FPGA Implementation of Glaucoma Detection using Neural Networks

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 10 | Oct -2017 www.irjet.net p-ISSN: 2395-0072
© 2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 591
FPGA Implementation of Glaucoma Detection Using Neural Networks
Hitesh Shirke1, Dr. Nataraj Vijapur2
1 IV Semester M.Tech. Student, Dept. of Electronics and Communication Engineering.
K.L.E. Dr. M.S.Sheshgiri College of Engineering and Technology, Belagum-590008
Karnataka, India-590018
2Faculty Dept. of Electronics and Communication Engineering.
K.L.E. Dr. M.S.Sheshgiri College of Engineering and Technology, Belagum-590008
Karnataka, India-590018
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Glaucoma is an eye disease that occurs because
of the increased IntraOcular Pressure(IOP) which results in
damaged optic nerve. Because of glaucoma the optic cup size
gets increased and the optic cup to optic disk ratio(CDR)
increases. This paper proposes glaucoma disease detection
from retinal images using artificial neural network as the
classifier. Optic cup area, opticdiskareaandneuro-retinal rim
area are the features that are used for the classification. FPGA
implementation of neural networkoffersboththeadaptability
in reconfiguration and parallel architectureof neurons. Hence
FPGA is better choice for the neuralnetworksinsteadofDSP or
ASIC implementation. The feed forward back propagation
algorithm is used for the neural network. Matlab R2015a
software is used to extract the features and the neural
network is implemented on Spartan 3a FPGA kit.
Key Words: Glaucoma, Cup to disk ratio, Neuroretinal rim,
Artificial neural network, FPGA implementation
1. INTRODUCTION
The human eye is the most important part of the human
body. Therefore detection of human eye diseases is always
needed. Diseases like glaucoma, cataract, macular
degeneration, diabetic retinopathy affect the patient’s eye.
Glaucoma is a very silent disease of the eye. The scary part
about this disease is that it is so silent that it normally
doesn’t announce itself. It’s a disease which affects the optic
nerve of the eye. It causes thinning up and drying up of the
optic nerve. This results in visual loss which can be
irreversible.
Glaucoma can happeninanyoneacrossthepopulationwhich
means any person and age. The easiest way to prevent
glaucoma is to get it diagnosed in the very early stages.
There is no prevention of glaucoma if you are the person
who is going to have glaucoma in your lifetime. Glaucoma is
best treated medically using eye drops or using tablets to
reduce the pressure inside the eye. This work aims to detect
glaucoma using FPGA hardware. The artificial neural
network is used as a classifier to detect glaucoma.
1.1 Glaucoma
Measurably just about 45 million individuals worldwide
have glaucoma and around 79 million individuals are
expected to suffer from glaucoma by the year 2020. The
optic nerve is in charge of conveying the data seen by eye to
the brain. If there should be an occurrence of Glaucoma,
optic nerve is harmed and the data got by the brain gets
debased henceforth the vision loss. The essential reason
behind this harm of the optic nerve is increased IntraOcular
Pressure(IOP). Keeping in mind the end goal to keep up a
steady eye pressure,human eyeceaselesslydeliversaqueous
humor, a liquid which moreover continually streams out of
the eye. In the event of glaucoma, there is an increase in
liquid pressure of the eye on the grounds that the aqueous
humor does not stream out of the eye as it ought to be which
at last results in damage to the optic nerve fiber. This optic
nerve harm can be recognized by Optical Coherence
Tomography (OCT) and Heidelberg Retina Tomography
(HRT). The cost of glaucoma identification utilizing OCT and
HRT is high. Colour Fundus Image (CFI) technique has been
broadly used to analyze glaucoma and other visual
sicknesses.
1.2 Glaucoma Effects
Fig-1: normal fundus image

Fig-1. shows the fundus image of a normal eye taken from a
fundus camera. The bigger circular area istheoptic disk area
and the smaller circular area within the optic disk is the
optic cup area. The difference between theoptic disk andthe
optic cup area is called the Neuro-Retinal Rim(NRR).
Fig-2: (A) Healthy optic disk (B) Unhealthy optic disk
The glaucoma disease can be determined by observing the
changes in the optic disk area. As you can see in fig-2. (A)
indicates the healthy eye in which the optic cup size is very
small as compared to the optic disk. But because of
glaucoma the optic cup size gets increasedandtheopticcup
to optic disk ratio(CDR) increases this is shown in fig-2.(B).
2. METHODOLOGY
Methodology of proposed glaucoma detection system is
shown in above fig-3. Input fundus image is initially pre-
processed for noise removal. The pre-processed image is
subjected to feature extraction. The features such as optic
disk area, cup area and neuro retinal rim (NRR) area are
extracted. Then the next stage involves classification using
neural networks
2.1 Fundus Retinal Images
To determine the performance of the proposed method, we
have used 30 fundus images from HRF image database,
containing 15 healthy and 15 glaucoma images. The images
are in jpeg format with resolution of 3504×2336 pixels. An
input fundus image is shown in below fig-4.
Fig-3: Methodology of the proposed system
Fig-4: input fundus image
2.2 Image Pre-processing
Goal of the pre-processing is to enhance the picture by
improving certain elements and/or by removing unwanted
noise. The input images are of high resolution 3504×2336
pixels. Processing such high resolution images can take a lot
of processing time hence wehaveresizedtheinputimagesto
875×563 pixels. The image is then converted to HSV colour
map (Hue, Saturation,Value) byusingrgb2hsv(N)command.
To highlight the optic disk the red colour is removed from
the image and unwanted pixels are removed.
It is important to mark the region of interest (ROI) to reduce
the processing time also to initially mark the temporary
boundary of the optic disk. ROI is defined by approximately
marking the centre of ROI and drawing a rectanglearoundit.
Then the region of interest is cropped from the original
image. The pre-processing steps are shown in below fig-5.
Fig-5: image pre-processing

2.3 Feature extraction
2.3.1 Optic Disk Extraction
Feature extraction from the pre-processed image is done
using the segmentation technique calledmulti-thresholding.
This technique is similar to thresholding technique. This
technique is based on partitioning animageintoregionsthat
are similar according to the thresholding values. The main
part of this technique is to determine the threshold value.
For example, the threshold value will be between 0 and 255
for an 8 bit resolution image. After determining the
threshold values for all regions, every pixel in that region is
assigned a fix value. For example, if there are two regions
then the pixels belonging to a region can be assigned with a
value 1 and the remaining pixels with value 0. When we
establish more than one threshold values in the sameimage,
the technique is called multi-thresholding. Here higher and
lower limits are set for each region.Thusinordertosegment
the optic disk we have set the threshold value as shown
below;
Idisk(x,y) = {255 if a(x,y) > TD
= { 0 otherwise
Where TD is the threshold value for optic disk extraction.
Fig-6: optic disk extraction
First histogram equalization technique is used to enhance
the ROI image. Histogram equalization technique is used for
adjusting the image intensities so that its contrast can be
enhanced. After that thresholding is applied to extract the
optic disk as shown in above fig 6.
2.3.2 Optic Cup Extraction
The optic cup extraction is similar to the optic disk
extraction, by applying thresholding technique to the ROI
image. In order to segment the optic cup we have set the
threshold value TC higher than TD since the opticcuparea is
more brighter than the optic disk area as shown below;
Icup(x,y) = {255 if a(x,y) > Tc
= { 0 otherwise
Where TC is the threshold value for optic cup extraction.
Fig-7: optic cup extraction
2.3.3 Neuro Retinal Rim Extraction
The neuro retinal rim area are is area between the optic cup
area and the optic disk area. Hence it can be found by
subtracting the cup area from the disk area.
Fig-8: NRR extraction
Finally optic cup to disk ratio(CDR) and neuro retinal rim
area is calculated as follows.
Cup to disk ratio = [cup area] / [disk area]
Neuro retinal rim area = [optic disk area – optic cup area]
2.4 Artificial neural networks
Neural networks can be actualized utilizing analog or digital
systems. digital implementation is more prevalent as it has
the benefit of higher precision, better repeatability, lower
noise sensitivity, better testability, higher adaptability,also,
compatibility with different sorts of preprocessors. The
digital neural network hardware implementation can be
classified as (i) FPGA-implementation (ii) DSP-
implementation (iii) ASIC-implementation. DSP
implementation of neural network doesn't support the
parallel architecture which is required for neural networks.
The re-configurability is not supported by the ASIC based
neural network. FPGA implementation of neural network
offers both the adaptability in reconfiguration and parallel
architecture of neurons. Hence FPGA is better choice for the
neural networks.
Neural network consists of a set of nodes called neurons
which are connected to each other. There can be ‘n’ inputs to
a neuron and there is a weight associated with each
interconnection between neurons. The neural network can
learn by changing the weights of the connections based on
the inputs to the neurons. The neural network has three
layers namely input layer, hidden layer and output layer.

Based on the complexity of the problem there can bea single
hidden layer or multiple hidden layers. But a single hidden
layer will be enough to solve most of the problems.
2.4.1 Feed Forward Propagation
There are two algorithms used for neural network feed-
forward algorithm and back propagation algorithm. There
can’t be a neural network based on onlyfeedforwardoronly
on back propagation algorithm. In feed forward the output
sum is calculated for the network by taking the summation
of the product of weights and the inputs to corresponding
neurons. The activation functionisappliedtotheoutputsum
to get the output. For the first feed forward iteration the
weights are set randomly.
Fig-4 shows the structure of a neuron. The circle represents
the neuron which has x1, x2, x3, ….xn inputs and w1, w2,
w3,…wn are the weights on the connections respectively.
The output of a neuron is given by the equation
y = f (p) (1)
and the value of p is given by
(2)
Where xi is the ith input of the network and wi is the weight
on the ith connection. The function f(p) is called the
activation function of the neuron.
Fig-9: structure of a neuron
There are three types of activation functions namely Linear,
Log-sigmoid and Tan-sigmoid functions as shown below.
Linear function,
(3)
Sigmoid-function,
(4)
Tan-sigmoid function,
(5)
2.4.2 Back Propagation
In back propagation we calculate the error between the
actual output and desired output then we try to adjust the
weights to decrease the error. Same as the feed forward the
back propagation calculations are applied at each level but
the execution is carried out in reverse order i.e. from output
layer to the input layer.
Let’s first consider the output layer and hidden layer. First
the error is calculated at the output by subtractingtheactual
output from the desired output
error = desired – actual
Change in the output sum is calculated by takingtheproduct
of derivation of activation functionandthe error. Wewill call
this as the delta output sum.
Delta output sum = S`(output sum) * ( error )
Where S`(output sum) isthe derivationofactivationfunction
applied to the output sum.
The change in the nets corresponding to output neurons is
calculated by dividing the delta output sum by the hidden
layer results.
Delta weights = (delta output sum) / (hidden layer results)
Change in the hidden layer sum is denoted by delta hidden
sum and is calculated by following equation.
Delta hidden sum = [delta output sum / hidden-to-output
weights] * [ S`(hidden sum) ]
Now the change in the weights between the input and
hidden layer is calculated as follows
Delta weights = delta hidden sum / input data
This completes the back propagation.Afterthisfeedforward
is applied and the process is repeated thousands of times to
get more accurate results.
3 RESULTS
We have tested our methodology on images of high
resolution fundus (HRF) database. Sample outputs are as
given in the following tables. Results obtained for sample
glaucoma images are shown in table-1 and for sample
healthy images are shown in table-2.
Table-1: Glaucoma eye results
Filenam
e
Cup area Disk
area
Rim area Cup/dis
k ratio
prediction
01_g 804.248 3019.071 2214.823 0.266 Glaucoma
02_g 1256.637 3019.071 1762.433 0.416 Glaucoma

03_g 1385.442 3959.192 2573.75 0.35 Glaucoma
04_g 1194.591 4185.387 2990.796 0.285 Glaucoma
05_g 1520.531 4656.626 3136.095 0.327 Glaucoma
06_g 1194.591 4185.387 2990.796 0.285 Glaucoma
07_g 1134.115 4071.504 2937.389 0.279 Glaucoma
Table-2: Healthy eye results
Filenam
e
Cup area Disk
area
Rim area Cup/dis
k ratio
Prediction
01_h 1075.21 5026.548 3951.338 0.214 Healthy
02_h 907.92 3739.281 2831.36 0.243 Healthy
03_h 0 4185.387 4185.387 0 Unknown
04_h 530.929 3848.451 3317.522 0.138 Healthy
05_h 962.113 4901.67 3939.557 .196 Healthy
06_h 855.299 4185.387 3330.088 0.204 Healthy
07_h 346.361 4901.67 4555.309 0.071 Healthy
The extracted disk area, cup area and neuro retinal rim area
are fed to the artificial neural network. The performance of
ROC is as shown below.
3.1 Receiver Operating Characteristic curve (ROC)
We have used 30 fundus images fromHRF imagedatabaseto
determine the performance of our system. For training
purpose 16 images (8 healthy and 8 glaucoma images) were
used. Remaining 14 images were used during the testing
phase, seven images were glaucoma eye images and
remaining seven were healthy images. For all glaucoma
images the system predicted the result as glaucoma affected
and for 6 out of 7 healthy images it predicted the result as
healthy eye.
Therefore, accuracy can be given as
here,
Tp = true positive = glaucoma image being identified as
glaucoma image = 7
Tn = true negarive = healthy image being identified as
healthy image = 6
Fp = false positive = glaucoma image being identified as
healthy image = 1
Fn = false negative = healthy image being identified as
glaucoma image = 0
Therefore,
accuracy= 13 / 14 = 92.85 %
We obtained the accuracy of 92.85 %. The ROC graph is
shown in fig-10.
Fig-10: ROC graph
CONCLUSION
The proposed system will be very useful to detect glaucoma
efficiently so that the disease can be diagnosed in the early
stages. The system does not depend on trained glaucoma
specialist and expensive HRT/OCT machines. In this project
the glaucoma detection is done by extracting three features
i.e. optic disk, optic cup and neuro-retinal rim from a digital
fundus image. The artificial neural networks is used as a
classifier to identify the disease. The thresholding approach
is used for optic disk and optic cup segmentation.
To determine the System performance 30 retinal images(16
images for training and 14 images for testing) were
processed and their optic disk area, cup area and NRR area
were computed. The system proposed can identify the
presence of glaucoma to the accuracy of 92%. The accuracy
of the system can further be increased by using the same
camera settings for capturing the fundus images. Also it can
be improved by using advanced segmentation techniques.
ACKNOWLEDGEMENT
I thank the management, Principal,HODandstaffofVLSI and
EC department, K.L.E. Dr. M. S. Sheshgiri College of
engineering and technology, Belagavi, Karnataka, India and
my special thanks to my guide Dr. Nataraj Vijapur, Dept of
VLSI for encouraging me for this work.
REFERENCES
[1] Sheeba O., Jithin George, “Glaucoma Detection Using
Artificial Neural Network” IACSIT International Journal, Vol.
6, No. 2, April 2014.
[2] Kurnika Choudhary, Shamik Tiwari,” ANN Glaucoma
Detection using Cup-to-Disk Ratio and Neuro-retinal Rim”,
International Journal of Computer Applications (0975 –
8887) Volume 111 – No 11, February 2015

[3] Akhade, Snehal B., V. U. Deshmukh, and S. B. Deosarkar.
"Automatic optic disc detection in digital fundus images
using image processing", 2014 International Conference on
Green ComputingCommunicationandElectrical Engineering
(ICGCCEE), 2014.
[4] Gopal Datt Joshi, “Optic Disk and Cup Segmentation from
Monocular Color Retinal Images for Glaucoma Assessment”,
IEEE transactions on medical imaging, vol. 30, no. 6, june
2011
[5] Gayathri Devi T. M., Sudha. S. and Suraj. P., ,” Artificial
Neural Networks in retinal image analysis” 2015 3rd
International Conference on Signal Processing,
Communication and Networking (ICSCN).
[6] Megha Lotankar, Kevin Noronha, Jayasudha Koti,
“Detection of Optic Disc and Cup from Color Retinal Images
for Automated Diagnosis of Glaucoma”, 2015 IEEE UP
Section Conference on Electrical Computer and Electronics
(UPCON).
[7] Esraa Zeki Mohammed, Haitham KareemAli,,”Hardware
Implementation of Artificial Neural Network Using Field
Programmable Gate Array”, International Journal of
Computer Theory and Engineering, Vol. 5, No. 5, October
2013.
[8] Philippe Dondon, Julien Carvalho and Rémi Gardere, ,”
Implementation of a Feed-forwardArtificial Neural Network
in VHDL on FPGA”, neural network applications in electrical
engineering 2014.
[9] Muthuramalingam, A. Himavathi, S. Sriniv. "Neural
network implementation using FPGA: issues and
application.(field programmable gate array)", International
Journal of
Information Tec, April 2007 Issue
[10] www.explainstuff.com/introduction -to-neural-
networks.html.
[11]
https://www.xilinx.com/support/documentation/boards_an
d_kits/ug330.pdf
[12] Megha Lotankar, Kevin Noronha, Jayasudha Koti.
"Detection of optic disc and cup from color retinal images for
automated diagnosis of glaucoma", 2015 IEEE UP Section
Conference onElectrical ComputerandElectronics(UPCON),
2015.
[13] High Resolution Fundus Image database
https://www5.cs.fau.de/research/data/fundus-images/

FPGA Implementation of Glaucoma Detection using Neural Networks

More Related Content

What's hot (20)

Similar to FPGA Implementation of Glaucoma Detection using Neural Networks (20)

More from IRJET Journal (20)

Recently uploaded (20)

FPGA Implementation of Glaucoma Detection using Neural Networks