Irjmets - Analysis of Diabetic Retinopathy in Fundus Images Using CNN

e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science

( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:03/Issue:09/September-2021 Impact Factor- 6.752 www.irjmets.com
ANALYSIS OF DIABETIC RETINOPATHY IN FUNDUS IMAGES USING CNN

Humaira Sadiq*1, Ankur Gupta*2
*1M-TECH, Computer Science Department, RIMT, Chandigarh, India.
*2Assistant Professor CSE Department, School of Engineering RIMT University Chandigarh, India.
ABSTRACT
Diabetes, often mentioned by doctors as Diabetes Milletus, describes a group of metabolic diseases during
which the person has high blood sugar (blood sugar), either because insulin production is inadequate, or
because the body's cells don't respond properly to insulin, or both. Patients with high blood sugar will typically
experience polyuria (frequent urination), they go to become increasingly thirsty (polydipsia) and hungry
(polyphagia). Diabetic Retinopathy may be a diabetes complication that affects eyes. It's caused by damage to
the blood vessels of the light-sensitive tissue at the rear of the eye (retina). Each eye of the patient are often in
one among the 5 levels: from 0 to 4, where 0 corresponds to the healthy state and 4 is that the most severe
state. Different eyes of an equivalent person are often at different levels (although some contestants managed
to leverage the very fact that two eyes aren't completely independent). Our aim was to detect the presence of
Diabetic Retinopathy in fundus images using Convolution Neural network, for which we used a pre-trained
Image-Net model, Inception V3 and Tensor-flow. We achieved the accuracy of approximately 80 percent in
training the model.
Keywords: Polydipsia, Polyphagia.
I. INTRODUCTION
Diabetic retinopathy is a microvascular complication of diabetes mellitus and is a critical reason for new-onset
blindness. Diabetic macular changes as yellowish spots and full or halfway thickness extravasations through the
retina were noticed for the first time by Eduard Jäger. In 1855, he published "Beiträge zur Pathologie des
Auges" where he incorporated his fundus artworks. Jaeger's discoveries were dubious until 1872, when
Edward Nettleship published his fundamental paper on "Oedema or cystic infection of the retina", giving the
first histopathological evidence of "cystoid degeneration of the macula" in patients with diabetes. In 1876,
Wilhelm Manz portrayed the proliferative changes happening in diabetic retinopathy and the significance of
tractional retinal detachments and vitreous hemorrhages. However, it was not until 1943 that crafted by Arthur
James Ballantyne gave proof that diabetic retinopathy addresses a remarkable type of vascular infection.
Various multi-focused clinical preliminaries during the most recent ten years have contributed considerably to
the comprehension of the normal history of diabetic retinopathy and have set up the worth of concentrated
glycaemic control in diminishing both the danger of beginning and the movement of diabetic retinopathy
Diabetes is the constant state brought about by a strange expansion in the glucose level in the blood and which
makes the harm the veins. The minuscule veins that feed the retina are harmed by the expanded glucose level.
Diabetes is the fifth deadliest sickness in the USA, and still there is no fix. The complete yearly financial expense
of diabetes in 2002 was assessed to be US $132 billion, or one out of each 10 medical care dollars spent in the
USA [1]. Diabetic retinopathy (DR) happens when diabetes harms the small veins inside the retina, the light-
touchy tissue at the rear of the eye. This small vein will spill blood and liquid on the retina, framing highlights,
for example, microaneurysms, speck and blotch Hemorrhage, hard exudates, cotton fleece spots, or venous
circles. Microaneurysms are the soonest clinical indication of diabetic retinopathy and happen optional to
slender divider outpouching due to pericyte misfortune. They show up as little red specks in the shallow retinal
layers, and there is fibrin and red platelet collection in the microaneurysm lumen. A break produces
smudge/fire hemorrhages. Influenced regions might seem yellowish on schedule, as endothelial cells multiply
and produce cellar layer. Speck and smear hemorrhages happen as microaneurysms crack in the more
profound layers of the retina, for example, the inward atomic and external plexiform layers. These seem like
microaneurysms in case they are little; fluorescein angiography might be expected to recognize the two. Hard
exudates are brought about by the breakdown of the blood-retina obstruction, permitting spillage of serum
proteins, lipids, and protein from the vessels.Cotton-fleece spots are nerve fiber layer areas of localized necrosis
www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

[77]
e-ISSN: 2582-5208
from impediment of precapillary arterioles. With the utilization of fluorescein angiography, there is no hairlike
perfusion. These are as often as possible lined by microaneurysms and vascular hyper-penetrability. Venous
circles and venous beading as often as possible happen nearby spaces of non-perfusion and reflect expanding
retinal ischemia. Their event is the main indicator of movement to proliferative diabetic retinopathy.DR can be
broadly classified as non-proliferative diabetic retinopathy (NPDR) and proliferative diabetic retinopathy
(PDR). Depending on the presence of features on the retina, the stages of DR can be identified.In the initial
stages of diabetic retinopathy, patients are generally asymptomatic; in the more advanced stages of the disease,
however, patients may experience symptoms that include floaters, blurred vision, distortion, and progressive
visual acuity loss.There are many screening tools available to diagnose Diabetic Retinopathy. Digital fundus
cameras are used to take the retinal vessel images; therefore, unnecessary brightness, environment, and the
process of acquisition of fundus image degrade the image quality to some extent. Hence image enhancement is
always required to improve the quality of desired image.The dataset used consists of images that come from
different models and types of cameras, which can affect the visual appearance of left vs. right. Some images are
shown as one would see the retina anatomically (macula on the left, optic nerve on the right for the right eye).
Others are shown as one would see through a microscope condensing lens (i.e. inverted, as one sees in a typical
live eye exam). Like any real-world data set, we will encounter noise in both the images and labels.Images may
contain artefacts, be out of focus, underexposed, or overexposed. A major aim of this project is to develop
robust algorithms that can function in the presence of noise and variation and deduce features to classify the
images in the data set. Analysis of disease is a major part of study today in Medical Science. Machine Learning
Algorithms can be used for learning the trends & nature of disease based on various parameters.
II. LITERATURE REVIEW
A couple of years prior, a few of us started contemplating whether there was a way Google innovations could
further develop the DR screening measure, explicitly by exploiting ongoing advances in Machine Learning and
Computer Vision. "Being developed and Validation of a Deep Learning Algorithm for Detection of Diabetic
Retinopathy in Retinal Fundus Photographs", distributed in JAMA, Google introduced a profound learning
calculation fit for deciphering indications of DR in retinal photos, conceivably assisting specialists with
screening more patients in settings with restricted assets. Perhaps the most widely recognized approaches to
identify diabetic eye illness is to have an expert look at photos of the rear of the eye (Figure 1) and rate them for
sickness presence and seriousness. Seriousness is dictated by the kind of injuries present (for example
microaneurysms, hemorrhages, hard exudates, and so on), which are characteristic of draining and liquid
spillage in the eye. Deciphering these photos requires particular preparing, and in numerous locales of the
world there aren't sufficient qualified graders to screen each and every individual who is in danger.
Fig 2.1: Examples of retinal fundus photographs that are taken to screen for DR. The image on the
left is of a healthy retina.
(A) Whereas the image on the right is a retina with referable diabetic retinopathy.
(B) Due a number of hemorrhages (red spots) present.
Nikolaos Maniadakis et al. (2019): described interpretation of cost-effectiveness data should be treated with
caution in this case; details of the therapeutic regimen, such as dosage and frequency, and clinical efficacy of the
[78]
e-ISSN: 2582-5208
treatments should be considered in relation to policy-making decisions. Given the scarcity of resources, the
ever increasing significance of health technology assessment and the substantial differences in the
methodologies of the studies presented in this review, there is a pressing need for more advanced and
standardized approaches to assessing the effectiveness and cost effectiveness of the emerging anti-VEGF
pharmacotherapies for the treatment of DMO.
Hagos et al. (2019): attempted to prepare Inception-Net V3 for 5- class characterization with pretrain on
ImageNet dataset furthermore, accomplished precision of 90.9%.
Carson Lam et al. (2018): described Automated detection and screening offers a one of a kind chance to
forestall a critical extent of vision misfortune in our populace. Lately, specialists have added CNNs into the
arrangement of calculations used to evaluate for diabetic sickness. CNNs guarantee to use the a lot of pictures
that have been amassed for doctor deciphered screening and gain from crude pixels. The high fluctuation and
low inclination of these models could permit CNNs to analyze a more extensive scope of nondiabetic sicknesses
also. Clinical pictures are laden with inconspicuous components that can be critical for finding. Luckily, the
regularly sent designs have been enhanced to perceive plainly visible provisions, for example, those present in
the ImageNet dataset. We may subsequently require another worldview for diagnosing illnesses by means of
CNN models. This could be a two phase injury location pipeline that includes highlight confinement followed by
order and further preprocessing steps to section out pathologies hard to observe by manual investigation, lastly
rebalancing network loads to represent class irregular characteristics found in clinical datasets. Generally
speaking, our future objectives include further developing discovery of gentle illness and progressing to really
testing and useful multi-grade sickness location.
Yung-Hui Li et al. (2018): described it is attainable to prepare a profound learning model for programmed
analysis of DR, as long as we have sufficient information for factual model preparing. Moreover, the information
base planning stage just necessities an absolute mark for each preparation picture. It doesn't need definite
explanation for retinal vessel following in each picture. Henceforth, the time has come proficient contrasted
with the customary AI based technique for programmed analysis of DR.The last exactness can accomplish
86.17% and 91.05% for five-class and binary class classifications separately.
Mapa Mudiyanselage Prabhath Nishantha Piyasena et al. (2018): described Analytic test exactness for the
recognition of any degree of DR showed that DRS utilizing two fields conveyed at non-essential consideration
settings is an attainable methodology. Dilatation of the students didn't work on the location of any degree of DR
for those with gradable pictures, however a wide scope of ungradable were introduced in these investigations
that this perspective should be considered when setting up DRSP. There was no sufficient proof in essential
examinations to remark on DTA of non-ophthalmological HR on DRS, so this perspective requires further
exploration. Great quality advanced imaging has the potential for constant understanding of retinal pictures,
which along with advising for hazard elements might work on the agreeableness of DRS and take-up of
reference for ophthalmic evaluation whenever directed in a socially worthy manner.
Dr.Ruksar Fatima, et al.(2017): descibed the quick and productive early discovery of Diabetic Retinopathy is
just conceivable in case there is a successful technique for portioning the diabetic highlights in the fundus
picture. The proposed strategies presents a quick, successful and powerful method of distinguishing diabetic
highlights in the fundus pictures which can be utilized for characterization of the pictures dependent on the
seriousness of the sickness..Various upsides of the boundaries such as precision, affectability, positive prescient
worth (PPV) what's more, explicitness were gotten which only not really set in stone powerful location of
Diabetic Retinopathy.
Anjali R. Shah et al. (2017): depicted massive upgrades under the watchful eye of patients with diabetes and
DR in the course of recent years are an illustration of the critical effect lab and clinical exploration can make on
the administration of persistent fundamental sickness. Regardless of extraordinary advances, however, the
extended expansion in the quantity of patients with DR in the coming many years advises us that there is still
advancement to be made. Ebb and flow research prompting a superior comprehension of atomic pathways,
advancement of novel restorative targets, and utilization of nanotechnology, combined with continually

[79]
e-ISSN: 2582-5208
working on analytic and imaging innovation and shared medical services conveyance frameworks, vows to
additional our capacity to upgrade and keep up with vision in diabetic patients.
Karami, et al. (2017): proposed a programmed DR discovery approach for computerized fundus pictures
which was a word reference learning (DL)- based methodology [15]. This discovery approach was produced
dependent on the best nuclear portrayal of fundus pictures relying upon the learned word references made
utilizing K-SVD Algorithm. The area of class which included minimal quantities of best specific particles was
considered to incorporate the test picture. In view of the directed examinations what's more, accomplished
results it was seen that around 70% exactness was accomplished for ordinary pictures and 90% of exactness
was accomplished for diabetic pictures when theproposed strategy was tried on 30 shading fundus pictures
Carrera, et al. (2017): proposed a novel methodology utilizing computerized preparing to recognize the DR
from retinal pictures at beginning stage with the goal that it very well may be controlled [16]. This methodology
meant to characterize the grade of NDPR in the retinal picture in programmed way. The exhibition of proposed
approach was tried on an information base which included 400 retinal pictures. In view of the 4-grade scale, the
pictures were arranged by this proposed approach. Through the test results it was seen that around 94% of
prescient limit and 95% of affectability were accomplished as yield. Higher power was likewise accomplished
by this proposed approach according to the assessments performed toward the finish of this examination
Problem Statement
1. India is one of the leading countries with the occurrence of diabetes patients. This deduces that other
diseases related to diabetes also prevail amongst the masses.
2. One such disease is Retinopathy, where patients often lose the sight due to lack of knowledge and late
detection of the disease.
3. This project would aid the Doctors to detect the advent of this disease at the earliest, without undergoing
costly imaging clinical tests.
III. RESEARCH METHODOLOGY
Fundus Photography
Fundus photography includes catching a photo of the rear of the eye for example fundus. Particular fundus
cameras that comprise of a perplexing magnifying instrument joined to a blaze empowered camera are utilized
in fundus photography. The fundamental designs that can be envisioned on a fundus photograph are the focal
and fringe retina, optic circle and macula. Fundus photography can be performed with hued channels, or with
specific colors including fluorescein and indocyanine green.
Fig 3.1 Retinal fundus image

The models and innovation of fundus photography has progressed and developed quickly throughout the last
century. Since the types of gear are refined and testing to produce to clinical norms, a couple of makers/brands
are accessible on the lookout: Digisight, Volk, Topcon, Zeiss, Canon, Nidek, Kowa, CSO, CenterVue, and Ezer are
some illustration of fundus camera makers. Fundus photos are visual documentation that record the presence
of a patient's retina. The photos permit the clinician to consider a patient's retina, distinguish retinal changes
[80]
e-ISSN: 2582-5208
and survey a patient's retinal discoveries with a collaborator. Fundus photos are regularly called upon in a wide
assortment of ophthalmic conditions. Fundus photography is utilized to investigate irregularities related to
infections that influence the eye and to screen the movement of the illness. It is imperative for illness cycles, for
example, macular degeneration, retinal neoplasms, choroid unsettling influences and diabetic retinopathy.
Furthermore it helps in distinguishing glaucoma, different sclerosis, and other focal sensory system anomalies.
It assesses abnormalities in the fundus, screens the movement of an illness, the executives and helpful result.
They are significant to make a beginning stage to more readily comprehend an illness' movement. Fundus
photos might be valuable in case there is another infection influencing the fundus and for the arranging of extra
administration alternatives. The clinical need of fundus photography and other indicative imaging should be
recorded in a deliberate style with the goal that the clinician can look at photos of a patient from various
timetables. Reports of a patient's clinical record should comprise of a new, important history, progress notes
and fundus photos portraying and supporting the pertinent analysis. The photos should be named fittingly, for
example, which eye, the date, and patient subtleties. The patient's records should contain reported results of
the fundus photography just as a portrayal of varieties from past photos. They ought to contain an
understanding of those outcomes and the important changes it could have on treatment plan. Fundus photos
without a translation are viewed as out of date. The records ought to be readable, and contain appropriate
patient data and clinician subtleties. The understanding of fundus photos that are glaucomatous should contain
a depiction of the vertical and even cup to circle proportion, vessel design, diffuse or central paleness,
unevenness and improvement of the above factors. The retinal nerve fiber layer ought to likewise be examined
and remarked on [3]. It is likewise a helpful apparatus in equitably estimating twist just as in reporting and
recording movement of infections over the long run. Fundus photography doesn't supplant binocular aberrant
ophthalmoscopy; it is an apparatus to enhance and supplement existing discoveries and to keep a record of
infection movement. Fundus photography is basically used to screen the movement of a retinal or optic nerve
head problem. It is likewise useful for giving photograph documentation to the continuum of care and to screen
the patient's visual condition.
Convolution Neural Networks:
In AI, a convolutional neural organization (CNN or Conv-Net) is a class of profound, feed-forward fake neural
organizations, most normally applied to investigating visual symbolism. CNNs utilize a variety of multi-facet
perceptrons intended to require insignificant preprocessing. They are otherwise called shift invariant or space
invariant fake neural organizations (SIANN), in light of their common loads engineering and interpretation
invariance qualities. Convolutional networks were roused by natural cycles in that the availability design
between neurons looks like the association of the creature visual cortex. Individual cortical neurons react to
improvements just in a confined locale of the visual field known as the open field. The open fields of various
neurons halfway cross-over to such an extent that they cover the whole visual field. CNNs utilize moderately
minimal pre-preparing contrasted with other picture characterization calculations. This implies that the
organization learns the channels that in conventional calculations were hand-designed. This autonomy from
earlier informaltion and human exertion in include configuration is a significant benefit. They have applications
in picture and video acknowledgment, recommender frameworks and regular language preparing. A CNN
comprises of an information and a yield layer, just as numerous secret layers. The secret layers of a CNN
commonly comprise of convolutional layers, pooling layers, completely associated layers and standardization
layer. [4]

[81]
e-ISSN: 2582-5208
Fig 3.2: convolutional neural network

Depiction of the cycle as a convolution in neural organizations is by show. Numerically it is a cross-connection
instead of a convolution. This just has importance for the records in the grid, and consequently which loads are
put at which list.
Convolutional Layers
Convolutional layers apply a convolution activity to the information, passing the outcome to the following layer.
The convolution imitates the reaction of an individual neuron to visual boosts. Each convolutional neuron
measures information just for its open field. Albeit completely associated feed-forward neural organizations
can be utilized to learn includes just as group information, it isn't down to earth to apply this engineering to
pictures. An exceptionally high number of neurons would be fundamental, even in a shallow (inverse of
profound) engineering, because of the extremely huge information sizes related with pictures, where every
pixel is an important variable. For example, a completely associated layer for a (little) picture of size 100 x 100
has 10000 loads for every neuron in the subsequent layer. The convolution activity carries an answer for this
issue as it lessens the quantity of free boundaries, permitting the organization to be more profound with less
parameters.[8] For example, paying little heed to picture size, tiling districts of size 5 x 5, each with similar
shared loads, requires just 25 learnable boundaries. Thusly, it settle the evaporating or detonating slopes issue
in preparing conventional multi-facet neural organizations with numerous layers by utilizing back engendering.
Pooling layers
Convolutional organizations might incorporate nearby or worldwide pooling layers, which join the yields of
neuron bunches at one layer into a solitary neuron in the following layer For instance, max pooling utilizes the
greatest worth from every one of a group of neurons at the earlier layer .Another model is normal pooling,
which utilizes the normal worth from every one of a group of neurons at the earlier layer. Pooling layers
decrease the components of information by joining the yields of neuron bunches at one layer into a solitary
neuron in the following layer. Nearby pooling joins little bunches, tiling sizes, for example, 2 x 2 are normally
utilized. Worldwide pooling follows up on every one of the neurons of the element map.
Fully connected layers
Completely associated layers interface each neuron in one layer to each neuron in another layer. It is on a
fundamental level equivalent to the conventional multi-facet perceptron neural organization (MLP). After a few
convolutional and max pooling layers, the last characterization is done through completely associated layers.
Neurons in a completely associated layer have associations with all enactments in the past layer, as found in
normal (non-convolutional) counterfeit neural organizations. Their enactments would thus be able to be
[82]
e-ISSN: 2582-5208
registered as a relative change, with framework duplication followed by a predisposition offset (vector
expansion) of a learned or fixed inclination term.
Weights
CNNs share loads in convolutional layers, which implies that a similar channel (loads bank) is utilized for each
open field in the layer; this decreases memory impression and further develops execution. Every neuron in a
neural organization processes a yield esteem by applying a particular capacity to the information esteems got
from the responsive field in the past layer. The capacity that is applied to the info esteems is dictated by a
vector of loads and an inclination (regularly genuine numbers). Learning comprises of iteratively changing
these predispositions and loads. The vector of loads and the predisposition are called channels and address
specific highlights of the information). A distinctive element of CNNs is that numerous neurons can have a
similar channel. This lessens the memory impression in light of the fact that a solitary inclination and a solitary
vector of loads are utilized across all open fields that share that channel, instead of each responsive field having
its own predisposition and vector weighting.
History
CNN design follows vision processing in living organisms.
Receptive fields
Work by Hubel and Wiesel during the 1950s and 1960s showed that feline and monkey visual cortexes contain
neurons that independently react to little districts of the visual field. Given the eyes are not moving, the locale
of visual space inside which visual boosts influence the terminating of a solitary neuron is known as its open
field. Adjoining cells have comparative and covering open fields. Responsive field size and area fluctuates
methodicallly across the cortex to shape a total guide of visual space. The cortex in every half of the globe
addresses the contralateral visual field.
Their 1968 paper distinguished two essential visual cell types in the mind:
 simple cells, whose yield is augmented by straight edges having specific
 directions inside their responsive field
 complex cells, which have bigger responsive fields, whose yield is unfeeling toward the specific situation of
the edges in the field.
Neocognitron
The neocognitron was presented in 1980, The neocognitron doesn't need units situated at numerous
organization positions to have similar teachable loads. This thought shows up in 1986 in the book form of the
first backpropagation paper. Neocognitrons were created in 1988 for fleeting signs. Their plan was worked on
in 1998, summed up in 2003 and improved around the same time. The neocognitron was presented by
Kunihiko Fukushima in 1980. It was roused by the previously mentioned work of Hubel and Wiesel. The
neocognitron presented the two essential sorts of layers in CNNs: convolutional layers, and downsampling
layers. A convolutional layer contains units whose open fields cover a fix of the past layer. The weight vector
(the arrangement of versatile boundaries) of such a unit is frequently called a channel. Units can share
channels. Downsampling layers contain units whose responsive fields cover patches of past convolutional
layers. Such a unit ordinarily figures the normal of the enactments of the units in its fix. This down testing
serves to effectively arrange objects in visual scenes in any event, when the items are moved. In a variation of
the neocognitron called the cresceptron, rather than utilizing Fukushima's spatial averaging, J. Weng et al.
presented a technique called max-pooling where a downsampling unit figures the limit of the enactments of the
units in its fix. Max-pooling is frequently utilized in current CNNs. A few regulated and unaided learning
calculations have been proposed throughout the a long time to prepare the loads of a neocognitron. Today, be
that as it may, the CNN design is typically prepared through back spread. The neocognitron is the primary CNN
which requires units situated at various organization positions to have shared loads.
Convolutional neural organizations were introduced at the Neural Information Processing Workshop in 1987,
naturally breaking down time-shifting signs by supplanting learned augmentation with convolution on
schedule, and exhibited for discourse acknowledgment.
[83]
e-ISSN: 2582-5208
LeNet-5
LeNet-5 CNN architecture is comprised of 7 layers. The layer organization comprises of 3 convolutional layers, 2
sub-examining layers and 2 completely associated layers. The main layer is the information layer — this is for
the most part not considered a layer of the organization as nothing is learnt in this layer. The info layer is worked
to take in 32x32, and these are the components of pictures that are passed into the following layer. The
individuals who know about the MNIST dataset will know that the MNIST dataset pictures have the
measurements 28x28. To get the MNIST pictures measurement to the meet the prerequisites of the info layer,
the 28x28 pictures are padded.
LeNet-5, a spearheading 7-level convolutional network by LeCun et al. in 1998, that orders digits, was applied by
a few banks to perceive transcribed numbers on checks digitized in 32x32 pixel pictures. The capacity to handle
higher goal pictures requires bigger and more convolutional layers, so this method is obliged by the accessibility
of processing resources.
Fig 3.3: LeNet-5 Architecture

Neural Abstraction Pyramid
The feed-forward design of convolutional neural organizations was reached out in the neural deliberation
pyramid by parallel and criticism associations. The subsequent repetitive convolutional network takes into
account the adaptable fuse of logical data to iteratively resolve neighborhood ambiguities. As opposed to past
models, picture like yields at the most noteworthy goal were produced. e.g., for semantic division, picture
remaking, and article restriction assignments.
Fig 3.4: Neural Abstraction Pyramid

GPU implementations
A graphics processing unit (GPU) is a specific electronic circuit intended to quickly control and modify
memory to speed up the production of pictures in a casing cradle proposed for yield to a showcase gadget. GPUs
are utilized in inserted frameworks, cell phones, PCs, workstations and game control center.

[84]
e-ISSN: 2582-5208
Current GPUs are exceptionally proficient at controlling PC illustrations and picture handling. Their profoundly
equal construction makes them more productive than universally useful focal preparing units (CPUs) for
calculations that interaction enormous squares of information in equal. In a PC, a GPU can be available on a
video card or implanted on the motherboard. In specific CPUs, they are implanted on the CPU pass on.
Following the 2005 paper that set up the worth of GPU for AI, a few distributions portrayed more productive
approaches to prepare convolutional neural organizations utilizing GPUs. In 2011, they were refined and
carried out on a GPU, with great outcomes. In 2012, Ciresan et al. essentially enhanced the best execution in the
writing for numerous picture information bases, including the MNIST data set, the NORB data set, the HWDB1.0
dataset (Chinese characters), the CIFAR10 (dataset of 60000 32x32 named RGB pictures), and the ImageNet
dataset. [4]
Tensor-Flow™
Tensor-Flow™ is an open source programming library for elite mathematical calculation. Its adaptable design
permits simple arrangement of calculation across an assortment of stages (CPUs, GPUs, TPUs), and from work
areas to bunches of workers to versatile and edge gadgets. Initially created by specialists and designers from
the Google Brain group inside Google's AI association, it accompanies solid help for AI and profound learning
and the adaptable mathematical calculation center is utilized across numerous other logical spaces.
We utilized a model prepared on the Image-Net Large Visual Recognition Challenge dataset. These models can
separate between 1,000 distinct classes, similar to Dalmatian or dishwasher. You will have a decision of model
designs, so you can decide the right compromise between speed, size and exactness for your concern. We
utilized this equivalent model, however retrain it to distinguish few classes dependent on our own examples.[5]
Tensor-Flow is Google Brain's second-age framework. Variant 1.0.0 was delivered on February 11, 2017. While
the reference execution runs on single gadgets, Tensor-Flow can run on different CPUs and GPUs (with
discretionary CUDA and SYCL expansions for broadly useful figuring on designs handling units). TensorFlow is
accessible on 64-cycle Linux, macOS, Windows, and portable figuring stages including Android and iOS.
Tensor-Flow calculations are communicated as stateful dataflow diagrams. The name Tensor-Flow gets from
the tasks that such neural organizations perform on multidimensional information clusters. These exhibits are
alluded to as "tensors". In June 2016, Dean expressed that 1,500 vaults on GitHub referenced TensorFlow, of
which just 5 were from Google .[5][6][7]
Image-Net
ImageNet is a picture dataset coordinated by the WordNet progression. Each significant idea in WordNet,
perhaps depicted by various words or word phrases, is known as a "equivalent set" or "synset".
There are in excess of 100,000 synsets in WordNet, greater part of them are things (80,000+). In ImageNet, we
expect to give on normal 1000 pictures to represent every synset. Pictures of every idea are quality-controlled
and human-explained. In its consummation, we trust ImageNet will offer huge number of neatly arranged
pictures for the vast majority of the ideas in the WordNet chain of importance.
The ImageNet project is enlivened by a developing slant in the picture and vision research field – the
requirement for additional information. Since the time the introduction of the computerized time and the
accessibility of web-scale information trades, analysts in these fields have been striving to plan an ever
increasing number of refined calculations to list, recover, coordinate and clarify mixed media information. Be
that as it may, great exploration needs great asset.
To handle these issue in enormous scope (think about your developing individual assortment of computerized
pictures, or recordings, or a business web crawler's data set), it would be massively useful to analysts if there
exists a huge scope picture information base. This is the inspiration for us to assemble ImageNet. We trust it
will end up being a valuable asset to our examination local area, just as anybody whose exploration and training
would profit with utilizing a huge picture database[8]
Inception V3
The "Inception" miniature engineering was first presented by Szegedy et al. in their 2014 paper, Going Deeper
with Convolutions: The objective of the initiation module is to go about as a "staggered include extractor" by
[85]
e-ISSN: 2582-5208
figuring 1×1, 3×3, and 5×5 convolutions inside a similar module of the organization — the yield of these
channels are then stacked along the channel measurement and prior to being taken care of into the following
layer in the organization. The first manifestation of this design was called GoogLeNet, however resulting signs
have just been called Inception vN where N alludes to the form number put out by Google. The Inception V3
design remembered for the Keras center comes from the later distribution by Szegedy et al., Rethinking the
Inception Architecture for Computer Vision (2015) which proposes updates to the commencement module to
additional lift ImageNet characterization accuracy.[9] The loads for Inception V3 are more modest than both
VGG and ResNet, coming in at 96 MB.
Figure 3.5: The original inception model used in Google Le Net

IV. RESULTS AND DISSCUSSION
We are having 300 images in the trained set and 300 images in the test set. Batch of 20 images set were taken
here because our algorithm processes only 20 images at a time. After that repeated iterations are done, our
algorithm was run again and again (20 images in one iteration). For every batch of 20 images (10 diseased and
10 non diseased), the iteration to collect the accuracy parameters were run. Sensitivity and specificity define
the ability of a clinical test to correctly identify people with and without a specific disease. For low prevalence
diseases, a high specificity is required to avoid large numbers of false positive results. The British Diabetic
Association (now Diabetes UK) has set a required screening standard for DR of at least 80% sensitivity and
95% specificity. Sensitivity and specificity are inversely proportional, meaning that as the sensitivity increases,
the specificity decreases and vice versa.
A true positive is a result where the model accurately predicts the positive class. Additionally, a true negative is
a result where the model accurately predicts the negative class. A false positive is a result where the model
inaccurately predicts the positive class. Also, a false negative is a result where the model mistakenly predicts
the negative class.
True positive: Sick people correctly identified as sick
False positive: Healthy people incorrectly identified as sick
True negative: Healthy people correctly identified as healthy
False negative: Sick people incorrectly identified as healthy

[86]
e-ISSN: 2582-5208
TABLE 4.1
One Two Three Four Five
sensitivity 78.95% 85.71% 66.67% 90.91% 76.92%
specificity 90.91% 88.24% 90.91% 83.33% 100.00%
PPV 90.91% 88.24% 90.91% 83.33% 100.00%
NPV 78.95% 85.71% 66.67% 90.91% 76.92%
120.00%
100.00%
80.00%
60.00%
40.00%
20.00%
0.00%
Sensitivity Specificity PPV NPV
Fig 4.1: Accuracy of one batch of images

TABLE 4.2
False Negative 8 5 15 3 9
False Positive 3 4 3 6 0
A false positive is an error in binary classification in which a test result incorrectly indicates the presence of a
condition such as a disease when the disease is not present, while a false negative is the opposite error where
the test result incorrectly fails to indicate the presence of a condition when it is present.
16
14
12
10
8
6
4
2
0
False Negative False Positive
Fig 4.2: Accuracy of another Batch of images.

[87]
e-ISSN: 2582-5208
The overall accuracy of the model as calculated as gross of running the model for maximum number of iteration
was calculated as:
TABLE 4.3
POSITIVE
TRUE POSITIVE 414
FALSE NEGATIVE 104
NEGATIVE
FALSE NEGATIVE 26
TRUE NEGATIVE 492
Sensitivity 79.92%
Specificity 94.98%
Positive predictive value 94.09%
Negative predictive value 82.55%
500
450
400
350
300
250
200
150
100
50
0
Fig 4.3: Accuracy of overall model.

V. CONCLUSION
Machine learning algorithms always looks for the most relevant paths to achieve accuracy. However, the
accuracy with analysis & pattern generation is low as the standardization of images is a greater challenge than
training the model to predict the disease. The quality of images and the angle at which the image is taken,
causes a huge impact on accuracy of the model training. Training the ground staff for taking images with such
accuracy is not possible. Using Machine learning from the ground, for training the computer itself to take the
fundus image with great accuracy would be the best beginning of a project like this Also the preprocessing of
images took a lot of time, which in future can be made efficient by using faster algorithms.
FUTURE SCOPE
In future [Insha Allah], we may train the dataset that might be collected locally, using the newly trained
computer aided fundus photography for greater accuracy of images. Also this would help in understanding the
probability of people locally, suffering from diabetic Retinopathy, at early stages, to avoid the damage as much
as possible.

[88]
e-ISSN: 2582-5208
VI. REFERENCES
[1] Frank, R. N. Diabetic retinopathy. Prog. Retinal Eye Res., 1995, 14(2), 361–392.
[2] Acharya, U. R., Ng, E. Y. K., and Suri, J. S. Image modelling of human eye, 2008 (Artech House,
Norwood, Massachusetts).
[3] Deep Learning for Detection of Diabetic Eye Disease available from
https://ai.googleblog.com/2016/11/deep-learning-for-detection-of-diabetic.html.
[4] Al‐Rubeaan, Khalid, et al. "Diabetic retinopathy and its risk factors in a society with a type 2 diabetes
epidemic: a Saudi National diabetes Registry‐based study." Actaophthalmologica 93.2 (2015): e140-
e147.
[5] Myśliwiec, Małgorzata, et al. "The role of vascular endothelial growth factor, tumor necrosis factor
alpha and interleukin-6 in pathogenesis of diabetic retinopathy." Diabetes research and clinical
practice 79.1 (2008): 141-146. (2004)
[6] Fong, Donald S., Lloyd Aiello, Thomas W. Gardner, George L. King, George Blankenship, Jerry D.
Cavallerano, Fredrick L. Ferris, and Ronald Klein. "Retinopathy in diabetes." Diabetes care 27, no. suppl
1 (2004): s84-s87.
[7] Li YH, Yeh NN, Chen SJ, Chung YC. Computer-assisted diagnosis for diabetic retinopathy based on
fundus images using deep convolutional neural network. Mobile Information Systems. 2019 Jan
2;2019.
[8] Eftekhari N, Pourreza HR, Masoudi M, Ghiasi-Shirazi K, Saeedi E. Microaneurysm detection in fundus
images using a two-step convolutional neural network. Biomedical engineering online. 2019
Dec;18(1):1-6.
[9] Harangi B, Toth J, Baran A, Hajdu A. Automatic screening of fundus images using a combination of
convolutional neural network and hand-crafted features. In2019 41st Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (EMBC) 2019 Jul 23 (pp. 2699-2702). IEEE.
[10] Saranya P, Prabakaran S. Automatic detection of non-proliferative diabetic retinopathy in retinal
fundus images using convolution neural network. Journal of Ambient Intelligence and Humanized
Computing. 2020 Sep 15:1-0.
[11] Liu YP, Li Z, Xu C, Li J, Liang R. Referable diabetic retinopathy identification from eye fundus images
with weighted path for convolutional neural network. Artificial intelligence in medicine. 2019 Aug
1;99:101694.
[12] Masood S, Luthra T, Sundriyal H, Ahmed M. Identification of diabetic retinopathy in eye images using
transfer learning. In2017 International Conference on Computing, Communication and Automation
(ICCCA) 2017 May 5 (pp. 1183-1187). IEEE.
[13] Stene LC, Midthjell K, Jenum AK, Skeie S, Birkeland KI, Lund E, Joner G, Tell GS, Schirmer H: Hvor mange
har diabetes mellitus i Norge?. [Prevalence of diabetes mellitus in Norway]. Tidsskr Nor Laegeforen.
2004, 124 (11): 1511-1514
[14] Hapnes R, Bergrem H: Diabetic eye complications in a medium sized municipality in southwest
Norway. Acta Ophthalmol Scand. 1996, 74 (5): 497-500
[15] Porta M, Bandello F: Diabetic retinopathy A clinical update. Diabetologia. 2002, 45 (12): 1617-1634.
[16] Jeppesen P, Bek T: The occurrence and causes of registered blindness in diabetes patients in Arhus
County, Denmark. Acta Ophthalmol Scand. 2004, 82 (5): 526-530.

[89]

Irjmets - Analysis of Diabetic Retinopathy in Fundus Images Using CNN

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Irjmets - Analysis of Diabetic Retinopathy in Fundus Images Using CNN

Uploaded by

Copyright:

Available Formats

e-ISSN: 2582-5208

International Research Journal of Modernization in Engineering Technology and Science

ANALYSIS OF DIABETIC RETINOPATHY IN FUNDUS IMAGES USING CNN

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

Fig 3.1 Retinal fundus image

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

Fig 3.2: convolutional neural network

Fig 3.3: LeNet-5 Architecture

Fig 3.4: Neural Abstraction Pyramid

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

Figure 3.5: The original inception model used in Google Le Net

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

Sensitivity Specificity PPV NPV

Fig 4.1: Accuracy of one batch of images

False Negative False Positive

Fig 4.2: Accuracy of another Batch of images.

Fig 4.3: Accuracy of overall model.

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

www.irjmets.com @International Research Journal of Modernization in Engineering, Technology and Science

You might also like