Chapter 75% of patients with longstanding diabetes and



Diabetic retinopathy is an eye
disease that is caused by diabetes. It is a medical condition where patients
with diabetes tend to show abnormalities in the affected eyes because of the
fluid leaks from blood vessels of the light-sensitive tissue at the back of the
eye (retina). Ophthalmologists recognize diabetic retinopathy based on features
such as blood vessel area, exudes, hemorrhages, micro aneurysms and texture.


1.1 Project Overview

In our project we focus on
diagnosis through Fundus photographs, which includes careful observation of
photographs taken with expensive equipment by highly trained clinicians. This
detection technique is very sophisticate and requires very specialized
clinician knowledge 1. Here we want to approach a computer vision method that
matches human performance.

For more information on the Diabetic
retinopathy (DR) and an analysis of how early detection of the disease can help
slow or even stop its spread, consult 2. For a comparative research paper on
studies of risk factors of DR, consult Yau et al. 3. Fluorescein angiography
is not the only a technique for diagnosis of DR; a extensive analysis on other
detection techniques to diagnose DR can be found in 4.

The general format of our method is
as follows. We took some sampled images to a tractable size as input images. These
images are preprocessed by normalizing and denoising technics and then we use a
convolutional neural network. After that, test images are passed through the
network and the model attempts to classify
the progression of DR.


1.2 Motivation of the Project Effort

For Diabetic Retinopathy (DR)
diagnosis there exist multiple techniques, an ocular manifestation of diabetes
that affects more than 75% of patients with longstanding diabetes and is the
leading cause of blindness for the age group 20-64 5. Researches shows that
it contributes around 5% of total cases of blindness. WHO estimates that 347
million of world population is having the disease diabetes and about 40-45% of
them have some stage of the disease. By seeing below image one can
differentiate between image produced by normal eye and DR eye 6.



                    (a)Normal Eye6                                                                (b)
DR affected eye6


There are various factors affecting
the disease like age of diabetes, poor control, pregnancy but Researches shows
that if we can detect DR in early stage of the disease progression to vision
impairment can be slowed or averted.

So the aim of our project is to
provide a automated, suitable and sophisticated model using image processing
technic we can detect DR at early levels easily so that damage to retina can be












Background Literature

Previous work has been done in using machine learning and various
models for automated DR screening. For development of our
method and result analysis, we have conducted a literature survey describing DR
features and past work done to detect DR.

Giri Babu Kande et al. 7 represented Segmentation of Vessels in
Fundus Images using Spatially Weighted Fuzzy c-Means Clustering an algorithm
for the extraction of Blood Vessels from Fundus images. They used a set of
linear filters sensitive to vessels of different thickness and orientation.
A vessel  detection methods
recently reported in the literature is simple and an experimental evaluation
demonstrates  excellent performance over global
thresholding. Their algorithm were expected to be applicable to a variety of
other applications due to its simplicity and general nature.

Faust et al. 8 approaches an algorithms for the automated detection of diabetic
retinopathy using digital fundus images which
provides a brief analysis of models that use explicit feature extraction to DR
screening. These studies are in the magnitude of their scope derived from less
than 400 total data points, the homogeneity of the dataset and the narrowness
of the explicit features extracted from the images.

Yuji Hatanaka et al 9 presented the improvement of automatic
hemorrhages detection methods using brightness correction on fundus Images. They indicates the importance of developing several automated
models for finding out the abnormalities in fundus images. The purpose of this
paper was to improve their automated hemorrhage detection model to diagnose diabetic
retinopathy. They represented a new method for preprocessing and false positive
elimination. They removed false positives by using a 45-feature analysis. To verify
their new method, they examined 125 fundus images, including 35 images with
hemorrhages and 90 normal images. The sensitivity and specificity for the
detection of abnormal cases was were 80% and 88%, respectively. These verified results
indicate that their new method may effectively improve the performance of their
diagnosis system for hemorrhages.

Vujosevic et al. 10 build a binary classifier on a dataset of 55
patients by explicitly forming single lesion features. The scope of this study is limited in that the dataset.

Wang et al. 11 used a CNN(LeNet-5 architecture) as a feature
extractor for addressing blood vessel segmentation. The model has three heads
at different layers of the convnet which then feed into three random forests.
The final classifier essembled the random forests for a final prediction
achieving an accuracy and AUC on 0.97/0.94 using a standard dataset for
comparing models addressing vessel segmentation

Lim et al. 12 where the authors represent building a
convolutional neural network for lesion-level classification and then use the
learned feature representations for image-level classification. This scope of
the study is limited in that the dataset which contains 200 images.

Clara I. Sánchez et al. 35 presented an automatic detection of diabetic
retinopathy exudates from non-dilated retinal images using mathematical
morphology methods. They showed the performance of a
system on a publicly available database which is independent. The performance
of the method on that dataset was comparable to that of human experts and with
the results obtained in previous studies. Their method shows retinopathy
screening programs a very fast solution to lessen the burden of screening
diabetes while maintaining a high specificity and sensitivity.









National Eye Institute provides a standardized description of the severity
class of DR patients (which are the classes that our classifier predicts).
There are four severity classes, the first three describe non-proliferative DR
(NPDR) and the last proliferative DR (PDR). The severity scales are
characterized through a progression of four stages14,15:

Mild NPDR – Lesions of micro-aneurysms,
small areas of balloon-like swelling in the retinas blood vessels.

Moderate NPDR – Swelling and distortion of blood
vessels, extensive micro-aneurysm, retinal hemorrhage, and hard

Severe NPDR – Various abnormalities, large blot hemorrhages, cotton
wool spots and many blood vessels are blocked, which causes abnormal growth factor

PDR – Growth factors induce
proliferation of new blood vessels inside surface of retina, the new vessels
are fragile and may leak or bleed, scar tissue from these can cause retinal















Fig: The severity scales of DR14

This is an ongoing problem on
kaggle16 which tries to develop a model for DR detection. Dataset is taken
from the challenge-data part. Data set consists of high resolution eye images
and graded by trained professionals in 5 classes(0-4) which is according to
below table and figure below that17.


Table: Class name descriptions



Class 0


Class 1

Mild DR

Class 2

DR Eye

Class 3

Severe DR

Class 4

DR Eye



(a) Normal eye                                                  (b) Mild DR eye                                     (c)Moderate DR Eye




Severe DR eye                                                                            (e)
Poliferative DR Eye


                                                 Fig: 5 classes(0-4) of DR affected






4.1 Hemorrhages

 The earliest sign of
retinopathy is small red dots in the superficial layers of retina. These are
termed as microaneurysms when they are small and depending on their depth
within the retina they are termed as hemorrhages.  This occurs because of the leakage of blood
vessels of retina and indicates mild retinopathy. But when macula edema
thickens within 2 disc diameters of the centre of macula this creates
microvascular changes and causes leaking of plasma components in the area. This
represents moderate type of retinopathy. Though hemorrhage is a hard work to
detect we need some preprocessing in order to get a noise free and bright,
contrast, enhanced image.

The steps including
preprocessing to detect hemorrhages are :

Resize the image into 512 x 512 pixels

Convert the RGB image into grey scale image.

Use Median Filtering to remove artifacts such as vignetting.

Equalise the image and enhance contrast by Histogram Equalization.

first we extract the green channel from the image because in this channel the
affected area is seen clearly and easy to identify. Then we apply median filter
with a radius of 8 pixels to create a background and remove the background from
original image. This produces an image with blood vessels and hemorrhages. Using
a vessel mask we remove the blood vessel which results in an image with
hemorrhages indicated.


                             Fig: Retina with
hemorrhages and exudates18


Exudates detection

method we have applied to detect exudates on human retina is inspired by the
work described in 35.  Since the data
set is of completely different characteristic as we have changed in various
sides. That is why we are going to describe every step and the reason behind
taking it. Here we need to mention that we have implemented some library
provided by 19. We have also used MATLAB version 2017a for this project and
this detection consists of the following steps:

       (a) Preprocessing the image.

       (b) Detection of Optic disc and other

 (c) Detection of exudates in terms of optic
disc and artifacts.

the preprocessing step first we extract intensity constituents from an image.
Here we are going to work with gray-scale images because exudates are mostly
visible in such images. We then apply Median Filtering for reducing the noise
and apply Histogram Equalization to enhance contrast and brightness. The
resulting image helps us to detect optic disc and accordingly exudates. This
works as input image. Exudates are high intensity values as well as optic disk.
Therefore in order to go for exudates detection we need to find optic disc and
then we need to differentiate between optic disc and exudates near and inside
the optic disc area. To do this we consider that optic disc is the largest and most
circular part in brightest portion of the image. We apply Gray Scale Closing to
remove blood vessels in the retina mostly in the optic disc area. Here we take
a flat disc shaped structure element and consider the radius is eight. We
threshold the image to binaries it and use the resulting image as a mask. Then
the mask is inverted by pixels before overlaying into the original image. We
then apply reconstruction by dilation was on the overlaid image. We threshold
the image and find the difference between the original image and the
reconstructed image by the algorithm. Consequently, high intensity optic disc
is detected and rests are removed.


this part we faced a big problem of this approach. At the beginning of the
process, vessels were removed by the Gray Scale Closing and reconstruction was
applied on the image created from the original image. Therefore we are going to
reconstruct vessels in the optic disc area. But we face a problem is that we
are not getting one big circular optic disc. Rather we are actually detecting
two or three big connected components in this step. To solve this problem we
applied an addition dilation of the final mask. As a result the independent
areas are connected together into a circular shape. Here we note that we have
already detected artifacts and other bright spots in the image. That is why if
we use too big dilation, it can lead to merge the optic disc with those areas.

the proper additional dilation we have considered a flat disc shaped structured
element with a radius of four. Since the optic disc and also some bright
artifacts are detected in this process, we have estimated for every component
of the mask in order to distinguish between the features some extra values.
These additional values are termed as scores. Thus we have,

                                    Score =
area  circularity3


we have some case to give attention. Since we have situation that the feature
rather than optic disc can become much larger than optic disc, we needed to
give circularity more importance. We take elements of size more than 1100
pixels as an optic disc keeping the rest as artifacts. Here we do not classify
small areas which can become exudates as artifacts. At this stage after optic
disc extraction and artifacts detection we are going to detect exudates. As
before, high intensity blood vessels are removed by Grey Scale Closing. Then we
go for to get a standard deviation image which shows the main characteristics
of nearly arranged exudates. The resulting image is being threshold by taking
the radius is six. We than remove the outside shape of the retina and fill the
holes by imfill(). We consider threshold to remove optic disc and artifacts.
Finally the result is achieved when we apply a threshold at a level 0.01
between the original and the reconstructed one. The produced exudates mask
image is overlaid into the main image to get a proper vision.






                                     ( a )


                                   ( b )

Exudates detection. (a) Original image (b) Exudates15




After all
the feature extraction has been done, now we are going to perform binary
classification. Here we have used deep neural nets with two input layers, a
total of three layers, one representing the output. For this we have created a
feed-forward backpropagation network (newff) . Here the terms ‘newfit’ is for
‘regression’ and ‘newpr’ is for ‘classification’. They together are called
‘newff’ the generic name which is still available and gives better output in
our classification20.
First of all we have created a two later feed
forward network. The first layer consists of three ‘transig’ neurons and the
second layer has one ‘purelin’ neuron. Thus we have,

                 net = newff(p,t,3,1,{‘tansig’,’purelin’});

Here, p is the matrix of input vectors and t is the matrix
for target vectors. For the inputs vectors we have used three components
namely, optic disk mask, artifacts mask and exudates mask.

Then the network is simulated and its output is plotted.
Thus we have,

                  y = sim(net,p);


need to mention here that the network is trained for 5000 epochs, train-  parameter goal is 0.01.


the training has been done a .mat file is created and further loaded to test
our datasets. Once loaded we are able to distinguish our image as a good one
and as a bad one. Then the corr2 library function is used to find the
correlation between four classes of images. The test image belongs to the class
with which it correlates most.

The related algorithm we took is based on feed forward neural network
described in 21





Result and Discussion


using 250 images as training dataset we get the results on classification. Here
we get results satisfactorily according to our analysis. The test process took
around 5-6 hours to run over the images. Here the whole data set is also run
for validation.

we have used a confusion matrix which determines the accuracy of our
classification and list the correct of them as ‘true positives’ or ‘true
negatives’ and incorrect of them as ‘false positives’ or ‘false negatives’.
This method also determines sensitivity, specificity, positive predictive
values and many more things. Here, we have taken plot confusion (targets,
outputs) that
returns a confusion matrix plot for the target and output and the plot is given
below :   


Fig : confusion matrix

the figure we see that our procedure gives 87.5% accuracy in the results.

are satisfied by our results although they provide slightly inaccurate findings
as they provide a pretty good approach based on morphological methods.




Future Scope


we have faced many problems dealing with the findings, feature we have picked
may seem rational but they are efficient according to our study. So possible
approach would be based on feature extracting. One can study further in order
to detect hemorrhage more specifically. 
We would expect such feature engineering to make the method better to
perform the whole 5-class classification perfectly. Moreover, one can be done
in exploring more nuanced data normalization and denoising techniques.






Our project is an analysis
of a model to identify the severity of DR from Fundus Photographs. Our method performed well in
comparison to other method. It
is a fact that better and accurate the diagnosis, the more exact will be the
treatment plan. So diagnostic measures should aim towards accuracy for an
effective treatment regimen. In our study we were able to establish a good
accuracy in the diagnosis results.