IRJET- Automatic Detection of Diabetic Retinopathy using R-CNN

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 05 | May 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 5595
Automatic detection of Diabetic Retinopathy using R-CNN
Athira T R1, Athira Sivadas2, Aleena George3 , Amala Paul4 , Neethu Radha Gopan5
1,2,3,4Student, Dept. of Electronics and Communication Engineering, Rajagiri School of Engineering and
Technology, Cochin, Kerala, India
5Assistant Professor, Dept. of Electronics and Communication Engineering, Rajagiri School of Engineering and
Technology, Cochin, Kerala, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - Diabetic Retinopathy (DR) is an eye
abnormality caused as a result of long term diabetes. As the
disease progresses it leads to distortion and blurred vision.
The diagnosis of DR using color fundus image requires
skilled clinicians to identify the presence of critical features
which makes this a difficult and time consuming task. In our
paper we propose a R-CNN (Regional Convolutional Neural
Network) approach to diagnose DR from digital fundus
images. In our research we implemented a new approach
were the whole image was segmented and only the regions
of interest were taken for further processing. In our method
we have used 10 layers for R-CNN, trained it on 130 fundus
images and tested on 110 images. All the images were
classified into two groups i.e., with DR and without DR. This
R-CNN (Regional CNN) approach was found to be efficient
in terms of speed and accuracy. An accuracy of
approximately 93.8% was obtained from R-CNN.
Key Words: Convolutional neural network, Regional
convolutional neural network, Diabetic retinopathy.
1. INTRODUCTION
Diabetes has now become a worldwide disease which
ultimately leads to complete vision loss. Diabetic
Retinopathy is a complication of type 2 diabetes.
According to the IAPB (International Agency for the
Prevention of Blindness) [1] report published in 2017,
there were 422 million people diagnosed with diabetes. 1
in 3 people diagnosed with diabetes will have diabetic
retinopathy up to a certain degree and 1 in 10 people will
suffer from vision loss. DR results in the damage of blood
vessels in the retinal layer of the eye. It forms
microaneurysms due to the focal dilation of weakened
walls [2]. The capillaries may become leaky forming
yellow white flecks which are commonly referred to as
exudates.
The main issue with DR is that it doesn’t usually cause
sight loss until it has reached the advanced stage. Due to
the lack of any significant symptoms normal DR screening
will only help the patients with high risk of progression. In
order to identify DR we analyze the fundus images for the
lesions and exudates. Normal approach for diagnosing DR
is time consuming and requires experienced clinicians to
identify critical features from the fundus images.
An automatic method for detection of diabetic retinopathy
would help people with diabetes to recognize the
symptoms at its earlier stage. It can greatly reduce the
clinical burden on retina specialists. This also helps to
monitor the dynamics of the lesions. Countries with huge
population like India, China, Indonesia and Bangladesh
contributes to 45% of the global burden in diabetes [2].
Since the counts are expected to move up, an automatic
clinical detection would be of much help.
Fig-1. (a) A normal fundus image with no sign of DR,
(b)Fundus image with red lesions, (c) Fundus image
with exudates.
The study for automatic detection of DR becomes more
and more crucial in the past few years.
In our study we are focusing on anomalies in the retina in
the form of exudates and red lesions. Due to the similar
color characteristics of red lesions with the retinal blood
vessels it is hard to locate these lesions using normal
image processing techniques. Considerable work has been
done in blood vessel extraction and optic disc removal,
both of which may also result in a false detection.
Techniques of morphological closing and opening can be
used for the removal of structures that may result in false
detection, but they may still produce residues [3]. SVM
classifier can be used to discriminate retinal images with
bright lesions to detect DR [4]-[5].

The short comings of the above mentioned methods can
be overcome through CNN and R-CNN techniques. CNN
comes under the class of deep neural networks whose
hidden layers include several convolution layers. CNN has
the ability to learn the features during its training phase.
At different layers of the network various features with
varying complexities can be learned. There is no need for
a manual feature extraction since it automatically extracts
features while passing through each of the layers.
The hidden layer may include ReLU (Rectified linear unit),
which is an activation function, convolutional layer, fully
connected layer, softmax layer etc.
2. METHODOLOGY
CNN has been widely recognized for applications such as
image processing, pattern recognition and video
recognition. CNN in image classification takes an image as
the input and classify it into the appropriate category. It
has a number of hidden layers in which convolution is
done to extract features and other valuable information
from the image. The output is obtained from the
classification layer [8].
In R-CNN, the image is segmented into many regions (or
segments) and the CNN is compelled to focus on these
segments. The accuracy of object detection is very high
compared to that of CNN due to extraction of region of
interest [12].
Fig-2. Block level representation of the methodology
2.1. Preprocessing
Initially the original fundus images are resized to a
dimension of 336 x 448. Due to the immense information
and varying contrast of images obtained from the fundus
cameras preprocessing is necessary. Without
preprocessing the images suffer from vignetting effects
and image distortion [7]. Since the images are obtained
from different fundus cameras they will be having non
uniform illumination, illumination normalization
techniques have to be incorporated [6]- [7].
2.1.1 Green channel extraction
Every image consists of three channels namely red, green
and blue. Green channel is extracted from the fundus
image. Extraction of green channel provides better
contrast between maximum and minimum intensities in
an image. It has less noise compared to that of red or blue
channels. Blue channel is not normally considered because
it provides less contrast and does not contain much
information.
2.1.2. Adaptive Histogram Equalization
This image processing technique improves the contrast of
the image. It is usually performed on small regions of the
image called tiles. This helps in enhancing the edges of the
image.
2.2. Optic disc removal
Automatic Optic Disc (OD) removal is a vital step because
optic disc act as false positives as they greatly resemble
exudates in intensity. Firstly Images are converted into
grayscale whose intensity values ranges between 0 to 255
and then thresholding is done. The pixels that have
intensities lower than the threshold were converted to 0
(black), while pixels above the threshold were converted
to 255 (white). the dilate-erode combinations are used to
remove unconnected small regions while preserving large
unconnected regions. This provides an initial region from
which the segmentation process begins. A disk structuring
element of size 30 is applied over the image and an OFF
pixel is set to ON if any of the structuring elements overlap
ON pixels of the image. The threshold is continuously
updated and this process is repeated until only one region
remains.
2.3. Extraction of region proposals
Region of interest is extracted using 2 methods by means
of thresholding and blocking (Pixel based) [14]. Based on
intensity differences the region of interests (i.e., those
segments which contain lesions) are extracted. For further
processing only these segments are considered.

2.3.1 Threshold based
Regions are extracted from the preprocessed image by
thresholding the intensity values of red lesions and
exudates separately [13]. For thresholding, lab color space
is used since it is the most accurate way of representing
colors. It is a 3 axis color space in which L represents
lightness, a & b represents color dimensions.
Fig-3. Thresholing
2.3.2 Pixel based
Images are divided into blocks equal segments of size 128
x 128 pixels and the blocks which are suspected to have
signs of DR are extracted. Only these extracted blocks are
considered for training
Fig-4. Pixel based blocking
2.4 CNN classification
The number of input layer neurons is equal to the number
of pixels in the input image. Convolutional layer makes use
of the convolutional features and computes the product
between the image patches and the filter. For the
activation layer ReLU(Rectified Linear Unit) can be used.
ReLU layer perform a threshold operation to each element
of the input where any value less than zero is set to zero.
Element wise activation is applied to the output of the
convolutional layer. The function of the pooling layer is to
down convert the volume so as to make the computation
faster and to reduce the memory requirement. The
convolutional layer that is present just before the fully
connected layer holds the information about the features
including the edges, contrast, blobs and shapes. And the
fully connected layer contains the aggregated information
from the previous layers. A softmax layer is used as the
final layer of CNN[9]. It assigns decimal probabilities to
each class. The output layer classifies the input images
into 2 classes namely normal (without DR) and abnormal
(with DR). Fully connected layer takes the output of all the
neurons from the former layer.
The designed layers are used for training and testing of
images [10]-[11]. The images for training can be obtained
from IDRiD (Indian Diabetic Retinopathy Image Dataset)
which provides marked images of Diabetic lesions.
In our method we have used 10 layers for R-CNN, trained
it on 130 fundus images and tested on 110 images. All the
images were classified into two groups i.e., with DR and
without DR.
Fig-5. CNN layers

3. PERFORMANCE PARAMETERS
The parameters that are used to evaluate the performance
of the proposed model are accuracy, sensitivity,
specificity, False positive rate and precision.
If P is the number of positive samples (patient with DR), N
is the number of negative samples (patient without DR).
True Positive (TP) is the number of patients with DR
which is correctly classified. False Positive (FP) is the
number of patients without DR which is incorrectly
classified. True Negative (TN) is the number of patients
without DR which is correctly classified False Negative
(FN) is the number of patients with DR which is
incorrectly classified.
Accuracy = (1)
Sensitivity = (2)
Specificity = (3)
False positive rate = (4)
Precision = (5)
4. EXPERIMENTAL RESULTS
The experiment was done on the images obtained from
IDRiD online database along with 50 images collected
from the regional hospitals in Kerala to form a total of 240
images. After preprocessing, and candidate region
extraction, the images were trained with the CNN network
which consisted of 10 layers. The training was done on
130 images for 30 epochs. Testing with the trained
network resulted in an accuracy of 93.8 percent..
In this section we evaluate the performance and results R-
CNN techniques. The performance is evaluated by the
various performance parameters like accuracy, sensitivity
and speed.
Input layer takes in a whole color image in CNN and only
suspected blocks in R-CNN. While thresholding the
intensity of suspected blocks, the range of L component is
taken to be greater than 80, a component to be less than -5
and b component to be greater than 70. First convolution
layer contains a 7 x 7 kernel with a stride of 2. ReLU acts
as the activation function for the previous layer's output.
Fig-6. Lab color space
A 2 x 2 max pooling layer is used with stride [2 2] thus
reducing the complexity of the computation. A second
convolutional layer with 32 filters of size [5 5] is used. It
also does zero paddings of size 2 along all the edges of the
layer input. The bias learning rate factor is set to 2 which
gives the learning rate for the bias in the layer to be twice
the current bias global learning rate. A second pooling
layer with pool size [3 2] is used with stride 2. Since we
aimed at classifying the whole dataset into two classes
namely ‘with DR’ and ‘without DR’ the fully connected
layer is designed for two outputs. The softmax layer
applies a softmax function to the input.
Table-1. Description of CNN layers
Layer # Layer Type Description
Layer 1 Input image Input image
Layer 2 Convolution
Filter size= 7 x 7,
No. of filters= 30,
Stride = 2
Layer 3 RELU
Introduces non-
linearity
Layer 4 Max Pooling
Pool size= 2 x 2,
Stride = 2
Layer 5 Convolution
Filter size= 5 x 5,
No. of filters= 32,
Padding = 2
Bias linear rate
factor= 2
Layer 6 RELU
Introduces non-
linearity
Layer 7 Max Pooling
Pool size= 3 x 2,
Stride = 2
Layer 8 Fully
connected
2 nodes
Layer 9 Softmax 2 classes
Layer 10 Classification
output
Output is the
message box
indicating the class
normal
or abnormal

4. CHALLENGES
The major challenge was about the computational
complexity that the laptops that we have with us could
handle. The program had to deal with several thousands of
neurons. Initially while working with a laptop that was
running on an intel Celeron n3050 processor which had a
clock speed of 1.6 GHz, it was not possible to train more
that 37 images. Matlab (software) crashed every single
time we tried to add a new image for training. Then we
moved to using a laptop that had a dual core intel i3
processor which had a clock frequency of 3.4GHz. With
this we were able to train more images but it took
approximately 20 to 25 minutes for training each new
image. Finally we moved to the systems systems running
on quad core intel i5 processor which took only less than
two minutes to train with an image.
5. CONCLUSION
Our project proved that machine learning techniques like
neural networks have very high future scope in disease
detection from medical images. Researches have already
proved the efficiency of R-CNN technique in the field of
object detection. With this project it is proved that R-CNN
can also the used for detecting very tiny features. R-CNN is
found to be highly accurate and sensitive for lesion
detection. An accuracy of about 93.8% is obtained for R-
CNN. It was due to the manual features that increased the
accuracy of R-CNN.
REFERENCES
[1] IAPB Resources “What is avoidable blindness
:Diabetic Retinopathy”
[2] Donald S FongMD, MPH Fredrick, L
FerrisIIMD, Mathew D DavisMD, Emily Y
ChewMD, “Causes of severe visual loss in the early
treatment Diabetic Retinopathy Study ETDRS
Report No.24,”American journal of
Ophthalmology, vol. 127, Issue 2, pp.137–
141,February 1999.
[3] Sudeshna Sil Kar and Santi P. Maity, “Automatic
Detection of Retinal Lesions for Screening of
Diabetic Retinopathy”IEEE Transaction on
Biomedical engineering , in press.
[4] Desire Sidibe, Ibrahaim Sadek,Fabrice
Meriaudeau, “Discrimination of retinal images
containing bright lesions using
sparse coded features and SVM,”
in Computers in Biology
and Medicine 62(2015), pp.175-184.
[5] Muhammad faisal, Djoko Wahono, Iketut
Eddy Purnama, Mochammad Hariadi, Mauridhi
Hery Purnomo, “Clasification of Diabetic
Rrtinopathy Patients using Support Vector
Machines based on digital retinal image,”Journal
Performance
Parameter
R-CNN (%)
Accuracy 95.5
Sensitivity 100
Specificity 93.8
False Positive Rate 6.3
Precision 85.7
Error Rate 4.5
Table-2. Performance evaluation
Chart-1. Dynamics of training process

of Theoretical and Applied Information
Technology,vol. 59 No.1,10th january 2014.
[6] Marco Foracchia, Enrico Grisan Alfredo Ruggeri,
“Luminosity and contrast normalization in retinal
images,”Medical Image Analysis, vol. 9, Issue 3,
pp. 179-190, June 2005.
[7] A. Hoover and M. Goldbaum, “Locating the optic
nerve in a retinal image using the fuzzy
convergence of the blood vessels.” IEEE
transactions on medical imaging, vol. 22, no. 8, pp.
951–8, 2003.
[8] Harry Pratt, Frans Coenen, Deborah M Broadbent,
Simon P Harding, Yalin Zheng, “Convolutional
Neural Networks for Diabetic
Retinopathy,”International Conference On
Medical Imaging Understading and Analysis 2016,
MIUA 2016, 6-8 July 2016,Loughborough,UK.
[9] Rishab Gargeya, Theodore Leng,“Automated
Identification of Diabetic Retinopathy Using Deep
Learning,”American Academy Of Ophthalmology,
in press.
[10] Arkadiusz Kwasigroch, Bartlomiej Jarzembinski,
Michał Grochowski, “Deep CNN based
decision support system for detection and
assessing the stage of diabetic retinopathy,”2018
International Interdisciplinary PhD Workshop
(IIPhDW),May 2018.
[11] Ratul Ghosh, Kuntal Ghosh, Sanjit Maitra, “
Automatic Detection and Classification of Diabetic
Retinopathy stages using CNN ,” 2017 4th
International Conference on Signal Processing and
Integrated Networks (SPIN), Feb. 2017.
[12] Georgia Gkioxari, Ross Girshick, Jitendra Malik, “
Contextual Action Recognition With R*CNN”, The
IEEE International Conference on Computer
Vision (ICCV), 2015, pp. 1080-1088.
[13] Ross Girshick, Jeff Donahue, Trevor Darrell,
Jitendra Malik , “ Rich Feature Hierarchies for
Accurate Object Detection and Semantic
Segmentation,” The IEEE Conference on
Computer Vision and Pattern Recognition
(CVPR), 2014, pp. 580-587.
[14] Shoji Kido, Yasusi Hirano, Noriaki Hashimoto,
“Detection and Classification of Lung
Abnormalities by Use of Convolutional Neural
Network (CNN) and Regions with CNN Features
(R-CNN),” International Workshop on Advanced
Image Technology , 2018.

IRJET- Automatic Detection of Diabetic Retinopathy using R-CNN

More Related Content

What's hot (19)

Similar to IRJET- Automatic Detection of Diabetic Retinopathy using R-CNN (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Automatic Detection of Diabetic Retinopathy using R-CNN