Extraction of Exudates from the Fundus Images A Review

Extraction of Exudates from the Fundus Images A Review

Vasanthi Satyananda

Dean Academics ,

Atria Institute of Technology, Bangalore

Research Scholar, VTU, Karnataka

K V Narayanaswamy

Principal,

Atria Institute of Technology, Bangalore

K Karibasappa

Principal,

Oxford College of Engineering, Bangalore.

Abstract The dawn of the 20th century has heralded the entry of technology into almost every sphere of our daily lives, specially healthcare. This paper deals with review of the various methods used to detect exudates present in the eye which is the indicator of diabetic retinopathy, a dreadful eye disease. The aim here is to go through the algorithms using image processing technique giving importance to highest possible accuracy. The paper concludes with an insight into the future scope of this research and how an economically viable solution can be further obtained.

Keywords Diabetes Mellitus; Diabetic Retinopathy; Exudates; Image Processing; Color Processing; Segmentation Techniques

1. INTRODUCTION

   Diabetes Mellitus is one of the most commonly found medical disorders in the world. It is a medical condition where the metabolism of carbohydrates, proteins, lipids, electrolytes and water is affected due to hormonal imbalance of insulin, resulting in variations in blood sugar levels [1]. According to a survey conducted by the World Health Organization in 2014, about 9% of people suffer from Diabetes [2]. Along with changes in the levels of blood sugar, Diabetes also affects other parts of the body such as heart, blood vessels, nerves, kidneys and eyes [3]. A very well known ophthalmological complication caused due to diabetes that can induce blindness is Diabetic Retinopathy (DR) [4]. In DR, the blood vessels that circulate nutrients in the retina rupture thus leaking the fluids that can solidify and cause blurred vision [5]. According to one of the surveys conducted in 2011, about 8.3% of the world adult population suffered from DR [6]. There are two types of DR Non-proliferative diabetic retinopathy (NPDR) and Proliferative diabetic retinopathy (PDR) [7]. The former is the early stage of DR, where tiny lesions form on the blood vessel of retina called microanuerysms that can cause fluid leakage. The next stage is PDR, where the retina is deprived of oxygen, thus making the blood vessels fragile. This can result in accumulation of gel-like fluid at the back of the eye. If DR is not treated at the right time, it can cause to irreversible loss of vision [8].

   Exudates are one of the parameters that deduce the presence of DR in a diabetic patient [9]. The lipid fluids that leak from the blood vessels accumulate in the retinal region, and these deposits are known as exudates [10]. There are two types of exudates Hard Exudates and Soft Exudates [11]. Soft exudates are the accumulation of the ruptured vessel debris. These are usually oval or round in shape and white or pearly white in color. These are found in fewer numbers and

   usually lie near the optic disc. Hard exudates are small or large deposits of the lipids. These appear waxy or yellowish white in color and are found in large numbers, usually underneath the blood vessels. The ophthalmologists acutely examine these exudates using a specialized camera called the Fundus camera [12].

   One of the most researched works conducted all around the globe, is to automate the process of exudate identification, to help the patients in the rural areas, who are deprived of ophthalmologists and good facilities to tackle eye related complications. The retina consists of dark and bright areas. Exudates, blood vessels and optic disk have higher contrast than any other parts of the eye. Thus extracting exudates based on only one parameter might always not result in true positives. Thus, extracting only the exudates is a complex task. This particular review paper discusses about various approaches implemented by several authors to solve our problem statement. Section 2 introduces the concepts of preprocessing and segmentation applied for automated exudate extraction as described in the papers reviewed mentioned in Section 3. Section 4 concludes the paper by presenting the scope for the future work related to automated exudate detection.

2. PREPROCESSING

   It is a well known fact that the fundus image obtained from the camera is prone to noise and uneven brightness [17] and hence before the actual extraction steps are performed, preprocessing is a must and should. Preprocessing is a crucial stage that nullifies the noise and other unwanted features of the retinal image such as blood vessels. In most of the research carried out, it was found that exudates have a higher contrast in the green (G) channel of the RGB image and therefore exudates are extracted using the green component. Few of the popular preprocessing methods have been described below.

   One major issue faced in image processing, is the problem of extracting spherical objects. The major reason for this is that they cannot be uniformly illuminated. The same concept holds good for the eye as well. The center of the fundus image is more illuminated than the rest of the retinal area and hence extraction becomes a major issue.

   Amongst various methods to obtain uniform illumination, one of the most popular methods is median filtering. The average of all the pixels included under a sliding window of size m x n is found out and this replaces the center pixel of the window as dealt in [13, 18, 19].

   Another methodology has been elaborated in [20] wherein authors state that by subtracting the results obtained from top- hat and bottom-hat transforms of the green channel, the image becomes uniformly illuminated.

   Other methods include the one explained in [21], wherein the fundus image is converted to the HSV color model. The V channel is extracted and median filtering is applied. It is then later enhanced in terms of contrast using the Contrast Limited Adaptive Histogram Equalization (CLAHE) method [22] where Adaptive Histogram Equalization (AHE) is a technique of dividing the entire image into small sections and performing histogram equalization for each segment. This is established using CDF to map the pixels in the region to enhance the contrast of the image [16]. AHE is highly prone to enhancing noise as well present in these segments. Contrast Limited AHE (CLAHE) is the improvised version of AHE, which does not amplify the noise. The transformed V component is combined with the H and S channels and converted back to the RGB color model, thus resulting in removal of noise.

   In a few cases such as [23], [24], [25], both median filtering and CLAHE method are utilized for noise removal on the grayscale component of the image which showed better performance than using one method alone.

   Papers such as [18], [26], [27], [28] perform contrast enhancement by converting RGB model to YIQ color space, replacing Y with weighted sum of Y, I and Q components and converting back to the RGB color model. Once the new model is obtained CLAHE is performed for better enhancement. [29], [18], [30], [31] also utilize CLAHE for image enhancement. This method is also comparable to the green channel extraction as discussed above.

   As per [32], the local contrast enhancement (LCE) process is applied on the median filtered output to enhance the contrast of the image to the best level. In [33], the RGB model is converted to HSV color space and the LCE process is induced for enhancement. According to [34], in order to enhance the image, histogram specification for even frequency distribution is appied followed by the LCE approach.

   [18], [26] also propose a method to brighten the exudates even more by performing white top-hat (WTH) transformation. Top-hat transformation is a procedure applied when small details have to be extracted from the image. This transformation comes in two variants white top-hat transformation and black top-hat transformation. The former is the difference between the original image and the opened image. The latter is the original image subtracted from closed image [14]. Black top-hat transformation is also known as Bottom-hat transformation [15]. WTH transformation showed better result than a simple averaging filter.
   [35] proposes to apply averaging filter to highlight only the bright areas of the retina. Later, the green channel is extracted and contrast stretching transformation is performed to brighten the high contrast areas even more. Contrast stretching transformation is also used in [36] on the grayscale component of fundus image.

   Morphological fuzzy based closing operation with a diamond shaped structuring element is performed on the contrast enhanced image, as explained in [36]. The resultant image is then added with the original image to highlight exudates. This is used in enhancement of exudates.

   Another major issue faced apart from non-uniform illumination, is the effect of blood vessels on the image utilized for exudates extraction. Several approaches have been discussed on this topic as well.

   In [37], the authors describe a methodology that utilizes median filters, to extract blood vessels. Several other approaches such as [20], [38] propose to remove blood vessels by extracting the green component and applying a high pass filter for contrast enhancement. Morphological closing with a large structuring element is performed for the complete vessel removal. The morphological closing concept has also been used in [18], [24], [39] as well.

   Once the preprocessing steps are complete, the next major step would be to extract the exudates and the procedures for the same have been elaborated in the next section.

3. SEGMENTATION OF EXUDATES

   The preprocessed image obtained using any of the methods explained in the above section, is now utilized for the segmentation of the exudates. The main challenge faced here is that the intensity-contrast characteristics of the exudates and the optic disk appear the same. Many authors have proposed algorithms where the optic disk is eliminated first and then exudates are extracted; while there is another set of authors who propose segmentation algorithms in association with feature extraction that are fed to classifiers to correctly identify the exudates and discard the optic disk and other bright areas of the retina. Both the categories have been explained below.

   [37] uses the histogram based thresholding approach, wherein the local minima of the histogram is considered. All small fluctuations are nullified, and the last minimum is considered as the threshold. This threshold is applied to extract exudates along with the optic disk. It is a well known fact that the blood vessels in the eye converge at the optic disk. Thus, the authors find out the converging point of the blood vessels and the bright area that houses this intersection point is considered as the optic disk, while the rest are declared as exudates. An accuracy of 89%, sensitivity of 92.87% and predictive value of 96.03% is achieved.

   Edge based segmentation methods can also be employed for exudates extraction. This can be seen in [23], where the authors use the Canny edge detector to identify the variations in grayscale intensity values of the retinal image that results in edges. These boundaries are optimized using a gradient image to eliminate the optic disk. The blood vessels are removed by morphologically closing the image. In order to refine the results, closing is performed once more to get only the exudates.

   In [28], edge detection is used for segmentation. The gradient of the original image is obtained. Once the edges are acquired, they are sharpened with the ramp width reduction technique, which utilizes gradient values previously attained.

   Another approach which utilizes edge based segmentation has been described in [40], wherein the green channel of the fundus image is morphologically dilated. Due to the bright features of exudates, the edges appear prominent. Dilation is then performed twice with different structuring element sizes and the difference between the dilated images results in the edges of the exudates. These edges are then filled via morphological reconstruction. Thresholding on the reconstructed image leads to a binary image that contains only exudates. The same concept is used in [41].

   A simple thresholding method can also be used to extract the exudates by choosing an appropriate threshold level, as showcased in [20]. In [42], the preprocessed image is complemented and with a threshold of 0.97, segmentation is performed. The binary image is then morphologically closed and opened for removal of blood vessels and optic disk, with a circular structuring element. Sensitivity of 96.89% and specificity 97.15% is achieved, but accuracy is not reported.

   Another approach for exudate segmentation, as demonstrated in [25] is to carry out marker based morphological reconstruction, which is performed after blood vessels and the optic disk are removed using morphological dilation. Closing operation is performed with a line structuring element of 15 and 9 corresponding to width of widest and narrowest vessels. The exudate extraction is reported to be good but there is no indicators measured.

   [35] discusses about the Watershed approach to extract exudates. After preprocessing, the contrast adjusted image is used to obtain gradient magnitude and internal markers are identified using the extended minima transformation. Also, an external marker is obtained to divide the image effectively into regions. Using these markers, the gradient image is modified using the minima imposition approach. The modified gradient image helps to get regional minima only in the marker areas. Watershed algorithm is then applied on this modified gradient for exudates extraction. Predictive value is used as the performance indicator which is 91.9%. The algorithm complexity is not discussed.
   [38] proposes to use K-means clustering method for segmentation. In most cases of data analysis, mean and covariance would not be sufficient to extract information from a huge set of data. Instead different data clustering algorithms can be implemented. One of the popular clustering algorithms is k-means clustering [45]. The data is classified by dividing it to k number of non-overlapping clusters. The grouping is done in such a way that there is a high degree of similarity within each group and low degree of similarity with different groups. Each cluster is identified with a centroid. The k-means algorithm is an iterative approach of refining similar kinds of data to be grouped together. During each trial of iteration, the centers of the clusters might change its position to obtain accuracy; but once all the groups are distinguished from each other and there is no chance of similarity, then the iterative process stops and the centers dont move further. The distance between the centers of the clusters obtained and the nearby pixels is calculated and the pixels with minimum distance are considered. The average of these pixel values is calculated till

   convergence is obtained. For machine learning, a trained set of images is acquired using the Principle Component Analysis (PCA). PCA is a data analyzing technique, which is often used for data reduction without any loss of information and data interpretation [46]. This is a multivariate technique where in the statistics or simultaneous observation and analysis of the outcome variables is conducted. The outcomes are studied and described by several quantitative dependent variables which are inter-correlated. The information from the data set is extrated and is stored as a set of orthogonal values known as Principal Components. The components are known as orthogonal values since they are the Eigen Vectors of the symmetric Covariance matrix. The number of these principal components are often lesser than or equal to the original number of variables. The characteristics of these components are plotted and a study on their similarities is conducted. The largest possible variance decides the first component. The succeeding components have the highest variance possible under the limitation that it is orthogonal to the preceding components. The resulting vectors are an uncorrelated orthogonal basis set. The PCA algorithm is sensitive to the relative scaling of the original variables. The Gaussian description method is used along with PCA and these images are used for exudate identification. Though the detection accuracy is good the performance of the algorithm has to be improved for early detection of DR.

   The preprocessed image obtained in [21] is converted to L*a*b* color space and using the Fuzzy C-means (FCM) clustering algorithm, clusters are identified in the a* and b*channels, which indicate exudates and optic disk. A specialized FCM clustering approach called the Spatially Weighted FCM clustering can also be applied for segmentation as in [32], [33]. The algorithm in [32] has to be improved to classify the segmented image into exudates and non-exudate patches.

   [18] proposes different methods through which exudates can be extracted. The first method explains that for a sliding window, the standard deviation (SD) is estimated, and this SD value is used as a threshold only in the window area. The binary image obtained is morphologically reconstructed to perform hole filling. The original and reconstructed images are subtracted and a simple threshold is applied on the difference image obtained to extract bright regions. A similar method has been proposed in [39], [31]. In the second method, global histogram analysis is performed and a threshold t1 is obtained. Similarly, using local histogram analysis a threshold t2 is acquired. This results in two binary images. These images are logically ANDed to get a segmented image that consists of only bright areas. Another approach explained in this paper, is that different features can be considered, which can be used to obtain clusters from the FCM clustering method [30]. This image is considered as a marker and the original image is considered to be a mask. Morphological reconstruction is performed and is subtracted with the original image. Thresholding is applied on the difference image to get only the bright regions. The resultant images of all steps are fed to an Adaboost classifier to extract only the exudates. Similar methods have been proposed in [26]. The accuracy is reported to be better than previous results, but there is a mixture of different pre-processing and segmentation methods and the whole process is not simple.
   [43] Proposes that once the image is preprocessed, histogram can be used to obtain the threshold for segmentation. Later different features of the bright areas are fed to a Multilayer Perceptron (MLP) classifier to yield only exudates.

   The authors of [34] use the Gaussian-smoothed histogram analysis and FCM clustering methods to extract exudates. In case of the Gaussian-smoothed histogram analysis, the histogram for all the coloured layers is estimated. Later, using a Gaussian function, the histogram is smoothed to remove all minor peaks and valleys. The valleys now obtained can be used for thresholding. The predictive value reported is 89.3% and the algorithm has to be improved for early stages of DR.

   Segmentation can also be performed by dividing the image into different regions and processing it further. According to the approach utilized in [29], the preprocessed image is resized to 800×800 pixels and is divided into sub-regions of 80×80 pixels. All the abnormal regions are identified and Gray-level Co-occurrence Matrix (GLCM) is introduced to obtain 13 features of the abnormalities, which are then fed to a Support Vector Machine (SVM) Classifier that is capable of identifying only the exudates. Similar concepts are used in [18], [21], [24], [44]. An accuracy of 71.94% for SVM method and 74.53% for multilayer perception method is reported which needs an improvement.

   [47] deals with minimum distance classifier which calculated the distance between unknown and each of the prototype pattern vector. Based on the smaller distance, the output is decided. Size, shape, color, brightness are some of the characteristics used to delineate the objects in an image.
   [48], [49] uses the combinations of the method discussed, that is thresholding and morphological operations to extract the exudates

   The severity of diabetic retinopathy is dealt in [49]. The segmentation is based on JSEG (abbreviated for J value Segmentation) algorithm , which is an unsupervised segmentation process for image and video. This includes color quantization and spatial segmentation. It is reported that the severity is high if the distance between the detected exudate and macula is smaller and vice versa.

   In this way, there exists several methodologies to perform both preprocessing and extraction of the exudates from the Fundus image of the eye. The next section describes the significant findings, conclusion and possible future scope of the research carried out in this paper.

4. CONCLUSION & FUTURE SCOPE

In this paper, a collation of various methodologies to perform both preprocessing and extraction of exudates have been presented. As explained in Section III, most of the methods utilize the green channel component along with algorithms such as the median filtering approach and Contrast Enhancement for preprocessing. With respect to exudates extraction, it can be concluded based on comparison of the results obtained from all the sources mentioned in Section IV, that utilizing the histogram based approach along with morphological operations give an accuracy of more than 90%. This ensures a total reduction in the number of false positives obtained.

Though all the techniques mentioned in this paper sound promising in terms of accuracy, a major cause of concern would be the necessity of a high speed algorithm which must complement this accuracy obtained. This is mainly due to the reason that along with an accurate result, there is also a need of a faster diagnosis in turn leading to quicker healing times and reduced delay in prognosis, thus saving eyesight as quick as possible. Hence, it becomes the need of the hour for future research work to concentrate on the same.

Also, most of the research carried out, perform deployment of algorithms on a computer, either a desktop or a laptop and none of the research actually concentrates on implementation on a handheld device. By employing an Application Specific Integrated Circuit (ASIC) approach, not only does it promise a low power solution but also an economically viable one. Research carried out in this direction would prove to be extremely advantageous especially for the eye disease treatment in developing nations where access to healthcare is a very costly affair.

REFERENCES

1. S.K.R Parvi, Introduction in Essentials of Diabetes Mellitus and its treatment by Homeopathy, B. Jain Publishers, New Delhi, India, 1995, pp iii.

2. World Health Organization (2015, January). Diabetes [Online].

   Available: http://www.who.int/mediacentre/factsheets/fs312/en/

3. National Institute of Diabetes and Digestive and Kidney Diseases (2014, April 23). Prevent diabetes problems: Keep your diabetes under control [Online]. Available: http://www.niddk.nih.gov/health- information/health-topics/Diabetes/prevent-diabetes- problems/Pages/index.aspx

4. National Eye Institute (2015, September). Facts About Diabetic Eye Disease [Online]. Available: https://nei.nih.gov/health/diabetic/retinopathy

5. University of Michigan Kellogg Eye Center. Dabetic Retinopathy [Online]. Available:

   http://www.kellogg.umich.edu/patientcare/conditions/diabetic.retinopat hy.html

6. The International Agency for Prevention of Blindness. Diabetic Retinoapthy[Online]. Available: http://www.iapb.org/vision- 2020/what-is-avoidable-blindness/diabetic-retinopathy

7. American Optometric Association. Diabetic Retinopathy [Online]. Available: http://www.aoa.org/patients-and-public/eye-and-vision- problems/glossary-of-eye-and-vision-conditions/diabetic- retinopathy?sso=y

8. Elia Duh, Non-proliferative Diabetic Retionopathy in Diabetic Retinopathy, Humana Press, Baltimore, ch. 1, pp 4.

9. Sopharak, A., et al.; Automatic detection of diabetic retinopathy exudates from non-dilated retinal images using mathematical morphology methods, Computerized Medical Imaging and Graphics,

   Volume 32, Issue 8, Elsevier, pp 720 – 727

10. Debra L Gordon, Mayer B. Davidson, M.D, Take Good Care of Ypur Eyes in The Complete Idiot's Guide To Diabetes, 2nd Edition,Alpha Books, pp 263-264

11. Ahmed E, Diseases of the Retina in Comprehensive Manual of Ophthalmology, Jaypee Brothers Medical Publishers, 2011. pp 250-252

12. George, S; Limbasiya, B; A Review Paper On Detection And Extraction Of Blood Vessels, Microaneurysms And Exudates From Fundus Images, Scientific & Technology Research International Journal of , Nov 2013, Vol. 2 Issue 11, p134.

13. Das, A., Image Enhancement in Spatial Domain, in Guide to Signals and Patterns in Image Processing: Foundations, Methods and Applications, Springer, 2015, ch. 2, sec. 4, pp. 58-60.

14. Wu, Q.; Merchant, F.; Castleman, K., Morphological Image Processing in Microscope Image Processing, Elsevier, 2008, ch. 8, sec. 3, pp. 131-133.

15. Cherifi, H.; Zain, J.M.; El-Qawasmeh, E., Individuals Identification Using Tooth Structure, Digital Information and Communication Technology and Its Applications (DIPTAP), 2011 International Conference on, Springer, New York, Part 2, Page 108

16. Bick, U.; Diekmann, F., Image Processing in Digital Mammography, Springer, Berlin, Germany, 2010, ch. 5, sec. 3, pp. 74-77.

17. Rajan, A., Detection of Diabetic Retinopathy in Fundus Image in International Journal of Science and Application, 2015, pp. 26-30.

18. Prentasic, P.; Loncaric, S., "Voting based automatic exudate detection in color fundus photographs," inSignal Processing Conference (EUSIPCO), 2014 Proceedings of the 22nd European , vol., no., pp.1816-1820, 1-5 Sept. 2014

19. Sreng, S.; Maneerat, N.; Isarakorn, D.; Pasaya, B.; Takada, J.; Panjaphongse, R.; Varakulsiripunth, R., "Automatic exudate extraction for early detection of Diabetic Retinopathy," in Information Technology and Electrical Engineering (ICITEE), 2013 International Conference on , vol., no., pp.31-35, 7-8 Oct. 2013

20. Shami, F.; Seyedarabi, H.; Aghagolzadeh, A., "Better detection of retinal abnormalities by accurate detection of blood vessels in retina," in Electrical Engineering (ICEE), 2014 22nd Iranian Conference on , vol., no., pp.1493-1496, 20-22 May 2014

21. Ravivarma, P.; Ramasubramanian, B.; Arunmani, G.; Babumohan, B., "An efficient system for the detection of exudates in colour fundus images using image processing technique," in Advanced Communication Control and Computing Technologies (ICACCCT), 2014 International Conference on, vol., no., pp.1551-1553, 8-10 May 2014

22. Guoliang Fang; Nan Yang; Huchuan Lu; Kaisong Li, "Automatic segmentation of hard exudates in fundus images based on boosted soft segmentation," in Intelligent Control and Information Processing (ICICIP), 2010 International Conference on , vol., no., pp.633-638, 13- 15 Aug. 2010

23. Mahendran, G.; Dhanasekaran, R.; Narmadha Devi, K.N., "Morphological process based segmentation for the detection of exudates from the retinal images of diabetic patients," in Advanced Communication Control and Computing Technologies (ICACCCT), 2014 International Conference on , vol., no., pp.1466-1470, 8-10 May 2014

24. Mahendran, G.; Dhanasekaran, R.; Narmadha Devi, K.N., "Identification of exudates for Diabetic Retinopathy based on morphological process and PNN classifier," in Communications and Signal Processing (ICCSP), 2014 International Conference on , vol., no., pp.1117-1121, 3-5 April 2014

25. Youssef, D.; Solouma, N.; El-dib, A.; Mabrouk, M.; Youssef, A.-B., "New feature-based detection of blood vessels and exudates in color fundus images," in Image Processing Theory Tools and Applications (IPTA), 2010 2nd International Conference on , vol., no., pp.294-299, 7-10 July 2010

26. Prentasic, P.; Loncaric, S., "Weighted ensemble based automatic detection of exudates in fundus photographs," in Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual International Conference of the IEEE , vol., no., pp.138-141, 26-30 Aug. 2014

27. Sanchez, C.I.; Mayo, A.; Garcia, M.; Lopez, M.I.; Hornero, R., "Automatic Image Processing Algorithm to Detect Hard Exudates based on Mixture Models," in Engineering in Medicine and Biology Society, 2006. EMBS '06. 28th Annual International Conference of the IEEE , vol., no., pp.4453-4456, Aug. 30 2006-Sept. 3 2006

28. Yazid, H.; Arof, H.; Mokhtar, N., "Edge sharpening for diabetic retinopathy detection," in Cybernetics and Intelligent Systems (CIS), 2010 IEEE Conference on , vol., no., pp.41-44, 28-30 June 2010

29. Hashim, M.F.; Mohd Hashim, S.Z., "Diabetic retinopathy lesion detection using region-based approach," in Software Engineering Conference (MySEC), 2014 8th Malaysian , vol., no., pp.306-310, 23- 24 Sept. 2014

30. Princye, P.H.; Vijayakumari, V., "Detection of exudates and feature extraction of retinal images using fuzzy clustering method," in Computational Intelligence and Information Technology, 2013. CIIT 2013. Third International Conference on , vol., no., pp.388-394, 18-19 Oct. 2013

31. Ranamuka, N.G.; Meegama, R.G.N., "Detection of hard exudates from diabetic retinopathy images using fuzzy logic," in Image Processing,

    IET , vol.7, no.2, pp.121-130, March 2013

32. Kande, G.B.; Savithri, T.S.; Subbaiah, P.V., "Extraction of exudates and blood vessels in digital fundus images," in Computer and Information Technology, 2008. CIT 2008. 8th IEEE International Conference on , vol., no., pp.526-531, 8-11 July 2008

33. Kande, G.B.; Subbaiah, P.V.; Savithri, T.S., "Segmentation of Exudates and Optic Disk in Retinal Images," in Computer Vision, Graphics & Image Processing, 2008. ICVGIP '08. Sixth Indian Conference on , vol., no., pp.535-542, 16-19 Dec. 2008

34. Osareh, A.; Shadgar, B.; Markham, R., "A Computational-Intelligence- Based Approach for Detection of Exudates in Diabetic Retinopathy Images," in Information Technology in Biomedicine, IEEE Transactions on , vol.13, no.4, pp.535-545, July 2009

35. Eswaran, C.; Reza, A.W.; Hati, S., "Extraction of the Contours of Optic Disc and Exudates Based on Marker-Controlled Watershed Segmentation," in Computer Science and Information Technology, 2008. ICCSIT '08. International Conference on , vol., no., pp.719-723,

    Aug. 29 2008-Sept. 2 2008

36. Bin Mansoor, A.; Khan, Z.; Khan, A.; Khan, S.A., "Enhancement of exudates for the diagnosis of diabetic retinopathy using Fuzzy Morphology," in Multitopic Conference, 2008. INMIC 2008. IEEE International , vol., no., pp.128-131, 23-24 Dec. 2008

37. Kavitha, D.; Shenbaga Devi, S., "Automatic detection of optic disc and exudates in retinal images," inIntelligent Sensing and Information Processing, 2005. Proceedings of 2005 International Conference on , vol., no., pp.501-506, 4-7 Jan. 2005

38. Silvia, R.C. Vijayalakshmi, R., "Detection of Non-Proliferative Diabetic Retinopathy in fundus images of the human retina," in Information Communication and Embedded Systems (ICICES), 2013 International Conference on , vol., no., pp.978-983, 21-22 Feb. 2013

39. Noronha, K.; Nayak, J.; Bhat, S.N., "Enhancement of retinal fundus Image to highlight the features for detection of abnormal eyes," in TENCON 2006. 2006 IEEE Region 10 Conference , vol., no., pp.1- 4, 14-17 Nov. 2006

40. Ravishankar, S.; Jain, A.; Mittal, A., "Automated feature extraction for early detection of diabetic retinopathy in fundus images," in Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on , vol., no., pp.210-217, 20-25 June 2009

41. Kumari, V.V.; Suriyaharayananm, N.; Saranya, C.T., "Feature Extraction for Early Detection of Diabetic Retinopathy," in Recent Trends in Information, Telecommunication and Computing (ITC), 2010 International Conference on , vol., no., pp.359-361, 12-13 March 2010

42. Punnolil, A., "A novel approach for diagnosis and severity grading of diabetic maculopathy," inAdvances in Computing, Communications and Informatics (ICACCI), 2013 International Conference on , vol., no., pp.1230-1235, 22-25 Aug. 2013

43. Garcia, M.; Hornero, R.; Sanchez, C.I.; Lopez, M.I.; Diez, A., "Feature Extraction and Selection for the Automatic Detection of Hard Exudates in Retinal Images," in Engineering in Medicine and Biology Society, 2007. EMBS 2007. 29th Annual International Conference of the IEEE , vol., no., pp.4969-4972, 22-26 Aug. 2007

44. Lili Xu; Shuqian Luo, "Support vector machine based method for identifying hard exudates in retinal images," in Information, Computing and Telecommunication, 2009. YC-ICT '09. IEEE Youth Conference on , vol., no., pp.138-141, 20-21 Sept. 2009

45. Morissette, L.; Chartier, S., The k-means clustering technique: General considerations and implementation in Mathematica in Tutorials in Quantitative Methods for Psychology, 2013, vol. 9(1), pp. 15-24.

46. Jolliffe, I.T., Introduction in Principal Component Analysis,

    Springer-Verlag, New York, 2002, 2nd edition, ch. 1, pp. 1-5

47. A. C. Somkuwar, T. G. Patil, S. S. Patankar and J. V. Kulkarni, "Intensity features based classification of hard exudates in retinal images," 2015 Annual IEEE India Conference (INDICON), New Delhi, 2015, pp. 1-5.

48. H. A. Nugroho, K. Z. W. Oktoeberza, T. B. Adji and M. B. Sasongko, "Segmentation of exudates based on high pass filtering in retinal fundus images," 2015 7th International Conference on Information Technology and Electrical Engineering (ICITEE), Chiang Mai, 2015, pp. 436-441.

49. M. Gandhi and R. Dhanasekaran, "Investigation of severity of diabetic retinopathy by detecting exudates with respect to macula," Communications and Signal Processing (ICCSP), 2015 International Conference on, Melmaruvathur, 2015, pp. 0724-0729.