the test image (Satisfying, SE=TP/TP+FN, SE corresponds to the proportion of correctly diagnosed exudates test subjects among all exudates test subjects, SP = TN/TN+FP, SP corresponds to the proportion of correctly diagnosed non-exudates test subjects among all non-exudates test subjects and AC = TP+TN/TP+FP+FN+TN, AC is overall per-pixel success rate of the methods).
For exudates test subject, the result is True Positive (TP) if the diagnostic test outcome is exudates and False Negative (FN) if the diagnostic test outcome is non-exudates. For non-exudates test subject, the result is true negative if the diagnostic test outcome is nonexudates and false positive if the diagnostic test outcome is exudates (Tomi et al., 2011).

RESULTS AND DISCUSSION
Many experiments have been performed on exudates and non-exudates retinal images to test and validate our method. The result from the fuzzy logical operator was validated using 2,084 images with their clinician hand-labelled images in the pixel level with 96.7, 71.4 and 79% sensitivity, specificity and accuracy respectively. Figure 6 shows a simulated view which an exudates using n = 8. Afterward, The resulting segmentation from the fuzzy logical operator is assign as input to the segmentation by mathematical morphology method. The average values of sensitivity, specificity and accuracy of validation results. The fuzzy mathematical morphology successfully with sensitivity, specificity and accuracy of 92.06, 92.92 and 92.49%, respectively. From simulated experiments results, if used fuzzy logical operator only; it gives a high TP value with a high FP value. Using fuzzy mathematical morphology, gives higher accuracy with a lower FP value.

CONCLUSION
There were two main issues that were addressed in this papers: (1) how good the identification of exudates in low-resolution. (2) how the performance of the identification of exudates should be evaluated to produce reliable. In addition, a robust and automated detection of optic disc method using Ostu's method and morphological method was presented as a supplementary item.
In this study, we have proposed a new methods to automatically identified of exudates from low-resolution images and multimodal images taken from diabetic retinopathy. The study is based on the fuzzy logical operator segmentation and mathematical morphology. Five characteristics of exudates from preprocessing step, namely, retinal color normalization, local contrast enhancement, median filtering, RGB color space was transformed to Luv values and detection of optic disc are selected to input feature. The performance of methods is measured against expert ophthalmologist hand-drawn ground-truth. The sensitivity, specificity and accuracy are also used to evaluate the method. The result shows that accuracy values increase when combine the fuzzy logical operator and mathematical morphology. Overall, the simulated experiments results show that careful preprocessing step and appropriate method together provide identification of exudates performance even on low-resolution images and multimodal images.