Available via license: CC BY 3.0
Content may be subject to copyright.
Journal of Physics: Conference Series
PAPER • OPEN ACCESS
Diabetic Retinopathy Detection Using Automated
Segmentation Techniques
To cite this article: S. Prabha et al 2022 J. Phys.: Conf. Ser. 2325 012043
View the article online for updates and enhancements.
You may also like
A Data Mining Inspired Methodology
towards the Identification of Diabetic
Retinopathy
KG Pradeep Kumar, K Karunakara, GS
Thyagaraju et al.
-
Detection of diabetic retinopathy using
deep learning methodology
Gazala Mushtaq and Farheen Siddiqui
-
Detection of Diabetic Retinopathy using
Deep Mining
KG Pradeep Kumar, K Karunakara, GS
Thyagaraju et al.
-
This content was downloaded from IP address 93.180.239.70 on 02/09/2022 at 01:56
Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
1
Diabetic Retinopathy Detection Using Automated
Segmentation Techniques
S. Prabha1, S. Sasikumar1 and Ch. Leela Manikanta1*
1Department of ECE, Hindustan Institute of Technology and Science, Chennai, India.
sprabha@hindustanuniv.ac.in, ssasik@hindustanuniv.ac.in,
manimanikanta.ch@gmail.com
Abstract. This paper contains a brief discussion about Diabetic Retinopathy. As the name
indicates, it’s a medical complication present in diabetic patients which affects the retina.
Diabetic Retinopathy acronymed as DR is a medical circumstance where the high glucose
levels in the blood start affecting the blood vessels in the retina. The paper discusses the non-
invasive technical method to detect diabetic retinopathy involving various algorithms in every
phase of the process. The input fundus images are taken from STARE Database. The
methodology conveyed in this paper involves contrast-limited adaptive histogram equalization
for noise cancellation purposes and enhancing the base contrast of the image. The
Segmentation consists of 2 steps and the first step consists of the Fuzzy C-Means clustering
primarily to find the coarse vessels present in the retina. Additionally, the Region-based active
contour is used to select the region of interest which is to highlight the blood vessels. As a
result, Our proposed segmentation method extracts the blood vessels accurately, resulting in
the similarity measure value of 85%. Furthermore, these segmented retinal blood vessels are
given as the input to CNN classifiers in order to detect Diabetic Retinopathy. For our proposed
method, an overall accuracy to detect DR was 92%. This methodology can be used for mass
screening processes in the field of ophthalmology.
Keywords: Diabetic Retinopathy, Contrast-Limited Adaptive Histogram Equalization
(CLAHE), Segmentation, Fuzzy C Means, Region based active contour, retinal images.
1. INTRODUCTION
Human eyes are commonly known as the organ of vision. It’s the organ that initially detects light and
transfers the signals through the optic nerves to the brain in order to perceive the visionary image.
Eyes are an essential part of the human body as a common fact. It's also a highly sensitive organ, with
23 neurons that transform light into electrochemical impulses. Sclera, Iris, Cornea, Pupil, Lens, Ciliary
Body and Muscle, Conjunctiva, Retina, Vitreous Body, Optic Nerve, Macula, and Retinal Blood
Vessels are the components of the eyes. These are the essential and common parts of the eyes and can
be further segmented if each organ is studied specifically. As it is known that the human body is an
interconnected system that is connected and functions by relying on other organs, the same it’s an
exclusive feature that the nerves in the eyes are to other organs of the body as well. The retina is a
fragile portion of the eye that is located in the base of the eye and is the innermost, light-sensitive
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
2
layer. The retina's job is to give the sharp, centre vision needed for tasks like reading, driving, and
anything else that requires minute concentration.
The world is evolving towards something new on a daily basis, leading to advancements in every
sector be it technical, business or even the overall life style of human beings. Along with these
advancements, an essential sector with regular advancements is the medical sector. There are
researches, inventions and discoveries of new medical circumstances often. As a result, the medical
world and advanced technology work hand-in-hand to resolve problems in an effective manner. The
same way diabetes is a medical condition wherein the sugar levels present in the blood are higher than
required. It is a disease that can be found in human beings of all ages certainly, including new born
babies.
Diabetes is further 48 divided into Prediabetes, Type 1 Diabetes, and Type 2 Diabetes. Prediabetes is a
phase wherein the glucose present in the blood is slightly higher than the amount needed.
Classification of Type 1 diabetes is as production of minute or nill amount of insulin by pancreas.
When the body's ability to metabolize, blood sugar is hampered by excessive amounts of glucose,
Type 2 Diabetes develops. Patients with diabetes commonly face other problems like foot ulcers,
cardiovascular complications, gum diseases, fatigue, stroke, nerve damage, chest pain, vision
complications. Diabetic Retinopathy is a familiar stage of diabetes. It is a medical disorder in which
excessive blood sugar levels impact the eyes in a number of ways. And one such condition is Diabetic
Retinopathy also known as DR. It is a medical condition wherein high blood sugar levels begin to
harm the blood vessels in the eyes, creating serious problems. The most prevalent concerns that
individuals with Diabetic Retinopathy may have include vessel obstruction or leakage, as well as the
initiation of new blood vessel growth in the Retina. When one end of a blood vessel becomes clogged
due to excessive sugar levels, a blockage occurs. And this further tends to lead to the next
complication of leakages of the blood vessels when there is continuous supply of blood to the vessel
and when there is a blockage, the blood cannot be present within the closed space, hence leading to the
leakage of blood as a result of the pressure. Sometimes there can even be new blood vessels formed. In
the short or long run, these conditions may result in various issues like blurred vision, impaired color
vision, eye floaters, etc., which restricts vision, poor night vision, dark or empty spots in vision.
There are 4 stages in Diabetic Retinopathy, and they are:
Stage 1- Mild Non-proliferative Retinopathy: In this stage, microaneurysms start to show
up. Microaneurysm is the swelling of blood vessels located in the retina. The blood vessels are in the
initial stage of bulging at this point of time.
Stage 2- Moderate Non-proliferative Retinopathy: The swelling of the blood vessels starts to increase.
This further starts to severely affect the blood flow in the retina. And this also prevents the proper
nourishment required in the eyes.
Stage 3- Severe Non-proliferative Retinopathy: In this stage the blockages in the blood vessels start
taking place in large amounts. This significantly reduces the amount of blood flow in the thin areas.
Stage 4- Proliferative Retinopathy: This is the critical stage where the birth of new blood vessels
occur, causing severe complications in the delicate areas.
These four stages of DR have their own complications and treatments for cure but this can be done
only if the problem can be detected accurately. In order to prevent all these complications, it is advised
for all the patients with diabetes Type 1 and Type 2 to take a DR test. Nevertheless, not every patient
agrees to do so as there are various factors involved in taking the DR test medically. The medical DR
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
3
tests are extremely time consuming they require the physical contact in the eyes which is a method
prone to infections if done without proper sanitization, and they are also high in cost. That is why there
are various technical methods proposed and carried out with computational algorithms to detect
DR. The engineers’ methods so far are extremely time effective so that treatment can be started out in
an early stage. They are non-invasive, meaning there is no physical contact required so safety is
always insured. And finally, they are also cost effective. Another added advantage of the engineer’s
methods is that the stages of DR can also be configured so that the diagnosis can be done accordingly.
2. RELATED WORKS
Using robust hybrid features Aslani [1] demonstrated a new supervised retinal vessel segmentation
method They took the hybrid feature vector to merge the supporting complementary local data. There
are thirteen Gabor filter responses that make up the 17D feature vector used. It is also noteworthy that
as the numbers are reduced, the accuracy also decreases. The proposed method tends to be designed
and executed to achieve the highest average accuracy in the DRIVE and STARE datasets.
Furthermore, the progress and scope of the development were discussed. Azzopardi [2] discusses
trainable COSFIRE filters for vascular description with attributes of retinal imaging. The dominant
feature here is that the COSFIRE filter has been structured. B in this case means bar, which is a
construction for a container. The implementation of the filter named COSFIRE depends on the
approximate values of required parameters. Christodoulidis[3] Multiscale Tensor Voting Framework
(MTVF) is the reference for the methodology proposed in this paper. It is added up with multi-scale
line detection to get rid of line detection limitations when managing smaller vessels. Although the
proposed methodology has produced excellent results, there were some issues that needed to be
explored and addressed. If the MTVF is attached to the false positive (FP) vessel-like structure of the
main vasculature, the main vessel may be lost.
Zhu[4] The document ultimately aims to improvise the standards of the segmented image. Using State-
of-the-art technology they enhanced 39 discriminating feature vectors for fundus imaging. The binary
retinal vessel segmentation is the acquired output by the classifier. As a result, the diagnosis of retinal
diseases is difficult. Therefore, it is crucial to use the right image segmentation algorithms to
accurately find all retinal blood vessels. Geetha[5] proposed his work based on the Principal
component analysis (PCA) that was used to create the feature vector in this paper. Additionally, k-
mean clustering is applied to this result to group the pixels into cup or non-cup groups. Vessel group
classification improved accuracy but reduced sensitivity in some cases. Hassanien[6] In this article, the
automated approach to segmentation is based on two-step optimization principles; retinal blood
vessels were created. was demonstrated. The suggested method combines artificial bee colony
optimization with a fitness function for fuzzy cluster compression with partial membership at first
level and approximate vasculature at aggregation. Roy Chowdary[7] conducted the research based on
blood vessel segmentation from fundus photographs by extracting major vessels and classifying sub-
images. In this document, a set of eight features was introduced to distinguish between vascular and
nonvascular pixels. The fine pixels of the vessels are found with both Gaussian and GMM classifiers.
Sil Kar [8] discussed this research study to track the numerous thresholds, they used their
understanding of the maximum response of the matched filter and the fuzzy conditional entropy. And
differential evolution calculations are used to find the ideal values. The discussed procedure is also
effective in removing the irregular and thin blood vessels in pathological retinal images.
Sreejini [9], this article discusses a PSO-based parameter that was introduced to determine the exact
values of the multi-scale coordinated filter parameters. The results generated by the multiscale
matching filter take precedence over the single scale matching filter for vessel
segmentation. Singh[10] this work analyzes the methodology for this particular experiment consisting
of PCA or principal component analysis during the pre-processing phase. In the next stage, CLAHE is
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
4
employed in order to improve the retinal picture. To obtain the segmented image, after running the
proposed combined filter, the post-processing phases include optimal entropy-based threshold and
length filtering. The results suggest that this method performs relatively well and efficiently compared
to other prominent Gaussian distribution functions. Saranya[11] The Convolution Neural Network
(CNN) is the predominant methodology utilized to observe and classify the images according to the
four phases of DR as it is capable of managing the pre-processing of images and normalization as
well. Chaudhary[12] The proposed methodology contains the raw fundus images which are carried out
for processing for sound reduction and further changed into gray image. Later the segmentation of the
optic disc and the retinal nerves take place. The extracted features will be used for the classification.
Post processing consists of erosion and dilation using a structuring element on the pre-processed
images. The CNN and the Fuzzy classifier are used to identify the status of the image. This will
confirm the presence of DR or a normal eye condition.
Imran Qureshi[14], demonstrated the computer assisted diagnosis (CAD) systems for treating
Diagnosis Retinopathy. It’s a combined study on the effectiveness of the different methodologies used
to detect DR. The segmentation and the method of extraction using the image combination, pre-
processing, improvement and segmentation methods have been briefly explained in the paper [18].
Akinboro[15] The project is carried out specifically to classify the stage of DR with the proposed
approach. The methodology chosen is by first generating the digital fundus image through the
convolution neural network CNN and 3662 datasets. The CNN model is built with the Keras infused.
Author [13][16][17] did the research work on threshold based on an iterative algorithm that has been
implemented in the research paper focusing on the accuracy of the segmented images. The proposed
methodology is made up of feature extraction, followed by image segmentation, image pre-processing,
and finally the classification based on CNN classifier which provides the results of whether DR is
present. The proposed methodology is designed in such a way to detect diabetic retinopathy in an
earlier stage in order to take appropriate diagnosis. Accuracy, precision, specificity, and the sensitivity
rates of this proposed system stays above 85% in all the four aspects. The overall procedure focuses on
maintaining the accuracy rates of the results and every algorithm chosen, carefully aids to it.
3. PROPOSED WORK
Figure 1: Proposed work architecture
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
5
The innovative technique flows through the following six phases as shown in figure 1.
1. The proposed method uses a fundus camera as an input image (STARE database).
2. For the first step of pre-processing, the input image is transformed to the green channel. The picture
is then subjected to CLAHE to boost the image's local contrast.
3. Fuzzy C-Means (FCM) is utilized to detect the coarse vasculature during the segmentation phase.
4. The region based active contour method is used to select the region of interest & aids in highlighting
the blood vessels.
5. The segmented retinal blood vessels are seen in the output image.
6. The segmented image is given as an input image to the deep learning classifier to find whether the
respected image is affected with DR or NOT.
The process of detecting DR will first be started out by first acquiring the retinal fundus image from
the STARE Database. STARE stands for Structural Analysis of the Retina. In an image, second
channel is the green channel, that second channel alone is extracted to correct non uniform
illumination. The pre-processing will be done through the CLAHE algorithm. This algorithm is
utilized to improve the image's contrast level and for noise cancellation characteristics. As a result, the
image is additionally enhanced so that the unseen features are also visible. The CLAHE has three
parameters: block size, histogram bins and max slope through these three methods the image will
undergo pre-processing. Then the optimization stage takes place in two steps. The first stage
of optimization includes the involvement of Fuzzy-C Means clustering to specifically spot the coarse
vessels present in the retina. Later the region of interest is selected utilizing the region based active
contour algorithm for better quality and higher accuracy levels of the blood vessels that exist in the
particular image. Now the output image will show the segmented retinal blood vessels which will be
sent as an input image through the CNN classifier and this is where the presence of diabetic
retinopathy is observed, reviewed and confirmed.
Resulted in the creation of the Designed analysis of the Retina database (STARE). The STARE
database photos were taken using a 35degree narrow field view camera and have a resolution of
700*605 pixels. Retinal imaging is a frequent procedure performed during an eye examination. To the
posterior inferior surface of the eyeball in order to see the pupil of the eye an optical camera is
assisted. The retinal layer, the optic nerve, the fovea, and the surrounding arteries are all imaged. The
ophthalmologist can then use this image as a reference point for evaluating any findings or to examine
the eye condition.
AHE (adaptive histogram) is a computer image processing approach for enhancing image contrast.
Traditional histogram equalization differs from the adaptive technique. It's designed to build a number
of histograms, each corresponding to a different portion of the image, and then use them to disperse
the image's brightness values. This makes it great for increasing local contrast and sharpening edge
sharpness in different areas of a photo. AHE also tends to exaggerate noise in relatively similar areas
of an image. By limiting gain, a variant of AHE known as CLAHE was developed to minimize or
prevent noise. Microaneurysm pixels were enhanced with the CLAHE filter. The CLAHE filter
produces adequate vein enhancement while removing excess noise. Other global enhancement
methods such as traditional contrast stretching and universal histogram equalization work less
efficiently compared to the CLAHE filter. The use of an independent elemental analysis is the second
way to increase the contrast of the veins. Various image enhancement techniques have been added in
recent research at the pre-processing stage, including grayscale conversion of image, spectrum
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
6
normalization, brightness enhancement using CLAHE, and gamma modification to improve overall
appearance. The CLAHE histogram has been clipped to a standard clipping value to avoid excessive
contrast enhancement that could result in an image with an odd appearance and unwanted detail. Also,
incorrect approaches to implementing contrast enhancement could result in poor appearance in
missing areas, such as small veins lead. By summing up a global threshold to the fixed point of
contrast the above consequence could be solved. Edge detection is a technique for detecting frequent
changes in pixel values. The brightness level of neighbouring pixels is used to examine the edge
information of a given target pixel. Without a clear difference in brightness, there can be no image
edge.
The FCM (Fuzzy C-Means) clustering technique allows a single data item to be assigned to two or
more clusters. Pattern Recognition also runs on the basis of the above Strategy. This algorithm assigns
the link to each data point corresponding to each group centre based on the distance between the group
and the data point. The closer the data is to the cluster centre, the more likely it is to belong to that
cluster centre. The iterative unsupervised Fuzzy C Means (FCM) method is the most extensively used
clustering technique for picture segmentation. Pattern recognition and clustering are used to analyze
and segment most medical images. Compared to the k-Means algorithm, fuzzy C-mean clustering is
considered a better option. The fuzzy C-means algorithm, unlike the k-means technique, which
requires data points to belong to only one group, allows data points to belong to many groups with
probability.
The fuzzy C means that the clustering process starts first by outputting the input image first. Then the
size of the image is retrieved. The possible range is then calculated. In addition, the possible number
of iterations is identified. Then the given dimension of the images is calculated. Then the iteration
begins. When the maximum number of iterations is reached, the procedure is repeated, and the picture
is submitted for iteration once more until the maximum number of iterations is achieved. This is the
complete procedure of the fuzzy pooling process C.
Active Contouring is a Segmentation approach that separates relevant pixels from an image for further
processing and analysis using forces and power constraints. An active contour is a model used in the
segmentation process. Contours are the lines that define the area of interest in an image. There’s a
need of converting segmented images into multiple similar images to get represented in the region-
based technique. The basic idea behind active contour models, sometimes referred to as snakes, is to
generate a curve to group objects in a given image while constrained by these constraints. The curve
approaches its interior normal and must end at the edge of the object. In classic active snake and
contour models, To come to a halt, an edge detector employs the curve evolution at the boundary of
the target element based on the gradient of the image. The following steps are used to evacuate the
papilla: i. To capture the retinal image mask, first select the underlying raw, apply the region-based
active contour to the segmented image mask. I. The mask from the input picture is removed, resulting
in a new mask. ii. The segmented image without a mask is the result of subtracting the new mask from
the previous one. The region based active contour methos uses statistical information to construct a
region stopping function which stops the contour evaluation between different regions. This method is
not sensitive to initialization of level set function and can recognition object boundaries efficiently.
The Structural Similarity Indexing Method (SSIM) is a tool that compares the structural similarity of
two photographs. If the second image is of perfect quality, the SSIM index can be regarded as a
quality indication for one of the photos under consideration. It is an improved version of the first
proposed Global Image Quality Index. The SSIM tool is used in this case to compare the actual terrain
image with the segmented image. The relevant output image for the fundus image retrieved from the
STARE database is the actual image of the floor. And the segmented image is the starting picture that
is received after the CNN after the final processing of the image. The SSIM tool will be used to
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
7
compare the differences present in the ground truth image and the segmented image which will prove
the accuracy and the efficiency of the methodology. Keras was used to build and compile the
Convolutional Neural Network (CNN) models. Except for their last layer, both models have the same
architecture; the first model, which was used to classify the severity of DR, has 5 outputs, while the
other model, which was used to identify DR, has 2. The models are divided into four stages:
i. Input
ii. Feature Learning
iii. Classification
iv. Output
The retinal fundus image is the CNN model's input, which was read by the computer as a matrix of 64
* 64 * 3 pixels, that represents the height, width, and dimension. The RGB values are represented by
the three dimensions of the images. The input phase's outcome is subsequently passed on to the feature
learning phase. The input image is passed through a series of convolution layers with filters in the
feature learning phase, and the model extracts and learns features about the images. The extracted
features are then forwarded to the classification phase, in which the image is categorized with
probabilistic values 26 between 0 and 1. The pre-processed image is fed into the input layers, which
produce an array of pixels as output, which is then transferred to the first convolution layer. The link
between/among of the pixels exists by learning visual attributes of minute squares and convolutions.
The first convolution layer uses a 64 * 64 * 3 picture as input and produces an output of 62 * 62 * 128
feature map with relu activation.
The output was fed into the second convolution layer as input. The second convolution layer, which
likewise uses a (3 * 3) filter matrix with relu activation, produced a (60 * 60 * 128) output that was
transferred to the pooling layer. By picking the largest element from the feature map, the max pooling
layer minimizes the feature map's dimensionality, producing an output of size (30 * 30 * 128) and
passing it to the dropout layer. Over fitting was reduced by using the dropout layer. During the
training, it switches off some neurons randomly. The output of the dropout is flattened using a flatten
layer, which transforms the entire feature map into a single column of size (115200). The dense layer
contains densely linked neurons and gets input from the prior layer's neurons. To reduce overfitting, a
dropout layer was added between the two thick layers. The last dense layer employed the SoftMax
activation function to assign probabilistic values between 0 and 1 to each class in the image.
4. RESULTS OBTAINED
4.1 Normal Images:
The normal images refer to the pictures of healthy eyes received from the STARE database. These
images do not contain or have the symptoms of DR. The retina will be clear in this image with normal
and regular functioning blood vessels without any complications or abnormalities. Images considered
to be normal will provide the image of the healthy retina when the image is fully processed.
4.2 Abnormal Images:
Abnormalities refer to images that contain Diabetic Retinopathy. These images will commonly show
up as abnormalities while retrieving from the STARE database. The images classified as abnormalities
will contain the image of the affected retina where the blood vessels may have blockages or leakages.
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
8
In other cases, there might also be new blood vessels formed or growing depending on the severity of
the stage as shown in figure 2.
Figure 2. Describes the Normal and the Diabetic Retinopathy Image.
4.3 CLAHE Images:
As we know that having the perfect contrast to an image will give the best results. To obtain best
results we are using Contrast Limited adaptive histogram equalization, The contrast to an image
shouldn’t be low meanwhile it shouldn’t be high. So, it should be in the limit. To maintain that certain
contrast limit to an image we are using CLAHE.
Figure 3. Describes the Normal and the Diabetic Retinopathy Image after pre-processing.
Here, by using the CLAHE we can clearly observe the Blood vessels in the image compared to the
input image as shown in figure 3.
4.4. Segmented Images:
Segmented images are the images received after the raw fundus image goes through the whole
segmentation process. This is the image that will be received from the Segmentation process. The
image of the blood vessels present in the retina can be seen in this image. The segmented image can
also be considered as the proof for the presence of Diabetic Retinopathy for the particular processed
image as shown in figure 4.
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
9
Figure 4. Describes the Normal and the Diabetic Retinopathy Image after segmentation.
4.5 Ground truth Images:
These images are the segmented Outputs of the STARE project. Ground truth is data that’s famed to
be real or true, provided by direct observation and measurement as shown in figure 5.
Figure 5. Describes the Normal and the Diabetic Retinopathy Image of ground truth images.
4.6 SSIM
SSIM stands for structural similarity index measure. The Structural Similarities Indexing Method
(SSIM) is a technique for assessing structural similarity between two photographs. If one of the photos
being compared is of perfect quality, the SSIM index can be considered as a quality indicator for that
image. It's an improved version of the global picture quality index that was first suggested. In the
present scenario the SSIM tool will be used to compare the ground truth image and the segmented
image. The ground truth image is the respective output image for the fundus image that is retrieved
from the STARE Database. And the segmented image is the output image received from the CNN
after the final processing of the image. The SSIM tool will be used to compare the differences present
in the ground truth image and the segmented image which will prove the accuracy and the efficiency
of the methodology.
Our proposed segmentation method extracts the blood vessels of each image accurately, resulting in
the similarity measure value of 85%. The primary reason for the high accuracy rates is the
combination of the Fuzzy C-means clustering and the Region based active contour algorithms.
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
10
4.7 Classification:
S. No
Input
Image
CLAHE
Image
Segmented
Image
Ground truth
Image
Accuracy
Classification
1
0.96
Affecting with
DR
2
0.93
Not Affecting
with DR
3
0.92
Not Affecting
with DR
4
0.95
Affecting with
DR
Table 1: Classification
The output images are the segmented images. The structural information in the segmented image and
the ground truth image are compared pixel by pixel using SSIM metrics. If there is a very good
similarity between segmented and ground truth image, it displays the value between 0.9 to 1. Those
segmented images are given as the input images to the CNN classifier to classify whether the image is
affected with diabetic retinopathy or not.
Total we have 88 segmented images. All these 88 images are divided into two stages, one is training
dataset and another stage is testing dataset. In that 88 images, 67 images are given to train the
classifier and the remaining 21 images are given to test the classifier. Each stage is divided into 2
classes, class0 and class1. Class0 represents the normal segmented images which are not affected with
Diabetic retinopathy. Class1 represents the segmented images affected with diabetic retinopathy.
The classifier will undergo the classification process with the help of training dataset and testing
dataset to provide the accurate result. For our proposed method, an overall accuracy to detect DR was
92%.
4.8 Performance measure:
The performance measure of our proposed method includes sensitivity, specificity and accuracy are
shown in the figure.6. The primary reason for the high accuracy rates is the combination of the Fuzzy
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
11
C-means clustering and the Region based active contour algorithms along with CNN classifier. The
visual representation of the performance measure has been provided in figure 6.
Figure 6. Shows the Performance Measure
5. CONCLUSION
In order to determine the severity level of DR, it is critical to accurately identify the retinal blood
vessels during the ophthalmological examination. Many algorithms aren't capable of differentiating the
depigmented abnormal retinal images from the retinal blood vessels. The research findings have
emerged from each of the STARE database's 88 photos. The results indicate that the proposed
methodology consisting of FCM clustering along with Region based active contour accurately
identifies all the blood vessels. Both these algorithms aid in increasing the accuracy rate of the
segmented blood vessels. There are no discontinuities between the minor vessels and they are also
identified in this process.
CLAHE minimizes the level of noise in depigmented retinal images. By examining the segmented
vessel structure through the proposed method, it is evident that it is capable of minimizing the
ophthalmologists' effort in analyzing the diabetic retinopathy affected blood vessels. The proposed
retinal blood vessels segmentation approaches can be used for datasets with similar attributes. Our
proposed segmentation method extracts the blood vessels of each image accurately, resulting in the
similarity measure value of 85%. For our proposed method, an overall accuracy to detect DR was
92%.
6. FUTURE SCOPE
In the long-run, the proposed methodology can be implemented in the field of ophthalmology for
retinal screening if it is further developed. There is still room for advancements in the present
procedure wherein there can be progressive algorithms designed to be involved in order to classify the
stage of Diabetic Retinopathy. If this advancement is added in this methodology, it will be a procedure
that will specifically convey the stage of DR the patient is currently in so that appropriate diagnosis or
treatment can be started at the earliest. Once the methodology is developed with the classification, the
procedure can be implemented for ophthalmological campaigns carried out especially in rural areas.
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
12
As the method discussed is time efficient, it can be introduced in campaigns and other screening areas
as the retinal examination can be done in a short span of time and the results can also be retrieved
soon.
7. REFERENCES
[1] Aslani, Shahab, and Haldun Sarnel. "A new supervised retinal vessel segmentation method based
on robust hybrid features." Biomedical Signal Processing and Control 30 (2016): 1-12.
[2] Azzopardi, George, et al. "Trainable COSFIRE filters for vessel delineation with application to
retinal images." Medical image analysis 19.1 (2015): 46-57.
[3] Christodoulidis, Argyrios, et al. "A multi-scale tensor voting approach for small retinal vessel
segmentation in high resolution fundus images." Computerized Medical Imaging and Graphics 52
(2016): 28-43.
[4] Zhu, Chengzhang, et al. "Retinal vessel segmentation in color fundus images using an extreme
learning machine." Computerized Medical Imaging and Graphics 55 (2017): 68-77.
[5] GeethaRamani, R., and Lakshmi Balasubramanian. "Retinal blood vessel segmentation employing
image processing and data mining techniques for computerized retinal image analysis." Biocybernetics
and Biomedical Engineering 36.1 (2016): 102-118.
[6] Hassanien, Aboul Ella, Eid Emary, and Hossam M. Zawbaa. "Retinal blood vessel localization
approach based on bee colony swarm optimization, fuzzy c-means and pattern search." Journal of
Visual Communication and Image Representation 31 (2015): 186-196.
[7] Roychowdhury, Sohini, Dara D. Koozekanani, and Keshab K. Parhi. "Blood vessel segmentation
of fundus images by major vessel extraction and subimage classification." IEEE journal of biomedical
and health informatics 19.3 (2014): 1118-1128.
[8] Kar, Sudeshna Sil, and Santi P. Maity. "Retinal blood vessel extraction using tunable bandpass
filter and fuzzy conditional entropy." Computer methods and programs in biomedicine 133 (2016):
111-132.
[9] Sreejini, K. S., and V. K. Govindan. "Improved multiscale matched filter for retina vessel
segmentation using PSO algorithm." Egyptian Informatics Journal 16.3 (2015): 253-260.
[10] Singh, Nagendra Pratap, and Rajeev Srivastava. "Retinal blood vessels segmentation by using
Gumbel probability distribution function based matched filter." Computer methods and programs in
biomedicine 129 (2016): 40-50.
[11] P.Saranya,K.M.UmaMaheswari(2021), “Classification of Different Stages of Diabetric
Retinopathy using Convolutional Neural Networks'', 2nd International Conference on Computation,
Automation and Knowledge Management (ICCA KM), pp 59-64.
[12] Chaudhary, Sonali, and H. R. Ramya. "Detection of Diabetic Retinopathy using Machine
Learning Algorithm." 2020 IEEE International Conference for Innovation in Technology (INOCON).
IEEE, 2020.
International Conference on Electronic Circuits and Signalling Technologies
Journal of Physics: Conference Series 2325 (2022) 012043
IOP Publishing
doi:10.1088/1742-6596/2325/1/012043
13
[13] Prabha, S., et al. "Machine Learning Technique for Analysing the Diabetic Retinopathy." 2022
International Conference on Sustainable Computing and Data Communication Systems (ICSCDS).
IEEE, 2022.
[14] Qureshi, Imran, Jun Ma, and Qaisar Abbas. "Recent development on detection methods for the
diagnosis of diabetic retinopathy." Symmetry 11.6 (2019): 749.
[15] Fayemiwo, M. A., S. A. Akinboro, and O. A. Adepegba. "Identification and Classification of
Diabetic Retinopathy Using Machine Learning." Adeleke University Journal of Engineering and
Technology 1.1 (2018): 245-259.
[16] Prabha, S., et al. "Retinal Image Analysis Using Machine Learning." 2020 International
Conference on Communication and Signal Processing (ICCSP). IEEE, 2020.
[17] Bindhu, V., and Villankurichi Saravanampatti PO. ""Semi-Automated Segmentation Scheme for
Computerized Axial Tomography Images of Esophageal Tumors."" Journal of Innovative Image
Processing (JIIP) 2, no. 02 (2020): 110-120.
[18] Rao, Smaran S., Gajanan Maske, and Antara Roy Choudhury. ""Iris Image Segmentation and
Localization using Dynamic Reconfigurable Processor."" Journal of Innovative Image Processing 2,
no. 3 (2020): 147-155.
