Diabetic Retinopathy Detection using Deep

Learning
Seema Hanchate
Supriya Mishra Zia Saquib
Department of Electronics and
Department of Electronics and Sr. Vice-President, Technology
Communications
Communications Inovation and Department
Usha Mittal Institute of Technology
Usha Mittal Institute of Technology Jio Platforms Ltd
Mumbai, India
Mumbai, India Mumbai, India
smhanchate.umit@gmail.com
supriya94mishra@gmail.com zsaquib@gmail.com

Abstract—Diabetic Retinopathy (DR) is human eye illness doctors used fundus camera which takes the picture of veins
which occurs in individuals who have diabetics which harms and nerves which is behind the retina. The initial phase of
their retina and in the long run, may lead visual deficiency. Till this disease has no signs of DR, so it turns into a real
now DR is being screened manually by ophthalmologist which challenge to recognize it into a starting stage. For early
is a very time consuming procedure. And henceforth this task detection we have used the different CNN (Convolutional
(project) focuses on analysis of different DR stages, which is Neural Network) algorithms, so that doctors can start the
done with Deep Learning (DL) and it is a subset of Artificial treatment at the correct time.
Intelligence (AI). We trained a model called DenseNet on an
enormous dataset including around 3662 train images to In this paper the dataset which we are using for the
automatically detect the DR stage and these are classified into project is collected from “Aravind Eye Hospital” and it is
high resolution fundus images. The Dataset which are using is available on kaggle that is “APTOS (Asia Pacific Tele-
available on Kaggle (APTOS). There are five DR stages, which Ophthalmology Society)”. We compare the two CNN
are 0, 1, 2, 3, and 4. In this paper patient’s fundus eye images architecture that is VGG16 architecture and DenseNet121
are used as the input parameters. A trained model (DenseNet architecture, and showing the results of these two
Architecture) will further extract the feature of fundus images architectures.
of eye and after that activation function gives the output. This
architecture gave an accuracy of 0.9611 (quadratic weighted In recent projects and researches, AI models, and in AI
kappa score of 0.8981) to DR detection. And in the end, we are specially “Deep Learning” gives the most accurate outputs in
comparing the two CNN architectures, which are VGG16 finding hidden layers in various AI tasks, particularly in the
architecture and DenseNet121 architecture. field of medical image analysis [1]-[3]. Based on the deep
learning models which are classify diseases and support
Keywords—Deep Learning, Diabetic Retinopathy (DR), medical decision making and can improve the persistent
DenseNet121 Architecture, VGG16 Architecture, Dataset, consideration (extra care) [4].
Fundus Camera.
The remaining paper is organized as follows; Section II
I. INTRODUCTION includes the litrature reviews of the DR image classification.
Section III tells all about the dataset information. Section IV
DR is the most debilitating form of diabetes in which includes the Methodology of DL architectures. Section V tell
serious damage occurs to the retina and causes visual us the main result of this project. Lastly the section VI
impairments. It harms the veins inside the retinal tissue, concludes the paper.
making them spill fluid and contort vision. Alongside
maladies prompting visual impairment like, waterfalls and
glaucoma, DR is one of the most continuous diseases. There II. LITERATURE REVIEW
are five stages of DR that is 0, 1, 2, 3, and 4. In a particular topic it includes an overview of existing
approaches that employed “Deep Learning” for DR
The below table gives the overall details about DR
automatic early detection.
stages:
A. Development and validation of a deep learning
algorithm for DR automatic detection
Applied a deep learning to learn an algorithm for
automatically detection of DR. Deep learning has ablity to
program an algorithm itself because it is a computational
methods and learning from a large set of examples that
demonstrate the desired behavior. These techniques are uses
in clinical imaging. The EyePACS-1 included 963 images
from 4997 patients, the Messidor-2 had 1748 images from
874 patients. For the accuracy detection the algorithm had an
area under the receiver operating curve of 0.991 (EyePACS-
Each stages has its own symptoms and specific 1) and 0.990 (Messidor-2) [5].
properties, now from normal images doctors can not specify
the DR stages. Moreover existing methods for diagnosing are The automatic detection of DR is of vital importance, as
very inefficient because it takes very large time, due to which it is the fundamental cause of irreversible vision loss in the
the treatment may go the wrong way. To detect retino-pathy working age or young age of populace in the world. The

978-1-7281-7213-2/20/$31.00 2020
c IEEE 515

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on September 06,2023 at 08:02:35 UTC from IEEE Xplore. Restrictions apply.
 classification of DR images is very difficult even for trained microscope in which camera is attached and designed to take
clinicians. Therefore, using DCNN (Deep Convolutional the picture of the interior surface of the eye [13]. The fundus
Neural Network) for the classification of DR with an image was used to document the DR condition that is images
accuracy of 94.5% [6]. gave the clear picture for detection.
The clinicians are divided these DR into five classes
Currently, a novel DCNN, which plays out the which shows the stages of DR :
beginning time identification by recognizing all
microaneurysms (MAs), the first indication of DR, x No DR (class 0)
alongside accurately allotting names to retinal fundus x Mild DR (class 1)
pictures which had five classes. The architecture was tested
on kaggle dataset and got the output of 0.851 QWK score x Moderate DR (class 2)
and 0.844 AUC score. In the early stage recognition, the x Sever DR (class 4)
model showed the sensitivity of 98% and specificity of 94%
which shows the effectiveness of technique [7]. x PDR (Proliferative DR) (class 5)
This dataset contains many folders like train.csv,
An ensuring dataset fidelity by master verification of test.csv, train_images, test_images, and
class labels improves acknowledgement of unobtrusive sample_submission.csv. The below figure shows the
highlights and found that preprocessing with contrast information of folders:
limited AHE. Transfer learning on models from ImageNet
improve accuracies to 74.5%, 68.8%, and 57.2% (2-ary, 3-
ary and 4-ary) classification models, respectively [8].
Fig. 1: List of folders in dataset
Starting stage of DR can prevent this type of disease
with correct tratment. A new feature extraction method that
is Modified Xception Architecture has shown in the picture CSV (Comma Separated Values) file gives all the
for the diagnosis of DR disease. This method shows that information of image and it is in excel sheet. Train.cvs
modified deep feature extractor improves DR classification contains the fundus eye image name and its severity level
with an accuracy of 83.09% versus 79.59% when compared (class) and test.csv includes only the eye image name
with the original xception architecture [9]. because it is going to be test after training the CNN
architecture. Now the below picture is the sample image of
The target is to automate the discovery of DR and access fundus camera and it is the sample from dataset :
the seriousness with high efficiency, through a general
possible methodology. Explore the utilization of different
CNN architectures on pictures from the dataset in the wake
of being subjected to suitable image processing techniques.
The final results acquired through training. VGG16 gave an
accuracy of 71.7% whereas the same for VGG19 gave
76.9% and Inception v3 was 70.2% [10]

Sadly the specific identification of the DR stage is

famously precarious and requires expert human Fig.2: Sample image
understanding of fundus pictures. Right now an automatic The above figure shows all the nerves which is behind
deep learning based method for DR stage identification by the eye. In our dataset all the image have 224X224 pixels
individual photography of human fundus. The method can and 3 channels that is RGB channel and divided into five
be utilized as a method for early stage detection with classes. Dataset includes 3662 train images and 1928 test
sensitivity and specificity of 0.99 and QWK score is images (in below figure).
0.925466 on APTOS Dataset [11].

III. DATASET
The image data used in this research was taken from
dataset. The dataset which we used an open dataset that is
this dataset can be used by anyone, which is collected from
“Aravind Eye Hospital” which was easily available on
Kaggle 4th APTOS (Asia Pacific Tele-Ophthalmology Fig.3: Number of train and test images
Society) 2019 Blindness Detection. This dataset was largest Again the fig.4 includes the counting’s of all the classes.
available on publicly to pre-training our CNNs architecture Class 0 has 1805 images (number of people), class 1 has 370
or model. images (number of people), class 2 has 999 images (number
The dataset which we are using was provided with a large of people), class 4 has 295 images and class 3 has 193
amount of high resolution retina images taken under a variety images.
of imaging condition. The images which are provide in
dataset are recorded from fundus camera which provides
color fundus image of DR. A fundus camera is a low power

516 International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE 2020)

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on September 06,2023 at 08:02:35 UTC from IEEE Xplore. Restrictions apply.
 A. Deep learning framework for DR

Fig.4: Number of images in each class

A. ImageNet Fig.5: Deep learning framework

Our CNN architecture is pre-trained with ImageNet The above fig.5 is all about the framework of deep
dataset. The ImageNet dataset improves the accuracy of learning for DR.
CNNs model in our case it improve the accuracy of
DenseNet121 architecture. 1) Preprocessing : There are few steps which we have
to follow during the preprocessing :
The ImageNet dataset is a very large set of photographs
designed for developing the algorithms or models like a) Take an image as an input.
computer vision, AI (Artificial Intelligence), ML (Machine b) Apply preprocessing technique to highlight the
learning) and DL (Deep learning). The Challenges, models important features.
and algorithms etc, uses the subsets means that images c) Cropping and resizing of image.
which we want to train from the ImageNet dataset when they
have annual competition. d) Proper data cleaning and removing black images.
Based on the statics about the dataset recorded on the e) Rotation and mirroring ofimages to balance the
ImageNet there are 14 million different images linke dataset, if the dataset is imbalance.
animals, medical images, plant data, etc in the dataset. The f) Conversion to numpy array.
goal of developing the dataset was to provide a resource to
promote the research and development of improved methods g) Now use for traing or testing.
for computer vision, AI, machine learning and deep
learning. 2) CNN model : After preprocessing, next step is train
our CNN model or architecture. There are many CNN
IV. METHDOLOGY models or architetures are available in deep learning
As we know that DR detection problem is a primary methods to train the network.
cause of blindness. To overcome from this problem early
detection is the first concern. So for early detection we are 3) Medical report : Once we train our model, now we
using the deep learning architecture called “DenseNet will get the final report that is output of input image. It
121Architecture”. means if we put any unseen image as a testing it will give
the report of that unseen image.

International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE 2020) 517

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on September 06,2023 at 08:02:35 UTC from IEEE Xplore. Restrictions apply.
 B. Flowchart of our project :
DenseNet architecture is an advance version of ResNet
architecture. This architecture do not summation or add the
outcome of the features of the layer with the incoming
features but concatenate them.

DenseNet121 are broke into DenseBlocks, where the

dimension of the featurs remains constant or unchange
within a block, but the number of filters changes between
the blocks, these layers are called transition layer.

As shown in the above figure, the measurement of each

volume represents the sizes of the 2D that is its depth and
width, whereas the numbers on the top which provides the
features dimension. Here 32 is the growth rate of model.
The volume of each block of denseblocks increases by the
growth rate multiply by the number of dense layers within
that denseblock. Every layer is adding to the previous of
these 32 growth rate which is the new feature adding to it.
By doing all this,layers are increasing from 64 to 256 after 6
layers. Furthermore transition block performed as 1 X 1
convolution with 128 filters . 2 X 2 pooling with a stride of
2, resulting on seperating the size of the volume and the
number of features on half.
D. VGG16 architecture : The below figure shows the
VGG16 architecture. We do not use the ImageNet in this
architecture.

Fig.6: Flowchart

The fig.6 is full flowchart of our project, which uses the

ImageNet for better accuracy with DenseNet architecture.
For VGG16 architecture we don’t use ImageNet and we will
see the difference between VGG16 architecture and
DenseNet 121 architecture. As we saw in above figure the
flowchart is self-explanatory it includes preprocessing step,
show the dataset information, display the shape of the image,
using of quadratic weighted kappa and ImageNet and at the
end running the epochs and got the output.

C. DenseNet 121 Architecture: The below figure shows the

block diagram of DenseNet 121 architecture
DenseNets are increasing the depth or layer of DCNN. Fig.8: VGG16 achitecture
DenseNets exploit the potential of the network by reusing
the feature. For DenseNet121 Architecture, there is no need The input of conv1 layer is of same size (224 X 224),
to learn feature maps and requires fewer or lesser maps. wherever we see the input it is of same size and it is a RGB
image. The image is gone through a layers (multiple layers)
of convolutional layers, where the filters were used. The
padding of convolutinal layer input is the spatial resolution
is preserved after convolution that is the padding is one
pixel for 3 X 3 conv layers. Spatial pooling is perform by
five max-pooling layers, which follow some of the conv
layers. Max-pooling had over a 2 X 2 pixel window and it
had stride 2. Fully connected (FC) layers which is almost
the last layer, follows a stack of convolution layers (which
has a different depth in different architecture). The FC
layers have 4096 channels each.
E. Quadratic Weighted Kappa:
The quadratic weighted kappa is very useful when codes
Fig.7: DenseNet 121 architecture are ordered. Three matrices are included the matrix of

518 International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE 2020)

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on September 06,2023 at 08:02:35 UTC from IEEE Xplore. Restrictions apply.
 observed score, the matrix of expected scores based on
chance agreement, and the weight matrix. There are few
steps to calculate the QWK, which is as follows :
Step 1: Create a multiclass confusion matrix
(confusion_matrix) 0 between predicted and actual values.
Step 2: In step 2 each element is weighted. Predictions that
are further away from actuals are marked harshly than
predictions that are closer to the actuals (construct the
weighted matrix which calculates the weight between the
actual and predicted values).
Step 3: Create two vectors, one for preds and another for
actuals, which provides how many values of each rating
exist in each vectors (calculate value_counts() for each Fig.9: (b)
rating in preds and actuals).
Step 4: E is the Expected Matrix which is exterior product
of the two vectors calculated in step 3 (calculate E, which is
the outer product of two value_count vectors).
Step 5: Normalize both matrices to have same sum.
Normalize E and 0 matrix.
Step 6: Calculate numerator and denominator of wighted
kappa and return the weighted kappa matrix as 1-(num /
den).
V. RESULTS AND ANALYSIS
After done with the experiments, we got the experiment
results in which we show the accuracy of our project. We Fig.10: (a)
used two architectures for same dataset and see the
accuracies of each.

Architecture Dataset QWK Loss Accuracy

VGG16 Kaggle Not 0.7874 0.7326

used
DenseNet121 Kaggle 0.8981 0.1197 0.9611

As we seen clearly in the above table VGG16 is used

without ImageNet and QWK and DenseNet is used with
ImageNet and QWK. So without ImageNet VGG16 gives
the less accuracy and with ImageNet DenseNet gives better Fig.10: (b)
accuracy than VGG16. Now will see the accuracy and loss
graph of VGG16 and DenseNet respectively. The above figures fig.9(a), fig.9(b) and fig.10(a),
fig.10(b) are shown the accuracies and losses of VGG16 and
DenseNet121 architectures respectively where VGG16
architectures do not used ImageNet and DenseNet
architecture used ImageNet.

Fig.9: (a)

Fig.11: Multiple test images

International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE 2020) 519

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on September 06,2023 at 08:02:35 UTC from IEEE Xplore. Restrictions apply.
 REFERENCES

Now if we want to detect the DR severity for at a time [1] S. H. Kassani, P. H. Kassani, M. J. Wesolowski, K. A. Schneider, and
multiple images then it is possible to do. The above fig.11 R. Deters, ““Breast cancer diagnosis with transfer learning and global
shows the multiple images DR detection. The fig 12 shows pooling,” arXiv preprint arXiv:1909.11839, 2019.
the single test image which give the output of one image [2] S. H. Kassani, P. H. Kassani, M. J. Wesolowski, K. A. Schneider, R.
Deters et al, “A hybrid deep learning architecture for leukemic
also. blymphoblast classification,” arXiv preprint arXiv:1909.11866, 2019.
[3] S. H. Kassani, P. H. Kassani, M. J. Wesolowski, K. A. Schneider, and
R. Deters, , “Classification of histopathological biopsy images using
ensemble of deep learning networks,” arXiv preprint
arXiv:1909.11870, 2019.
[4] Xiaomin Zhou, Chen Li, Md Mamunur Rahaman, Yudong Yao et al.
"A Comprehensive Review for Breast Histopathology Image Analysis
Using Classical and Deep Neural Networks", IEEE Access, 2020
[5] Varun Gulshan, Subhashini Venugopalan, Rajiv Raman,
“Development and Validation of a Deep Learning Algorithm for
Detection of Diabetic Retinopathy in Retinal Fundus Photographs,”
JAMA. 2016;316(22):24022410. doi:10.1001/jama.2016.17216.
[6] Kele Xu, Dawei Feng, and Haibo Mi, “Deep Convolutional Neural
Network-Based Early Automated Detection of Diabetic Retinopathy
Using Fundus Image,” Received: 10 November 2017; Accepted: 22
November 2017; Published: 23 November 2017.
[7] Sheikh Muhammad Saiful Islam, Md Mahedi Hasan, and Sohaib
Abdullah, “Deep Learning based Early Detection and Grading of
Fig 12: Single test image Diabetic Retinopathy Using Retinal Fundus Images,”
arXiv:1812.10595v1 [cs.CV] 27 Dec 2018.
[8] Lam C, Yi D, Guo M, Lindsey T., “Automated Detection of Diabetic
VI. CONCLUSION Retinopathy using Deep Learning,” AMIA Jt Summits Transl Sci
Proc. 2018 May 18;2017:147-155. PMID: 29888061; PMCID:
As we know that the DR (Diabetic Retinopathy) is PMC5961805.
primary concern for the diabetes patients, and manually it [9] Sara Hosseinzadeh Kassani, Peyman Hosseinzadeh Kassani, Reza
took a long time to detect DR. So we developed a Khazaeinezhad, Michal J. Wesolowski et al. "Diabetic Retinopathy
architecture for automatic detection of DR, here we took Classification Using a Modified Xception Architecture", 2019 IEEE
International Symposium on Signal Processing and Information
two architectures to compare them that which architecture is Technology (ISSPIT), 2019
best at what condition. The two architectures are VGG16 [10] Anuj Jain, Arnav Jalui, Jahanvi Jasani, Yash Lahoti, Ruhina Karani.
and DenseNet121 and the accuracies are 0.7326 and 0.9611 "Deep Learning for Detection and Severity Classification of Diabetic
respectively. The QWK helped us to give the confidence of Retinopathy", 2019 1st International Conference on Innovations in
Information and Communication Technology (ICIICT), 2019
accuracy which we got from DenseNet architecture.
[11] R Borys Tymchenko, Philip Marchenko and Dmitry Spodarets, “Deep
ACKNOWLEDGMENT Learning Approach to Diabetic Retinopathy Detection”.
[12] Weiguo Fan, Edward A. Chandan K. Reddy, “A Deep Learning
We wish to express our deepest gratitude to “Dr. Zia Based Pipeline for Image Grading of Diabetic Retinopathy”.
Saquib” who helped us a lot for this project and lastly [13] Eswar Kumar Kilari, Swathi Putta. " Delayed progression of diabetic
thanked to our college to co-operate with us for completion cataractogenesis and retinopathy by in STZ-induced diabetic rats ",
of this project. Cutaneous and Ocular Toxicology, 2016
[14] N. Yalin, S. Alver and N. Uluhatun, ”Classification of retinal images
with deep learning for early detection of diabetic retinopathy disease,”
2018 26th Signal Processing and Communications Applications
Conference (SIU), Izmir, 2018, pp.

520 International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE 2020)

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY ROURKELA. Downloaded on September 06,2023 at 08:02:35 UTC from IEEE Xplore. Restrictions apply.