ADCAIJ: Advances in Distributed Computing and Artificial Intelligence Journal. Vol. 8 N.

4 (2019), 41-48

ADCAIJ: Advances in Distributed Computing and Artificial Intelligence Journal

Regular Issue, Vol. 8 N. 4 (2019), 41-48
eISSN: 2255-2863
DOI: http://dx.doi.org/10.14201/ADCAIJ2019844148

Detection of Hard Exudates

in Retinopathy Images
Satya Bhushan Verma, Abhay Kumar Yadav
Department of Computer Science, BBA University (Central University), Lucknow, India
[email protected], [email protected]

KEYWORD ABSTRACT

Hard Exudates; The tissue layer located at the back of the eye is known as retina which converts the
Retinopath; incoming light into nerve signals and those signals are sent to the brain for understand-
Fundus Image; ing. The damage onto the retina is termed as retinopathy and that may lead to vision
Cotton Wool Spot; weakening or vision loss. The hard exudates are small white or yellowish white deposits
Retina with their edges being clear and sharp. In the proposed methods we take color image of
retina then extract the green channel of that image then apply top hat transformation and
bottom hat transformation on that image. The DIARETDB1 and High-Resolution Fun-
dus (HRF) databases are used for performance evaluation of the proposed method. The
proposed technique achieves accuracy 97%, sensitivity 95%, and specificity 96% and it
takes average 5.6135 second for detection of hard exudates in an image.

1. Introduction
Retina is the tissue layer located at the back of the eye that converts incoming light into nerve signals and
then sent that signal to the brain for understanding. The damage onto the retina is called retinopathy which
may lead to vision weakening or vision loss. Diabetes or hypertension is also detected by retinopathy. The
retina’s walls of blood vessel are thickened and narrow by constant high blood pressure. This places pressure
on optic nerve and cause human vision difficulties. This type of condition is known as hypertensive retinop-
athy (Diana L. Shechtmanet al. 2007). If any signs of hypertensive retinopathy are present on the retina then
it signifies the presence of soft exudates and hard exudates. Cotton Wool Spots (CWS) is also called soft
exudates. Cotton wool spots are small, yellowish-white or grayish-white slightly elevated lessons which look
like clouds on retina. CWS edges are blurry and not defined easily. The hard exudates are small white or
yellowish white deposits with sharp margins.
Nowadays the role of image processing in automated identification of diseases in the retina is vital. It of-
fers a non-invasive technique for the recognition of numerous retinal diseases like hypertensive retinopathy,
diabetic retinopathy etc. Detected result will help us to take quick conclusion for the automatic appointments
to the ophthalmologists. Detection of contributing signs of a diseased retina from the fundus image (Suma G.
et al. 2018) will helps in early diagnosis of the disease and necessary treatment can be carried on further. Our
proposed method tends to provide a quick and effective hard exudates detection technique. In United States the
Diabetic retinopathy (DR) is one of the most common reasons of visual damage among working-age adults.
The diabetic retinopathy happens when blood vessels are affected by diabetes in the retina (the light-sensitive

ADCAIJ: Advances in Distributed Computing

and Artificial Intelligence Journal
Satya Bhushan Verma and Abhay Kumar Yadav
Detection of Hard Exudates in Retinopathy Images Regular Issue, Vol. 8 N. 4 (2019), 41-48
eISSN: 2255-2863 - http://adcaij.usal.es
Ediciones Universidad de Salamanca - cc by nc dc

41
tissue in the back of the eye), causing them to leak and distort vision. If this is not found and treated early, then
diabetic retinopathy can causes permanent vision loss” (U. M. Akram et al. 2011).

Figure 1: (A) Healthy retina (B) Retina with Hard Exudates.

The severity of Diabetic retinopathy is categorized based upon number of micro aneurysms, hemorrhages,
exudates, and in neovascularization. The growth of Diabetic retinopathy is divided into normal retina, back-
ground DR, nonproliferative DR (NPDR), proliferative DR (PDR), and macular edema (ME). Diabetic ret-
inopathy is a general microvascular problem. Diabetic macular edema is more common in type 2 diabetes,
approximately 7.5% occurs in diabetic patients, and this is an important reason of blindness in working age
person (Rajashekar D.et al. 2016).

2. Literature Review
Shengchun Long et al. (Shengchun L. et al. 2019) proposed an algorithm for Hard Exudates detection. Initially
they make an automatic retinal image preprocessing method by using the active threshold technique and fuzzy
C-means clustering (FCM) technique, then they used a classifier named support vector machine. Their proposed
algorithm contains four stages first is preprocessing, second is localization of the optic disc, the third stage is
the determination of candidate hard exudates by using dynamic threshold in the grouping with global threshold
which is based on the FCM, and last stage is feature extraction. In the last stage, the candidate Hard Exudates
region extracted the eight texture feature, which was then fed into an SVM classifier for the automatic Hard
Exudates classification. They used DIARETDB1 (diaretdb1) and the e-ophtha EXretinal image database for the
examination of their proposed algorithm. On the e-ophtha EX and DIARETDB1 retinal database they trained
and cross-validated (10 fold) at the pixel-level. On e-ophtha EX database they achieved average sensitivity
76.5%, PPV 82.7%, and F-score 76.7%, and on the DIARETDB1 retinal database. They achieved average sen-
sitivity 97.5%, specificity 97.8%, and accuracy 97.7%.
Iqbaldeep Kaur et al. (IqbaldeepK., et al. 2016.) managed to describe different works that were needed for
the automatic identification of HEs and CWSs in retinal images for detection of diabetic retinopathy and images
classification based on support vector machine (SVM). They evaluated their system on a large dataset having
129 retinal images. Their proposed method achieved sensitivity 96.9%, specificity 96.1% and accuracy 97.38%
for the detected exudates from a database.
Another interesting approach was proposed by Abhilash Goud Marupally et al (Abhilash G. M., et al. 2017)
they put forward a Semi-automated quantification technique for detection of hard exudates in photographs of
color fundus that were diagnosed with the diabetic retinopathy. Color fundus photographs of 30 eyes were taken
from 30 subjects, 21 males and 9 females. They were able to develop a semi-automated algorithm to quantify

ADCAIJ: Advances in Distributed Computing

and Artificial Intelligence Journal
Satya Bhushan Verma and Abhay Kumar Yadav
Detection of Hard Exudates in Retinopathy Images Regular Issue, Vol. 8 N. 4 (2019), 41-48
eISSN: 2255-2863 - http://adcaij.usal.es
Ediciones Universidad de Salamanca - cc by nc dc

42
the area covered by HEs. Two different methodologies were given (i) top-hat filtering, second order statistical
filtering, (ii) color fundus images thresholding. They were able to detect 60–90% of the HEs area in 13 images
and 90–100% in other 17 images.
Kavitha et al. (Kavitha, D. et al. 2005) uses the histogram based thresholding approach, wherein the local
minima of the histogram were considered. They ignored the small fluctuations, and the last minimum was con-
sidered as the threshold value. This threshold value was applied to detect exudates with the optic disk of the
eye. The blood vessels converge at the optic disk. Thus, the authors were able to find out the converging point
of the blood vessels and the bright area that has kept this intersection point was considered as the optic disk,
while the others were declared as exudates. An accuracy of 89%, sensitivity of 92.87% and predictive value of
96.03% was achieved.
Punnolil, A. et al (Punnolil, A. 2013.), proposed a simple thresholding method that can be used to extract
the exudates by choosing an appropriate threshold level, their preprocessed image was complemented and had
a threshold value of 0.97, on which segmentation was performed. The binary image was 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% was achieved, but accuracy was not reported.

3. Proposed Methodology
The Proposed method involves following steps

3.1. Image Acquisition

The DIARETDB1 database [DIARETDB1] contains of 89 color fundus images, 84 images contain at least
mild non-proliferative signs of the diabetic retinopathy, and the other 5 images are normal with no any signs
of the diabetic retinopathy. The images were taken by the same 50-degree field-of-view using digital fundus
camera with the varying imaging settings. High-Resolution Fundus (HRF) Image Database (High-Resolution
Fundus (HRF)) contains total 45 images (15 images of healthy patients, 15 images of glaucomatous patients
and 15 images of patients with diabetic retinopathy). Online databases with ground truth annotations are very
important in helping researchers to compare the performance of different methods in medical imaging. Ground
truth annotation are prepared first by computer vision researchers and then used by medical experts to select
and annotate pathological signs (Kauppi T. et al. 2013.).This collaborative work creates ground truth images for
computer-aided diagnosis systems.

3.2. Image Preprocessing

Image acquisition in retinal image mainly focuses on the OD area (nasal view) and the macular area In DI-
ARETDB1, HRF, and local datasets, the captured images mainly focus on the macular area, which makes the
illumination high at the macular area and makes the pixels at the outer ring of the image saturated (Rajashekar
D.et al. 2016.). This might be the cause of image quality degradation and the presence of artifacts or noises
during the process of image acquisition. Image preprocessing step is an essential step to improve the image
quality (Verma, S. et al. 2019.). The green channel of RGB color space is selected because of its good contrast.
Then the image enhancement and noise filtering are considered to enhance the contrast of the image and remove
the noises respectively.

ADCAIJ: Advances in Distributed Computing

and Artificial Intelligence Journal
Satya Bhushan Verma and Abhay Kumar Yadav
Detection of Hard Exudates in Retinopathy Images Regular Issue, Vol. 8 N. 4 (2019), 41-48
eISSN: 2255-2863 - http://adcaij.usal.es
Ediciones Universidad de Salamanca - cc by nc dc

43
 Figure 2: (A) Original Retinal Image (B) Retinal Image with green Channel Component.

3.3. Image segmentation

Image segmentation is one of the main steps in image processing techniques (Akram F, et al. 2017.). Several
methods have been proposed to segment images based on attributes such as intensity, color, and texture. Thresh-
olding is a useful image segmentation method and is frequently employed in medical imaging. It can be further
classified into local thresholding and global thresholding.

3.4. Morphological Top Hat Transformation

The top-hat transformation is a process that removes small details from the images. There are two types of top-
hat transformation those are white top-hat transformation and black top-hat transformation (E. Aswini, et al.
2013.). The black top-hat transform is also known as bottom-hat transformation (A. Halder and P. Bhattacharya,
2015).
Let 𝑓: 𝐸 → 𝑅 be an image in grayscale,
Where, E is a mapping point from a discrete grid or Euclidean space. Let assume 𝑏(𝑥) be a grayscale struc-
turing element.
Then, the white top-hat transform of 𝑓 is given by: 𝑇�(𝑓) = 𝑓 – 𝑓 ∘ 𝑏, the black top-hat transform (bottom-hat
transformation) of 𝑓 is given by: 𝑇�(𝑓) = 𝑓 • 𝑏 – 𝑓.
Where
“∘” denotes the opening operation.
“•” denotes the closing operation.

Figure 3: (A) fundus region after applying Morphological Bottom Hat (B) Fundus region after applying
Morphological Top Hat.

ADCAIJ: Advances in Distributed Computing

and Artificial Intelligence Journal
Satya Bhushan Verma and Abhay Kumar Yadav
Detection of Hard Exudates in Retinopathy Images Regular Issue, Vol. 8 N. 4 (2019), 41-48
eISSN: 2255-2863 - http://adcaij.usal.es
Ediciones Universidad de Salamanca - cc by nc dc

44
3.5. Feature Extraction
In the Proposed method top-hat transformation and bottom-hat transformation is used for the feature extraction.

𝐻𝑒 = 𝑇�(𝑓) – 𝑇�(𝑓) (1)

Where 𝑇�(𝑓) is top-hat transformation and 𝑇�(𝑓) is bottom-hat transformation.

Figure 4: (A) Retina Image with Hard Exudates (B) Segmented Region of Hard Exudates.

4. Experiment and Result

High-Resolution Fundus (HRF) database and DIARETDB1 database are used for performance evaluation of
the proposed method. Accuracy, sensitivity, and specificity are three performance metrics used for confirm the
efficiency of the proposed method. Correctness of a diagnostic test is defined by accuracy (eq.2). The perfor-
mance of positive disease detection is defined by sensitivity (eq.3) and specificity is known as chance of correct
non-disease detection.

TP + TN
Accuracy = (2)
TP + TN + FN + FP

TP
Sensitivity = (3)
TP + FP

TN
Specificity = (4)
TN + FN

Where,
TP: Number of correctly classified Hard Exudates.
TN: Number of correctly classified Non-Hard Exudates.
FP: Number of wrongly classified Hard Exudates.
FN: Number of wrongly classified Non-Hard Exudates.

ADCAIJ: Advances in Distributed Computing

and Artificial Intelligence Journal
Satya Bhushan Verma and Abhay Kumar Yadav
Detection of Hard Exudates in Retinopathy Images Regular Issue, Vol. 8 N. 4 (2019), 41-48
eISSN: 2255-2863 - http://adcaij.usal.es
Ediciones Universidad de Salamanca - cc by nc dc

45
 Figure 5: (A) Healthy retina (B) There is no any sign of Hard Exudates in retina.

Figure 6: (A) Retina with Hard Exudates (B) Area of Hard Exudates in the retina.

The above figure 5(A) is an image of healthy retina and when this image passes with the proposed system
resultant is figure 5(B) that clearly show there no any sign of hard exudates, but image in figure 6(A) passes
through the proposed method then area of hard exudates is easily located. The following table1 shows the ex-
perimental results of the proposed method in the term of accuracy, sensitivity, and specificity those are 95%,
94%, 96% in the DIARETDB1 database; and 97%, 95%, 95% in the High-Resolution Fundus (HRF) database.

Table 1: Accuracy, Sensitivity, and Specificity in the DIARETDB1and High-Resolution Fundus (HRF) database

Name of Database Performance Metrics Results

DIARETDB1 Accuracy 95%
Sensitivity 94%
Specificity 96%
High-Resolution Fundus (HRF) Accuracy 97%
Sensitivity 95%
Specificity 95%

ADCAIJ: Advances in Distributed Computing

and Artificial Intelligence Journal
Satya Bhushan Verma and Abhay Kumar Yadav
Detection of Hard Exudates in Retinopathy Images Regular Issue, Vol. 8 N. 4 (2019), 41-48
eISSN: 2255-2863 - http://adcaij.usal.es
Ediciones Universidad de Salamanca - cc by nc dc

46
 The experiment of the proposed model is carried out by using MATLAB R2015a on laptop with Intel core
i3-23102.1 GHz processor and 4 GB RAM. The proposed method takes 5.6135 second for detection of hard
exudates in retinopathy image. The proposed method is quite fast compared to other techniques.

5. Conclusion
The role of image processing is vital in the automated detection of diseases in retina. It provides a non-inva-
sive methods for detection of many retinal diseases such as hypertensive retinopathy, diabetic retinopathy etc.
In retina if any indication of hypertensive retinopathy then it contains the presence of soft exudates and hard
exudates. Cotton wool spots (CWS) or soft exudates are small, light yellowish-white or grayish-white slightly
higher lesions which appears as clouds on retina and their edges are blurry and not clearly defined. The hard
exudates are small white or yellowish white deposits and their edges are clear and sharp. In the proposed meth-
ods we take color image of retina then extract the green channel of that image then apply top hat transformation
and bottom hat transformation on that image. In the proposed technique we achieve accuracy 95%, sensitivity
94%, and specificity 96% in the DIARETDB1 database; and accuracy 97%, sensitivity 95%, specificity 95% in
the High-Resolution Fundus (HRF) database. The proposed method takes average 5.6135 second for detection
of hard exudates in an image.

6. References
A. Halder and P. Bhattacharya, 2015. An application of Bottom Hat transformation to extract blood vessel from
retinal images in International Conference on Communications and Signal Processing (ICCSP), Melmaru-
vathur, pp. 1791-1795.
Abhilash G. M., et al. 2017. Semi-automated quantification of hard exudates in color fundus photographs diag-
nosed with diabetic retinopathy in BMC Ophthalmol 17, 172, doi:10.1186/s12886-017-0563-7
Akram F, et al. 2017. Active contours driven by local and global fitted image models for image segmentation
robust to intensity inhomogeneity.PLoS One12(4):e0174813
Diana L. Shechtmanet al. 2007. Hypertension: More Than Meets the Eye in Review of Optometry, Vol.No:
144:09.
DIARETDB1 - Standard Diabetic Retinopathy Database,https://www.it.lut.fi/project/imageret/diaretdb1/index.
html
E. Aswini, et al. 2013. Mathematical morphology and bottom-hat filtering approach for crack detection on relay
surfaces in International Conference on Smart Structures and Systems - ICSSS'13, Chennai, pp. 108-113.
High-Resolution Fundus (HRF) Image Database, https://www5.cs.fau.de/research/data/fundus-images/
IqbaldeepK., et al. 2016. Automated Identification of Hard Exudates and Cotton Wool Spots using Biomedical
Image Processing, International Journal of Computer Science And Technology, V. 7, I.4
Kauppi T. et al. 2013. Constructing benchmark databases and proto-cols for medical image analysis: Diabetic
retinopathy.Computational and Mathematical Methods in Medicine, pp-1 – 15.
Kavitha, D. et al. 2005. Automatic detection of optic disc and exudates in retinal images, inInternational Con-
ference on Intelligent Sensing and Information Processing, vol., no., pp.501-506.
Punnolil, A. 2013.A novel approach for diagnosis and severity grading of diabetic maculopathy in Internation-
al Conference on Advances in Computing, Communications and Informatics (ICACCI), 2013, vol., no.,
pp.1230-1235, 22-25.
Rajashekar D.et al. 2016.Comprehensive retinal image analysis for aggressive posterior retinopathy of prema-
turity.PLoS One11(10)
Shengchun L. et al. 2019. Automatic Detection of Hard Exudates in Colour Retinal Images Using Dynamic
Threshold and SVM Classification: Algorithm Development and Evaluation, in BioMed Research Interna-
tional, doi: 10.1155/2019/3926930

ADCAIJ: Advances in Distributed Computing

and Artificial Intelligence Journal
Satya Bhushan Verma and Abhay Kumar Yadav
Detection of Hard Exudates in Retinopathy Images Regular Issue, Vol. 8 N. 4 (2019), 41-48
eISSN: 2255-2863 - http://adcaij.usal.es
Ediciones Universidad de Salamanca - cc by nc dc

47
Suma G. et al. 2018. Detection of Neovascularization in Proliferative, Diabetic Retinopathy Fundus Images,
The International Arab Journal of Information Technology, 15, 1000-1009.
U. M. Akramet al. 2011. Automated Detection of Dark and Bright Lesions in Retinal Images for Early Detection
of Diabetic Retinopathy, Springer Science Business Media, LLC. doi: 10.1007/s10916-011-9802-2
Verma, S. et al. 2019.Contactless palmprint verification system using 2-D Gabor filter and principal component
analysis.The International Arab Journal of Information Technology.16(1), pp-23–29.

ADCAIJ: Advances in Distributed Computing

and Artificial Intelligence Journal
Satya Bhushan Verma and Abhay Kumar Yadav
Detection of Hard Exudates in Retinopathy Images Regular Issue, Vol. 8 N. 4 (2019), 41-48
eISSN: 2255-2863 - http://adcaij.usal.es
Ediciones Universidad de Salamanca - cc by nc dc

48