International Journal of Computer Applications (0975 – 8887)

National Conference on Recent Trends in Computer Applications NCRTCA 2013

Adaptive Neuro Fuzzy Inference System Assisted

Diagnosis of Diabetic Retinopathy from Fundus Image

G.Anidha U.Sabura Banu

BS Abdur Rahman University BS Abdur Rahman University
Assistant professor/EIE Professor/EIE
Vandalur, Chennai-48 Vandalur, Chennai-48

ABSTRACT a segmentation method to differentiate the contrast in larger

In this paper, it is proposed to detect the exudates in the and thin blood vessels. Adaptive local thresholding is used to
retinal image and to classify the severity stages caused by the produce the normalized image and to extract larger vessels.
exudates and non-exudates using Adaptive Neuro Fuzzy Thin vessel segments are classified using Support Vector
Inference System (ANFIS). The retinal image used for this Machine. Different stages of Diabetic retinopathy disease
project work is subjected to the preprocessing steps such as severity are detected by Morphological operation and Texture
green channel extraction, median filter, histogram Analysis methods applied on retinal images. The statistical
equalization and contrast enhancement. Then the image is features are extracted and classified using Bayes Minimum
subjected to morphological operation for dilation and erosion. Distant Discriminant (MDD) classifier and the classifier is
Optic disk is eliminated by connected component technique compared with original and brightness enhanced image.
and statistical features like exudates area, size, color, energy, Niemeijer et al. proposed a method to differentiate the bright
skewness, kurtosis, entropy, homogeneity and texture lesions such as exudates, cotton wool spots and drusen from
properties are extracted. The features are now classified to colour retinal images. Sinthanayothin C et al. developed a
identify the normal eye and affected eye. The classification is Recursive region growing segmentation technique which is
accomplished using ANFIS. used for the automated system of detection of diabetic
retinopathy stages. Huiqi Li and Opas Chutatape have
General Terms presented a method to detect exudates using region growing
Intelligent Classification and edge detection techniques. He also detected optic disk
using principal component analysis. Using a modified active
Keywords shape model the shape of optic disk was detected. Sánchez et
Retina, Exudates, Diabetic retinopathy, Adaptive Neuro al. detected hard exudates by colour, statistical classification
Fuzzy Inference System classification and sharpness of its edges using Kirsch operator.A fundus
coordinate system is used to describe the features. Usher et al.
1. INTRODUCTION found an adaptive intensity thresholding method for the
All people with diabetes mellitus are at risk – those with Type extraction of exudates. The features extracted are size, shape,
I diabetes and those with Type II diabetes. The longer a hue and intensity and are used as the input to artificial neural
person has diabetes, the higher the risk of developing some network. Prof B. Venkatalakshmi and V.Saravanan presented
ocular problem. Between 40 to 45 percent of diagnosed with a method to detect hard exudates using color and sharp edges
diabetes have some stage of diabetic retinopathy. Diabetic of lesion. Graphical user interface window created using
retinopathy is oozing of exudates in the retina. Alireza Osareh MATLAB used in examining the abnormal fundal retinal
et al . developed a tool to segment the retinal images using images automatically and thus reducing the examining time of
Fuzzy C-Means (FCM) clustering technique. The features are the physician. Walter et al. have presented and discussed an
extracted and ranked using Genetic Algorithm and classified algorithm for detection of exudates, as well as detection of
using neural network. Akara Sopharak et al. detected optic disk which was essential for this approach. Exudates
exudates from nonmydriatic, low contrast retinal digital were found using their grey level variation, and their contours
images using mathematical morphology techniques. Akara were determined by means of morphological reconstruction
Sopharak et al. proposed an automatic method for detection of techniques. The optic disk was detected by means of
exudates from the diabetic retinopathy images using fuzzy c- morphological filtering techniques and the watershed
means clustering technique. Asha Gowda Karegowda et al. transformation was to find its contours. V. Vijaya Kumari, N.
developed a Back propagation Neural network classifier to Suriya Narayanan developed a method for early detection of
detect exudates in retinal images. The significant features are Diabetic Retinopathy. Optic disk is extracted by propagation
identified by Decision Tree and GA- CFS method. Doaa through radii method and exudates are detected using feature
Youssef et al. found that Green channel have high contrast extraction, template matching and enhanced MDD classifier.
and necessary information (exudates). Canny edge detector In the proposed work, it is planned to obtain the retinal images
and Hough transform was used for optic disk elimination and using fundus camera, preprocess the image and separate the
morphological operation for detection of blood vessel tree. A extrudates in the image and classify the eye image as normal
Bayesian classifier was used for classification of images into and affected eye.
exudates and non-exudates was given by Ege et al.. Huan
Wang et al. Detected exudates by using a minimum-distance 2. DIABETIC RETINOPATHY
discriminant classifier based on statistical pattern recognition The retina is a multi-layered sensory tissue that lines the back
and used local window for classification. Lili Xu et al. use of the eye. It contains millions of photoreceptors that capture

32
 International Journal of Computer Applications (0975 – 8887)
National Conference on Recent Trends in Computer Applications NCRTCA 2013

light rays and convert them into electrical impulses. These pre-processing describes any type of processing performed
impulses travel along the optic nerve to the brain where they on raw data to prepare it for another processing procedure.
are turned into images. There are two types of photoreceptors Commonly used as a preliminary data mining practice, data
in the retina: rods and cones. pre-processing transforms the data into a format that will be
more effective to feed to the next process.
3.1 Grayscale Conversion
The RGB color model is an additive colour model in which
red, green, and blue light are added together in various ways
to reproduce a broad array of colors. The name of the model
comes from the initials of the three additive primary colors,
red, green and blue. In photography and computing a gray
scale or grey scale digital image is an image in which the
Figure 1: Eye affected by Diabetic Retinopathy value of each pixel is a single sample, that is, it carries only
intensity information. Images of this sort, also known
The retina contains approximately 6 million cones. The cones as black-and-white, are composed exclusively of shades
are contained in the macula, the portion of the retina of gray, varying from black at the weakest intensity to white
responsible for central vision. They are most densely packed at the strongest. Grayscale images are often the result of
within the fovea, the very centre portion of the macula. Cones measuring the intensity of light at each pixel in a single band
function best in bright light and allow us to appreciate colour. of the electromagnetic spectrum (e.g. infrared, visible light,
When the retina is been affected as a result of diabetes, this ultraviolet, etc.), and in such cases they are monochromatic
type of disease is called Diabetic Retinopathy (DR) , if not proper when only a given frequency is captured. But also they
properly treated it might eventually lead to loss of vision. can be synthesized from a full color image; see the section
Ophthalmologists have come to agree that early detection and about converting to grayscale. This input image is subjected to
treatment is the best treatment for this disease. The retinal the grayscale conversion (see Figure 2)
image affected by DR is shown (see Figure 1).DR occurrence
have been generally categorize into three main form viz,
Background DR, Proliferative DR, Severe DR. Retinopathy
(damage to the retina) is caused by complications of diabetes,
which can eventually lead to blindness. It is anocular
manifestation of diabetes, a systemic disease, which affects up
to 80 percent of all patients who have had diabetes for 10
years or more. Despite these intimidating statistics, research
Figure 2: Gray scale converted retinal image
indicates that at least 90% of these new cases could be
reduced if there was proper and vigilant treatment and 3.2 Green Channel Enhancement
monitoring of the eyes. The longer a person has diabetes, the The image is pre-processed with green channel enhancement
higher his or her chances of developing diabetic retinopathy. .Using this red and blue components in the image are
eliminated. The exudates appear bright in green channel
3. PROCESSING OF RETINAL IMAGES compared to red and blue (see Figure 3). The exudates appear
The retinal image used for this research work is subjected to bright in the green channel compared to red and blue channels
the pre-processing steps such as green channel extraction, in RGB image. Hence green channel is used for further
histogram equalization and contrast enhancement. The optic processing by neglecting other two components
disk is eliminated by connected component analysis and
features like exudates area, size, color, and homogeneity and
texture properties are extracted. The images for this project
work are taken from publicly available data on the website
“Drive”. The images forms a dataset of 100 colour fundus
images in which 32 are normal and 68 contains exudates
patches with different stages of severity. Each image was Figure 3: Green Channel Enhanced Image
captured using 24 bit per pixel at a resolution of 774 X 893
pixels in JPEG format. The retinal images in the dataset are 3.3 Histogram Equalization
often noisy and poorly illuminated because of unknown noise Histogram equalization is used to increase the
and camera settings. Also there is a wide variation of colour global contrast of the images, especially when the
of retina from patient to patient. Thus the images are usable data of the image is represented by close contrast
subjected to various preprocessing steps, which include green values. Through this adjustment, the intensities can be better
channel extraction, histogram equalization, median filter and distributed on the histogram. This allows for areas of lower
contrast enhancement. The retinal images and the local contrast to gain a higher contrast. Histogram
preprocessed images for normal, mild, moderate and severe equalization accomplishes this by effectively spreading out
stages of exudates are shown in below image respectively. the most frequent intensity values. Histogram equalization is
The exudates appear bright in the green channel compared to a specific case of the more general class of histogram
red and blue channels in RGB image. Hence green channel is remapping methods. These methods seek to adjust the image
used for further processing by neglecting other two to make it easier to analyze or improve visual quality. This
components. Histogram equalization and contrast method increases global contrast of images and hence the
enhancement are used to increase the contrast between the intensities can be better distributed on the histogram. So the
exudates and the image background. Median filter are used in hidden features of the image are more visible. Histogram
addition for removing noise in the image. The median filter provides economic hardware implementations. Histogram is
removes the noise like salt and pepper noise and etc. Data the basic for a number of spatial domain processing

33
 International Journal of Computer Applications (0975 – 8887)
National Conference on Recent Trends in Computer Applications NCRTCA 2013

techniques. They are simple to calculate in software. In this element to an input image, creating an output image of the
image technique, we are able equalize the image pixels. The same size. In a morphological operation, the value of each
histogram equalized image shows how the pixels are equally pixel in the output image is based on a comparison of the
distributed (see Figure 4, 5, 6). corresponding pixel in the input image with its neighbors. By
choosing the size and shape of the neighborhood, you can
construct a morphological operation that is sensitive to
specific shapes in the input image. The most basic
morphological operations are dilation and erosion. Dilation
adds pixels to the boundaries of objects in an image, while
erosion removes pixels on object boundaries (see Figure 8).

Figure 4: Histogram equalized retinal image

Figure 8: Retinal image after dilation and erosion

4.1 Optic Disc Elimination
The morphologically operated image has only exudates and
optic disk. The optic disk occupied maximum area in the
Figure 5. Histogram of the original image image and hence it is eliminated by connected component
analysis. The optic disk eliminated image for normal, mild,
moderate and severe stages of Diabetic Retinopathy
 The optic disk occupies the maximum area in
the image
 Now it is eliminated using connected
component analysis
 This is done using connected component
analysis
Connected-component labeling is used in computer vision to
detect connected regions in binary digital images,
although color images and data with higher dimensionality
can also be processed. When integrated into an image
Figure 6. Histogram of the equalized recognition system or human-computer interaction interface,
Image connected component labeling can operate on a variety of
3.4 Contrast Enhancement information. Blob extraction is generally performed on the
The contrast enhancement is used to increase the contrast resulting binary image from a thresholding step. Blobs may be
between the exudates and the background. (see Figure 7). counted, filtered, and tracked. Blob extraction is related to
but distinct from blob detection. A graph,
containing vertices and connecting edges, is constructed from
relevant input data. The vertices contain information required
by the comparison heuristic, while the edges indicate
connected 'neighbors'. An algorithm traverses the graph,
labeling the vertices based on the connectivity and relative
values of their neighbors. Connectivity is determined by the
medium; image graphs, for example, can be 4-connected or 8-
Figure 7: Retinal image after contrast connected. This image shown in figure 9 shows the exudates
Enhancement of the eye and optic disk has been removed by connected
component (see Figure 9).
4. EXUDATES SEGMENTATION
In computer vision, image segmentation is the process of
partitioning a digital image into multiple segments
(sets of pixels, also known as super pixels). The goal of
segmentation is to simplify and/or change the representation
of an image into something that is more meaningful and easier
to analyze. Image segmentation is typically used to locate
objects and boundaries (lines, curves, etc.) in images. More
precisely, image segmentation is the process of assigning a
label to every pixel in an image such that pixels with the same
label share certain visual characteristics. Morphology is a Figure 9: Image with exudates
broad set of image processing operations that process images
based on shapes. Morphological operations apply a structuring

34
 International Journal of Computer Applications (0975 – 8887)
National Conference on Recent Trends in Computer Applications NCRTCA 2013

5. FEATURE EXTRACTION
In image processing, feature extraction is a special form
of dimensionality reduction. Feature extraction involves
simplifying the amount of resources required to describe a
large set of data accurately. Morphological features like
exudates area, exudates perimeter and the statistical feature;
the standard deviation are used in this work. Table 1 lists the
sample of the features used for the classification.

Table 1: Results of extracted feature values of normal and

abnormal images

Figure 11: Rule viewer for the detection of the diabetic

retinopathy

6. CLASSIFICATION
The features extracted from the retinal images are classified
using ANFIS. Input to the ANFIS is the morphological Figure 12: Surface viewer to diagnose diabetic retinopathy
features like exudate area and the exudate perimeter and the from the morphological and statistical retinal image data
statistical data includes the standard deviation. The output of
the ANFIS is 0 for the normal retinal images and 1, 2 and 3 7. CONCLUSION
for Background retinopathy, pre-proliferative retinopathy and This research work presents the use of state of art algorithms
proliferative retinopathy respectively. Epoch is set to be 1e-4. for the processing of fundus images, with particular attention
Out of the 100 samples, there are 32 Normal data, 13 images to the aspects related to diabetic retinopathy. The retinal
of Background DR, 10 images of Pre-proliferative and 45 image is preprocessed by green channel extraction, histogram
images of Proliferative eye disease. 75% data in each group is equalization and contrast enhancement. The optic disk is
selected for training and the remaining 25% data is used for eliminated by connected component analysis and features like
testing and validation. Gbell membership function is used in exudates area, size, color, and homogeneity and texture
the input side and Takagi Sugeno Constant membership properties are extracted. 100 data is collected and
function is used in the output side. Hybrid optimization is preprocessed. Morphological and statistical features are
used for the training and 300 epoches were selected for the extracted and classified using ANFIS. Normal data,
training. Figure 10 shows the architecture of the ANFIS used background DR and pre-proliferative data was identified with
for detecting the diabetic retinopathy. Figure 11 and 12 show 100% accuracy, the proliferative data alone was identified
the rule viewer and surface viewer used to show the results of with 97% accuracy.
the trained output.
8. REFERENCES
[1] B.Ramasubramanian and G.Mahendran,“An Efficient
Integrated Approach for the Detection of Exudates and
Diabetic Maculopathyin Colour fundus Images”,
Advanced Computing: An International Journal (ACIJ),
Vol.3, No.5, September 2012, DOI :
10.5121/acij.2012.3509 83
[2] Neelapala Anil Kumar, Mehar niranjan Pakki,
"Analyzing The Severity of The Diabetic Retinopathy
and Its Corresponding Treatment”, International Journal
of Soft Computing and Engineering (IJSCE) ISSN:
2231-2307, Volume-2, Issue-2, May 2012.
Figure 10: Architecture for the ANFIS based retinal [3] Priyanka Mishra, Suyash Agrawal,”Recognition of
diabetes detection Retinal Image Using Redial Basis Function”,
International Journal of Engineering Research and

35
 International Journal of Computer Applications (0975 – 8887)
National Conference on Recent Trends in Computer Applications NCRTCA 2013

Applications (IJERA) ISSN: 2248-9622, Vol. 2, Issue 2, Using Fundus Image”, IJART, Vol. 2, Issue 4, 2012,55-
Mar-Apr 2012. 59
[4] NilanjanDey, Anamitra Bardhan Roy,Moumita Pal, [10] Jaspreet Kaur , H.P.Sinha, “An efficient blood vessel
Achintya Das, “FCM Based Blood Vessel Segmentation detection algorithm for retinal images using local entropy
Method for Retinal Images”, , International Journal of thresholding”, International Journal of Engineering
Computer Science and Network (IJCSN) ,Volume 1, Research & Technology (IJERT).,Vol. 1 Issue 4, June –
Issue 3, June 2012 2012, ISSN: 2278-0181
[5] Jaspreet Kaur, Dr.H.P.Sinha, “Automated Localisation [11] Shilpa Joshi, P.T. Karule.” Retinal Blood Vessel
of Optic Disc and Macula from Fundus Images”, Volume Segmentation” ISSN: 2277-3754, International Journal
2, Issue 4, April 2012 , International Journal of of Engineering and Innovative Technology
Advanced Research in Computer Science and Software (IJEIT)Volume 1, Issue 3, March 2012.
Engineering.
[12] M. Ponni Bala, S. Vijayachitra, “Computerised Retinal
[6] Jaspreet Kaur, Dr. H.P. Sinha, “Automated Detection of Image Analysis to Detect and Quantify Exudates
Retinal Blood Vessels in Diabetic Retinopathy Using Associated with Diabetic Retinopathy”, Volume 54–
Gabor Filter,” IJCSNS International Journal of Computer No.2, September 2012
Science and Network Security, Vol.12 No.4, April 2012
[13] Romany Fouad Mansour.,“Using Genetic Algorithm for
[7] Kamala Devi. K, Agnes Anto, K. John Peter, “Curvelet Identification of Diabetic Retinal Exudates in Digital
Transform and Multi Structure Elements Morphology by Color Images”, Journal of Intelligent Learning Systems
Reconstruction based Retinal Image Analysis”, and Applications, April 2012
International Journal of Soft Computing and Engineering
(IJSCE) ISSN: 2231-2307, Volume-2, Issue-3, July 2012 [14] Jyoti Patil , A. L. Chaudhari ,“ Detection of Diabetic
548 Retinopathy Using Sobel edge detection method in DIP”,
International Journal of Scientific & Engineering
[8] Neelapala Anil Kumar, Mehar Niranjan Pakki., Research Volume 3, Issue 7, July-2012 1 ISSN 2229-
“Analyzing The Severity of The Diabetic Retinopathy 5518
and Its Corresponding Treatment”, International Journal
of Soft Computing and Engineering (IJSCE) ISSN: [15] Jyoti Patil, “ Development of Digital Image Processing
2231-2307, Volume-2, Issue-2, May 2012 using Fuzzy Gaussian Filter Tool for Diagnosis of Eye
Infection” , International Journal of Computer
[9] S. Nithyaa, S. Karthikeyen, “Identification of the Applications (0975 – 8887) Volume 51– No.19, August
Diabetic Retinopathy by Detecting New Retinal Vessel 2012

IJCATM : www.ijcaonline.org
36