Automated classification of exudates from

digital fundus images

Ravitej Singh Rekhi, Ashish Issac, Malay Kishore Carlos M. Travieso
Dutta Signals and Communication Department, IDeTIC
Department of Electronics and Communication University of Las Palmas deGran Canaria
Engineering, Amity University, Noida, India Las Palmas deGran Canaria, Spain
[email protected], [email protected], [email protected]
[email protected]

Abstract—Diabetic Retinopathy and Diabetic Macular Edema [1], which grouped RGB values of the retinal image into fuzzy
are diseases that affect vision and eventually may lead to sets and membership functions. In [2], a supervised learning
blindness. Early detection is a must to prevent the progression of approach has been used followed by a grading algorithm for
the disease imploring the need for effective computer-aided Diabetic Macular Edema. A. Rocha et al. [3] proposed a
diagnostic techniques. In the following research paper, a robust common analytical approach to get anidea about both bright
method has been proposed to segment hard exudates from and red lesions using a visual word dictionary. Sohini
digital, color fundus images using anisotropic diffusion and Roychowdhury et. al. [4] reported analyzing the fundus with
adaptive thresholding followed by a support vector machine for varying degrees and fields of view. Further use of classifiers
classification. The geometrical, shape and orientation features
like the GMM, kNN, support vector machine (SVM) and
have been used to correctly classify the segmented objects as
exudates or false pixels. The proposed technique has a high
AdaBoost was incorporated. [5] reported using morphological
specificity and eliminates false positives correctly when applied features and classifying the features thus extracted usinga
across a wide range of images. The exudates segmented have a technique called soft margin SVM. A. S. Ladkat et. al. [6]
high degree of accuracy and no false positives are generated in reported using a modified matched filter for segmentation of
case of non-diseased images. The proposed method has been hard exudates. [7] proposes automatic exudate detection using
tested on a total 189 images of the DIARETDB1 and MESSIDOR a combination of texture features, extracted from different local
database and achieves an accuracy of 92.13% and 90% binary pattern variants with an artificial neural network (ANN)
respectively. The proposed method can be used in the classifier. In [8], an active contour based method has been
development for some computer aided technology for ocular reported. S. Ravishankar et. al. [9] reported segmentation based
diseases detection from fundus images. on morphological operations and automated feature extraction
using a correlation based on intensity an geometry. [10]
Keywords—Medical Imaging; Diabetic Macular Edema; reported an ingenuous algorithm based on genetics to sort the
Fundus Image; Anisotropic Diffusion; Classification; SVM; features extracted using fuzzy c-means clustering. A three stage
Exudates system using a filter bank has been reported in [11]. B V Shilpa
et. al. [12] proposed an ensemble approach to detect exudates
I. INTRODUCTION
by combining morphological and logical operations. A lot of
Diabetic Retinopathy (DR), a systemic condition of the eye work has also been done for the segmenting out the optic disc
caused in people having Diabetes Mellitus can cause several from fundus images. A. Singh et al [13] has proposed a method
complications- vision loss being the most imminent. Early that utilizes a region growing technique for segmentation of
symptoms of DR include the occurrence of small, red dots optic disc. M. Partha Sarathi et al [14] has proposed a
called micro-aneurysms in the retinal surface followed by technique for localization and segmenting the optic disc using
larger hemorrhages. When the blood vessels become fragile inpainting of blood vessels. A. Issac et al [15] [16] has
and are damaged, lipids and proteins leak out from them into proposed an intensity based adaptive threshold technique to
the retina, giving rise to yellowish-white exudates. Hard segment the optic disc from fundus image. M. K. Dutta et al
exudates further arise complications if they coalesce and [17] proposed a super pixel segmentation technique using
extend into the macular region causing Diabetic Macular histogram analysis for detection of OD. Franklin et al [18] has
Edema. This leads to rapid vision loss in the patients. Hence, proposed an ANN based classification for detection of exudates
regular check-ups are a must for people susceptible to diabetes. from the fundus images. Esmaeili et al [19] has proposed a
Early detection of the symptoms of DR can vastly reduce the curvelet transform and level set method for segmentation of
chances of progression of loss of vision with timely treatment. exudates and the optic disc. Although a considerable amount of
Hence, the need for automatic algorithms that can aid in research has been performed in this field, still a need exists for
detecting these conditions by identifying and segmenting hard a robust algorithm which is able to accurately obtain the
exudates, if present, in the fundus images of the eye. exudates from the fundus colour images.
Work has been continuing to segment out the exudates by The most significant contribution of the proposed method
academia. Classification using fuzzy logic was implemented in lies in the use of an adaptive threshold for segmenting out the
exudates from the fundus images. The mean and standard Also, the position of such abnormalities in the image is an
deviation of a particular image has been strategically combined important feature for diagnosis of the severity of a disease.
and used to calculate the threshold for segmentation of
exudates. The use of statistical features for threshold There is a difference in the intensity ranges of the regions
determination makes the algorithm image independent. The corresponding to exudates and the optic disc, as compared to
proposed algorithm is tested on images of two different the darker areas of the image like blood vessels. This
databases and is able to correct segment the exudates from the conventional technique has been primarily used to isolate the
images. region of interest (ROI).
Another significant contribution lies is the use of The proposed algorithm is divided into four parts-
geometrical features, orientation and distance from optic disc to preprocessing using image normalization, anisotropic diffusion
correctly reject the false positives from the segmented images. followed by thresholding, removal of optic disc, and
Apart from optic disc and exudates, there are many other pixels classification using support vector machine.
with similar intensities which are detected during the
segmentation process. The correct rejection of such falsy pixels
is necessary to increase the final accuracy of exudates detection
from the fundus images. So, the above mentioned features for
individual objects have been determined and are subjected to
an automatic machine learning algorithm to sort them as either
exudates or non-exudates pixels.
The remaining paper is structured in the folowing way:
Section II talks about the image processing and classification
techniques which are employed in the proposed method.
Section III presents segmentation and classification results for
the images of different databases used for experimentation.
Section IV discusses some conclusions which are derived
during the experimentation process.
II. PROPOSED METHODOLOGY
The fundus of the eye can be photographed as an RGB
image and analyzed for the presence of abnormalities like
micro-aneurysms and hemorrhages. Depending upon the
position of these abnormalities and their number in the retinal
image, diseases like Diabetic Retinopathy (DR) and Diabetic
Macular Edema (DME) can be diagnosed. Prior detection and
treatment through automatic computer-aided image processing
techniques is thus of vital importance, especially in countries
that are developing where there is a considerable dearth of
medical professionals.

Fig. 2. Flowchart for proposed method

A. Pre-processing
Owing to the high contrast provided, use of green channel is
employed for the fundus images. The input fundus images
may have color variations due to image acquisition process.
Fig. 1. Fundus Image of the Eye So, after applying the brightness correction technique on a
Fig.1 shows a labelled retinal image which consists of color space transformed image, the original image is updated
various objects and abnormalities such as optic disc, blood using the image that has been transformed. The input RGB
vessels, exudates and red lesions (micro-aneurysms and image is first subjected to a normalization process in which
hemorrhages). The optic disc and retinal blood vessels are the image is transformed to a different color space and
objects which characterize a fundus image as normal. smoothing operation is performed and a new RGB image is
However, the presence of abnormalities, like exudates and red obtained by updating and combining the individual channels.
lesions, in an image decides that a particular image is diseased. The method is similar as used in [20] which utilizes the V-
channel out of the HSV color-space followed by median Histogram analysis of the image obtained after subtraction
filtering. The three RGB channels are updated using the contributes in the empirical calculation to find a suitable
resulting smoothened V-channel and concatenated to form a threshold.
single color image and the green channel image from this T = [(a*mean(s))+(b*std(s))] (3)
resultant image is extracted.
Where, T is the adaptive threshold for segmentation of high
The smoothed V-channel updates all the 3 channels in the intensity pixels, and s is the subtracted image. ‘a’ and ‘b’ are
RGB color space using values calculated according to the the weights given to the mean and standard deviation
formula: respectively
2 C. Optic Disc Removal
VBC(x,y) = 1 - { V(x ,y)-1 } (1)
The thresholded image has three types of objects- exudates,
noise and optic disc. So, a need arises to remove the noise and
optic disc for correct detection of exudates. Localization of
optic nerve head is achieved by subjecting the grayscale image
to average blurring and finding the pixel with brightest
intensity. It has been observed that the optic disc corresponds
to one of the brightest regions in the digital fundus image. The
blurring operation is achieved using an averaging filter which
is approximately equal to the size of the optic disc. The
Fig. 3. (a) Original input image (b) Normalized image (c) Green channel of the
normalized image blurring filter size plays an important role. So, the blurred
image consists optic disc region as the brightest region while
B. Anisotropic Diffusion all other regions are blurred. The localized point will lie within
Non-linear Anisotropic Diffusion uses partial derivative the optic disc and depending on whether an image belongs to
equations to reduce noise in an image wherein a threshold right or left eye, this point is shifted by some pixels in order to
function prevents diffusion along the edges thereby preserving obtain the center of optic disc. It has been observed from the
them. Each image is the resultant of the convolution between images that optic disc diameter makes up almost one-tenth of
the previous image and a modified Gaussian filter defined as the image height, therefore, a standard radius is used to create a
follows. binary mask around the localized center. The mask thus created
is then multiplied by the thresholded image obtained after
R (x,y) = Rp (x,y) * G(x,y,t) (2) subtraction.
Where, R is the resulting image, Rp the previous image, and
G is the Gaussian kernel of size NxN.
The process can be likened to non-linear (non-uniform)
blurring unlike a regular Gaussian filter, which uniformly blurs
the given image. A direct comparison between the two methods
is illustrated in Fig 4. which also demonstrates the application
of this technique in the proposed method. By preserving the
edges of the blood vessels, the same can be easily eliminated
by subtracting from the accompanying green channel image.

Fig. 5. (a) Average Filtered Image (b) Localization of OD (c) Binary Mask (d)
Multiplication of subtracted image with mask

Fig. 5 explains the process of OD localization and creating

a mask for its removal from the binary segmented image which
contains exudates, noise and optic disc. Fig 5(a) shows an
average filtered image of the fundus image with the green
channel extracted. The size of the filter is chosen to be
approximately equal to the OD diameter. Fig 5(b) shows a ‘+’
sign in blue color which represents the OD center which is
determined by detecting the pixel of highest intensity. Fig 5(c)
shows the binary circular mask which is centered at the OD
Fig. 4. (a) Gaussian filtered image (b) Anisotropic diffusion (c) Subtracted center and its radius is equal to that of OD. Fig 5(d) shows the
image from Green Channel (d) Thresholded image
finally segmented image with OD pixels removed from it. This
image is finally fed to a classifier which is capable enough to the identification of abnormalities in fundus images. The
remove the remaining noise to obtain exudates from the fundus MESSIDOR program is funded by the French Ministry of
image. Research and Defense. There are a total of 1200 images
available that have been graded according to the severity of
D. Classification diabetic retinopathy and macular edema. The proposed
A Support Vector Machine is a supervised learning algorithm is tested on 100 images which are available in TIFF
classifier which is incorporated for classification or regression format.
analysis. SVM is a binary classifier which separates the data
samples in a multi-dimensional space into different classes
using a hyperplane which depends on the kernel used. Use of
SVM for classification of segmented pixels as exudates or
normal have been employed in [21] and [22]. To distinguish
between the exudates and the false positives, the following
features are employed for classification.

TABLE I. FEATURES USED FOR CLASSIFICATION

S.No. Feature Approach
1 Mean Statistical Parameter
2 Standard Deviation Statistical Parameter
3 Variance Statistical Parameter
4 Perimeter Object Boundary
5 Area No of white pixels
6 Eccentricity A measure of how elliptical an
object is
7 Solidity A measure of how solid an
object is
8 Aspect Ratio Ratio of Major Axis to Minor
Axis Length
9 Distance from OD Using coordinates of optic disc
10 Slope with OD Using coordinates of optic disc

The matrix of training data consists of values of parameters

associated with only exudates which are then compared against
non-exudates. When the trained SVM classifier is tested for a
new fundus image, only objects corresponding to exudates are
left behind in the image while the false positives are
successfully removed as shown in Fig. 6. The image shown
belongs to MESSIDOR database.

Fig. 6. (a) Image with exudates and false positives (b) Exudates Detected

III. EXPERIMENTAL RESULTS

The algorithm has been tested on 89 images of the
DIARETDB1 database and 100 images of the MESSIDOR
database. DIARETDB1 is a global database used for
benchmarking DR. Out of the 89 images, 84 exhibit some signs
of non-proliferative diabetic retinopathy while 5 images are
there that can be considered normal. The digital fundus camera
used has a field of view of 50 degrees.
Fig. 7. Segmentation results on DIARETDB1 database (a) Original Input
The primary purpose of the global database linked to the Image (b) Final Exudates
MESSIDOR project is to validate segmentation algorithms for
 The results for the proposed work are mentioned in terms of
parameters like sensitivity, specificity and accuracy. The
reported results consider the image based diagnosis. The image
based diagnosis implies that the algorithm tests a fundus image
for the presence/absence of exudates in an image and classifies
it as affected/normal respectively.

Table II shows the use of parameters like TP, TN, FP and

FN to calculate the accuracy of exudates detection. The
formulae used to determine the accuracy parameters are
determined as follows:
𝑇𝑃
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =
𝑇𝑃 + 𝐹𝑁

𝑇𝑁
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =
𝑇𝑁 + 𝐹𝑃

𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃 + 𝐹𝑁 + 𝑇𝑁 + 𝐹𝑃
TABLE II. RESULTS OF THE ALGORITHM ON THE DATASET
Database TP FN FP TN
DIARETDB1 44 4 3 38
MESSIDOR 33 5 5 57
Total 77 9 8 95

Table III demonstrates the number of images from both the

databases used for experimentation which have been
considered to contain exudates and normal pixels. The table
also gives a description of the number of images which have
been falsely classified. On the basis of this data, the accuracy
parameters are determined which are reported as under in
Table IV.

TABLE III. RESULTS OF THE ALGORITHM

Database Sensitivity (%) Specificity (%) Accuracy (%)
DIARETDB1 91.67 92.68 92.13
MESSIDOR 86.84 91.94 90

IV. CONCLUSION
The algorithm was tested on 189 images of standard and
publically available databases, i.e. DIARETDB1 and
MESSIDOR, and successfully segmented exudates which can
be useful to grade the severity of Diabetic Macular Edema
besides detecting early signs of Diabetic Retinopathy. The
proposed method gives a high degree of specificity and
accuracy and generates a very low number of false positives
which is a desirable quality for use in real time diagnostic
applications. The threshold is adaptive making the process of
diagnosis entirely automatic. The images obtained from the
Fig. 8. Segmentation results on MESSIDOR database (a) Original Input Image fundus camera can be analyzed to detect signs of disease
(b) Final Exudates which can help in early treatment.
 ACKNOWLEDGEMENTS Workshop on Computational Intelligence in Medical Imaging (CIMI),
Singapore, 2013, pp. 30-35.
This work has been supported in part by the Grants from [12] B. V. Shilpa and T. N. Nagabhushan, "An ensemble approach to detect
Department of Science and Technology, No. exudates in digital fundus images," 2016 Second International
DST/TSG/ICT/2013/37”, The authors further express their Conference on Cognitive Computing and Information Processing
(CCIP), Mysore, 2016, pp. 1-6.
gratitude to Dr. S.C. Gupta, Medical Director of Venu Eye
[13] A. Singh, M. K. Dutta, M. Parthasarathi, R. Burget and K. Riha, "An
research centre for his invaluable contribution in classification efficient automatic method of Optic disc segmentation using region
of the images. growing technique in retinal images," 2014 International Conference on
Contemporary Computing and Informatics (IC3I), Mysore, 2014, pp.
REFERENCES 480-484.
[14] M. Partha Sarathi, Malay Kishore Dutta, Anushikha Singh, Carlos M.
[1] N. G. Ranamuka and R. G. N. Meegama, "Detection of hard exudates Travieso, Blood vessel inpainting based technique for efficient
from diabetic retinopathy images using fuzzy logic," in IET Image localization and segmentation of optic disc in digital fundus images,
Processing, vol. 7, no. 2, pp. 121-130, March 2013. Biomedical Signal Processing and Control, Volume 25, March 2016,
[2] K. S. Deepak and J. Sivaswamy, "Automatic assessment of macular Pages 108-117, ISSN 1746-8094,
edema from color retinal images," in IEEE Transactions on Medical http://dx.doi.org/10.1016/j.bspc.2015.10.012.
Imaging, vol. 31, no. 3, pp. 766-776, March 2012. [15] A. Issac, M. Parthasarthi and M. K. Dutta, "An adaptive threshold based
[3] A. Rocha, T. Carvalho, H. F. Jelinek, S. Goldenstein and J. Wainer, algorithm for optic disc and cup segmentation in fundus images," 2015
"Points of Interest and Visual Dictionaries for Automatic Retinal Lesion 2nd International Conference on Signal Processing and Integrated
Detection," in IEEE Transactions on Biomedical Engineering, vol. 59, Networks (SPIN), Noida, 2015, pp. 143-147.
no. 8, pp. 2244-2253, Aug. 2012. [16] Ashish Issac, M. Partha Sarathi, Malay Kishore Dutta, An adaptive
[4] S. Roychowdhury, D. D. Koozekanani and K. K. Parhi, "DREAM: threshold based image processing technique for improved glaucoma
Diabetic Retinopathy Analysis Using Machine Learning," in IEEE detection and classification, Computer Methods and Programs in
Journal of Biomedical and Health Informatics, vol. 18, no. 5, pp. 1717- Biomedicine, Volume 122, Issue 2, November 2015, Pages 229-244,
1728, Sept. 2014. ISSN 0169-2607, http://dx.doi.org/10.1016/j.cmpb.2015.08.002.
[5] H. Tjandrasa, R. E. Putra, A. Y. Wijaya and I. Arieshanti, "Classification [17] M. K. Dutta, A. K. Mourya, A. Singh, M. Parthasarathi, R. Burget and
of non-proliferative diabetic retinopathy based on hard exudates using K. Riha, "Glaucoma detection by segmenting the super pixels from
soft margin SVM," 2013 IEEE International Conference on Control fundus colour retinal images," 2014 International Conference on
System, Computing and Engineering, Mindeb, 2013, pp. 376-380. Medical Imaging, m-Health and Emerging Communication Systems
[6] A. S. Ladkat, S. S. Patankar and J. V. Kulkarni, "Modified matched filter (MedCom), Greater Noida, 2014, pp. 86-90.
kernel for classification of hard exudate," 2016 International Conference [18] S. W. Franklin and S. E. Rajan, "Diagnosis of diabetic retinopathy by
on Inventive Computation Technologies (ICICT), Coimbatore, 2016, pp. employing image processing technique to detect exudates in retinal
1-6. images," in IET Image Processing, vol. 8, no. 10, pp. 601-609, Oct.
[7] M. Omar, F. Khelifi and M. A. Tahir, "Detection and classification of 2014.
retinal fundus images exudates using region based multiscale LBP [19] M. Esmaeili, H. Rabbani, A. M. Dehnavi and A. Dehghani, "Automatic
texture approach," 2016 International Conference on Control, Decision detection of exudates and optic disk in retinal images using curvelet
and Information Technologies (CoDIT), St. Julian's, 2016, pp. 227-232. transform," in IET Image Processing, vol. 6, no. 7, pp. 1005-1013,
[8] B. Harangi, I. Lazar and A. Hajdu, "Automatic exudate detection using October 2012.
active contour model and regionwise classification," 2012 Annual [20] P. N. Sharath Kumar, R. R. Kumar, A. Sathar and V. Sahasranamam,
International Conference of the IEEE Engineering in Medicine and "Automatic detection of exudates in retinal images using histogram
Biology Society, San Diego, CA, 2012, pp. 5951-5954. analysis," 2013 IEEE Recent Advances in Intelligent Computational
Systems (RAICS), Trivandrum, 2013, pp. 277-281.
[9] S. Ravishankar, A. Jain and A. Mittal, "Automated feature extraction for
[21] A. Narang, G. Narang and S. Singh, "Detection of hard exudates in
early detection of diabetic retinopathy in fundus images," 2009 IEEE
colored retinal fundus images using the Support Vector Machine
Conference on Computer Vision and Pattern Recognition, Miami, FL,
2009, pp. 210-217. classifier," 2013 6th International Congress on Image and Signal
Processing (CISP), Hangzhou, 2013, pp. 964-968.
[10] A. Osareh, B. Shadgar and R. Markham, "A Computational-Intelligence-
[22] Lili Xu and Shuqian Luo, "Support vector machine based method for
Based Approach for Detection of Exudates in Diabetic Retinopathy
identifying hard exudates in retinal images," 2009 IEEE Youth
Images," in IEEE Transactions on Information Technology in
Biomedicine, vol. 13, no. 4, pp. 535-545, July 2009. Conference on Information, Computing and Telecommunication,
Beijing, 2009, pp. 138-141.
[11] A. Tariq, M. U. Akram and M. Y. Javed, "Computer aided diagnostic
system for grading of diabetic retinopathy," 2013 Fourth International