Diabetic Retinopathy Tracker System
Dafwen Toresa
1
, Hazlyna Harun
2
, Juhaida Abu bakar
3
and Nur Haslinda Abdul Hasan
2
1
Faculty of Computer Science, Lancang Kuning University, Pekanbaru, Indonesia
2
Data Science Research Lab. School of Computing, University Utara Malaysia, Malaysia
Keywords:
Early Detection, Diabetic Retinopathy, CNN, Automated Intelligence.
Abstract:
The examination of retinal images is compulsory for ophthalmologists to spot features of eye diseases. Mul-
tiple studies and clinical trials have reported the benefit of early detection and timely treatment in reducing
the risk of vision loss from Diabetic Retinopathy (DR) and decreasing the global burden of blindness. Active
screening for DR is important because most patients who develop DR have no symptoms until the very late
stages, and by then it is often too late for effective treatment. Multiple patient barriers to DR screening exist,
including poor access to care rural healthy deprivation, lack of time, high out-of-pocket expenses, insufficient
patient knowledge and awareness of DR, and lack of care coordination, especially among low-income popu-
lations, and ethnical minorities specifically in Malaysia. Hence, this project aims at developing an automated
intelligence system based on fundus image captured by using a Convolutional Neural Network (CNN) trained
by a Deep Learning (DL) algorithm. The system is named Diabetic Retinopathy Tracker system (DR Tracker)
assists medical officers/technologists in rural areas in the screening process based on raw fundus images cap-
tured. In this project, 200 images were used for testing, while 2000 images were used for training. The results
achieved optimum accuracy is 93.75%. This product innovation known as DR Tracker will contribute to an
early screening process by improving the performance accuracy for the detection of DR based on the fundus
image. By utilizing classification algorithms, such solutions have the potential to provide quick, accurate deci-
sion support at a low cost. It can benefit the patients in medically underserved areas that have limited numbers
of ophthalmologists and rare medical resources.
1 INTRODUCTION
In 2020, the number of adults worldwide with Dia-
betic Retinopathy (DR), Vision-Threatening Diabetic
Retinopathy (VTDR), and Clinically Significant Mac-
ular Edema (CSME) was estimated to be 103.12 mil-
lion, 28.54 million, and 18.83 million, respectively;
by 2045, the numbers are projected to increase to
160.50 million, 44.82 million, and 28.61 million, re-
spectively (Sasongko, 2017). The World Health Or-
ganization (WHO) predicted that by 2030, Malaysia
would have 2.48 million people suffering from di-
abetes (W.H.O., 2014). However, Datuk Seri Dr.
Dzulkefly Ahmad, the previous Minister of Health,
declared in 2019 that 3.6 million Malaysians have dia-
betes. This occurrence leads to Asia’s highest and one
of the world’s highest rates. In Malaysia, diabetes has
become a major concern. As a result, the Malaysian
government is paying close attention to this problem
and has devised a comprehensive plan and method to
improve diabetes prevention, treatment, and control
as a matter of urgency. The workload of repetitive
tasks, on the other hand, is quite high.
Diabetic retinopathy (DR) is a serious eye disease
associated with long-standing diabetes that results in
progressive damage to the retina, eventually leading
to blindness (Toresa et al., 2021). The disease does
not show explicit symptoms until it reaches an ad-
vanced stage; however, if DR is detected early on,
vision impairment can be prevented with use of laser
treatments.
Regardless of the type of diabetes, all people di-
agnosed with it require yearly retinal screening in or-
der to detect DR early and effectively treat it. Sec-
ond, some of these patients are based in rural areas
and could not visit an eye care provider. Thirdly,
as such follow ups are required for years together,
the attitude, and/or behavioral aspects negatively im-
pact the patients practice despite knowledge of con-
sequences. These issues can be solved with provision
of an automated imaging system within easy reach
of the patient (Avidor et al., 2020). Hence, there
Toresa, D., Harun, H., bakar, J. and Hasan, N.
Diabetic Retinopathy Tracker System.
DOI: 10.5220/0012446600003848
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 3rd International Conference on Advanced Information Scientific Development (ICAISD 2023), pages 187-191
ISBN: 978-989-758-678-1
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
187

has been an increasing interest in the development
of automated analysis software using computer ma-
chine learning/artificial intelligence (AI) for analysis
of retinal images in people with diabetes thus solving
at least some part of the problem (Zhuang and Ette-
hadi, 2020). Second, some of these patients live in
remote locations and are unable to see an eye doctor.
Third, as a result, follow-up is essential.
For years, the attitude and/or behavioral traits
have been a source of contention despite awareness,
have a detrimental influence on the patients’ practice
of repercussions Provision can help to resolve these
challenges (Sharif, 2020). The presence of an auto-
mated imaging device within easy reach of patients
(Abed et al., 1999). As a result, there has been a surge
of interest in the computer-assisted creation of auto-
mated analytical software. For the analysis of data,
machine learning or artificial intelligence (AI) is used
to retinal images in diabetics, therefore resolving at
least a portion of the issue.
2 PREVIOUS WORK
Proposed classification of the grade of non-
proliferative diabetic retinopathy at any retinal
image. In order to extract features, used by a support
vector machine, an initial image processing stage
isolates micro aneurysms, blood vessels and hard
exudates. To figure out the retinopathy grade of
each retinal image used by a support vector machine.
The evaluation was implemented in the software
MatlabR. Getting better results from the support
vector machine (SVM) than other machine learning
algorithms (Carrera et al., 2017).
Proposed that the analysis which used the deep
CNN method is extremely particular during an auto-
mated DR severity grading and also very useful dur-
ing a large-scale database. Feature enhancing meth-
ods like matched filtering and region growing are
combined with SVM and Neural networks and are uti-
lized for various classification problems of DR sever-
ity. There has been growth of Graphics Processing
Units (GPUs) because of the growth of a Deep CNN
based feature extraction method. As a result, deep
learning models for DR detection are popular among
several researchers. Deep learning ensures high accu-
racy and performance and is a powerful tool for DR
detection (Rajesh et al., 2022).
In their paper used a CNN method with a bi-
nary tree-based ensemble of classifiers to increase the
performance. The model trained could provide us-
able point-of-care diagnostic services for DR. The
model was so well equipped that it could even be used
with other medical applications. The developed ap-
plication could only diagnose images from a phone’s
gallery (Hagos et al., 9 01).
Another study, proposed a multi-categorical dis-
ease detector system by utilizing deep learning tech-
niques. This study consisted of STARE (structured
analysis of the retina) database and detected differ-
entretina features like lesion, optic disk, and blood
vessels. Thus, CNN was utilized for disease classi-
fication. The augmentation and normalization of im-
ages have been done to enhance the number of im-
ages to prevent the overfitting issue. The blood ves-
sels were extracted using the morphological dilation,
erosion, adaptive histogram equalization (AHE), and
CLAHE. The optic disk removal was done with the
aid of the Canny edge detector and thresholding of
the images. The exudates appear due to leakage of
blood vessels because of hypertension, diabetes, and
vein obstruction. So, these yellowish exudates were
mainly detected with Gaussian blurring and binary
thresholding. The classification stage encompassed
two main stages, training and testing stages. The
training phase consisted of determining the represen-
tative classes and attributes from the training dataset.
Further, the dataset was associated with the class to
which it bears a resemblance; then the network deter-
mined the disease category of the conforming image
(Rajan and Sreejith, 2018).
Proposed an automated system using a convo-
lutional neural network (CNN) to detect diabetic
retinopathy and its severity stage. The model was
trained on approximately 88000 labeled retinal pho-
tographs, and a quadratic weighted-kappa metric was
used to determine the classification efficiency. The
proposed system acquired 82% accuracy in detecting
the DR, 51% for assessing its current stage and a de-
cent kappa value of 0.776. Results depicted that the
effectiveness of the neural networkcan be used for a
complex medical problem (Kwasigroch et al., 2018).
3 METHODOLOGY
The proposed method is developed and presented
here. The flowchart of the processes is illustrated in
figure 1. All the processes involved is explained in
this section.
Figure 1: Methodology.
ICAISD 2023 - International Conference on Advanced Information Scientific Development
188

Process
1. User Input Fundus Image.
A dataset of 2000 fundus images of type 2 di-
abetic patients were obtained from local hospi-
tal in Malaysia which is Hospital Universiti Sains
Malaysia from Kerian and Kaggle website. The
fundus images obtained have been categorized
into Normal and Diabetic Retinopathy by op-
tometrist. There are 850 normal retina images;
and 1250 images of diabetic retinopathy that were
used in this study. All the images had no link to
the patients’ identities.
2. Preprocessing.
The collection included photos from patients of
various ethnicities, ages, and illumination levels
in fundus photography. This has an effect on the
pixel intensity values in the photos, causing exces-
sive variance unrelated to categorization levels.
The photographs were likewise high resolution,
requiring a large amount of RAM. The dataset
was downsized to 512x512 pixels, which kept the
detailed characteristics we were looking for but
reduced the dataset to memory size.
3. Training.
The CNN was initially pre-trained on 2000 im-
ages until it reached a significant level. This was
needed to achieve a relatively quick classification
result without wasting substantial training time.
The network was trained using stochastic gradient
descent with Nestrov momentum. A low learning
rate of 0.0001 was used for 8 epochs to stabilize
the weights.
Figure 2 shows the results of 2000 images vali-
dation accuracy which is 88.46% until reached final
iteration.
Figure 2: Training Progress.
4 RESULT AND DISCUSSION
In this project, there are 200 images that have been
used for the testing phase. In this table, there are show
us the image and the outcome the images which is ac-
curacy. Based on this table, the result achieved opti-
mum accuracy is 93.75%.
Table 1 displays the degree of accuracy of the
test-phase photographs. The best accuracy for test-
ing phase, which is 93.75%, was obtained using 200
images during testing and 2000 during training. All
of the photographs used in this project’s training and
testing phases are raw images. In the testing step, we
blend the images of normal and diabetic retinopathy
to determine the accuracy. The outcome demonstrates
AlexNet’s convolutional neural network’s good accu-
racy. The majority of the accuracy results for the test
photos are 100%, and the lowest is 80.83%.
5 EVALUATION OF DR
TRACKER SYSTEM
15 respondents were asked to engage in this survey in
order to undertake the system evaluation. As a result,
the developer has taken the initiative to request that
educators (CNN-AI) from UNIMAP, UiTM, and Po-
liteknik institutions examine the system. Respondents
assessed the system’s functionality and completed a
questionnaire that was printed and made available be-
fore reviewing the system. The cornerstone for the
system assessment was usability testing. Usability
refers to how well something works, how well it com-
pletes its goal, and how happy the user is with it over-
all. The key aims of this testing are to determine
whether or not the system’s end-users can understand
it, operate it effectively, and ensure that it has been
well-developed. Following the recruitment of respon-
ders, each of them was evaluated using a 12 questions
which were separated into three sections: Part A: Per-
sonal Details, Part B: System Efficiency, and Part C:
Overall Evaluation.
Figure 3: Interface of DR Tracker system.
Diabetic Retinopathy Tracker System
189

Table 1: Accuracy of the test-phase photographs.
Image Accuracy (%)
Image 1 100
Image 2 100
Image 3 95.45
Image 4 92.04
Image 5 100
Image 6 98.42
Image 7 82.31
Image 8 100
Image 9 99.00
Image 10 100
Image 11 96.98
Image 12 92.76
Image 13 100
Image 14 100
Image 15 80.83
Image 16 84.97
Image 17 81.99
Image 18 91.96
Image 19 100
Image 20 100
Image 21 100
Image 22 96.98
Image 23 90.95
Image 24 100
Image 25 93.53
Image 26 93.19
Image 27 94.79
Image 28 100
Image 29 100
Image 30 95.00
Image 31 100
Image 32 100
Image 33 88.01
Image 34 90.76
Image 35 97.17
Image 36 84.79
Image 37 84.09
Image 38 100
Image 39 99.97
Image 40 100
Image 41 94.79
Image 42 100
Image 43 100
Image 44 100
Image 45 86.97
Image 46 100
Image Accuracy (%)
Image 47 95.97
Image 48 94.28
6 CONCLUSIONS
The AI DR tool can aid the doctor with fundus picture
analysis, which helps to guide the next stages in the
patient’s therapy more swiftly. Furthermore, without
mydriasis, clinicians may attend to more patients that
require attention. Emerging healthcare technologies
place a premium on decreasing visits to eye special-
ists, lowering total treatment costs, and increasing the
number of patients seen by each practitioner. AI can
assist health care professionals in reaching their ob-
jectives. Though technology can help in the health
care industry, it should not be used to replace clini-
cians at this time. New advancements in the field of
artificial intelligence are bringing new opportunities
for running DR detection and grading systems.
Doctors will be able to detect the early stages of
diabetic retinopathy by utilising the DR Tracker sys-
tem. Experts in significant Artificial Intelligent and
Convolutional Neural Network have reviewed this
system. Based on this system review, the average
score is around 4, indicating that the system was well
created, even though certain features need to be added
for future development.
When considering the future of Diabetic
Retinopathy, we may conclude that a smart phone
gadget attached to a retinal image capturing camera
will be effective in diagnosing DR in any place.
This can aid in overcoming the value and time
constraints of traditional techniques. There are two
general approaches to smartphone-based automated
DR diagnosis. The first and most obvious option
is a web-based diagnostic, while the second is a
stand-alone independent smart phone software-based
diagnosis. An independent smart phone application
might allow patients in rural and distant places
to readily request an automated diagnosis of DR.
Another solution would be to employ a cloud-based
diagnostic, which would allow users to constantly be
up to current and have the most up-to-date informa-
tion. Furthermore, the DR Tracker system may need
to be enhanced by adding more features, such as a
box to show where it identifies diabetic retinopathy in
a retina image, the requirement to emphasise disease
infection in the screening system, and a looking fine
feedback to know whether the screening is in process.
ACKNOWLEDGEMENTS
I’d want to convey my heartfelt gratitude to every-
one who made it possible for me to finish this report.
I would like to express my heartfelt appreciation to
Prof. Madya. Ts. Dr. Nor Hazlyna, Dr. Ts. Juhaida
ICAISD 2023 - International Conference on Advanced Information Scientific Development
190

Abubakar and Nur Haslinda Abdul Hasan. A special
thanks to Miss Amiera Syazlin and Siti Naquiah for
their assistance and support in completing the project.
REFERENCES
Abed, M., Muhammed, L., and Toman, S. (1999). Diabetic
retinopathy diagnosis based on convolutional neural
network. J. Phys. Conf. Ser, (1):1–8,.
Avidor, D., Loewenstein, A., Waisbourd, M., and Nutman,
A. (2020). Cost-effectiveness of diabetic retinopathy
screening programs using telemedicine: A systematic
review. Cost Eff. Resour. Alloc, 18(1):1–9,.
Carrera, E., Gonzalez, A., and Carrera, R. (2017). Au-
tomated detection of diabetic retinopathy using svm.
In Proc. 2017 IEEE 24th Int. Congr. Electron. Electr.
Eng. Comput. INTERCON 2017, pages 6–9,.
Hagos, M., Kant, S., and Bala, S. (2019-01). Automated
smartphone based system for diagnosis of diabetic
retinopathy. Proc. - 2019 Int. Conf. Comput. Com-
mun. Intell. Syst. ICCCIS 2019, (2005):256–261,.
Kwasigroch, A., Jarzembinski, B., and Grochowski, M.
(2018). Deep cnn based decision support system for
detection and assessing the stage of diabetic retinopa-
thy. Interdiscip. PhD Work. IIPhDW, pages 111–116,.
Rajan, K. and Sreejith, C. (2018). In Proceedings of the
International Conference on Advances in Computa-
tional Mechanics 2017. Springer International Pub-
lishing.
Rajesh, K., Santhanam, A., Sridhar, M., and Mohan, D. J.
(2022). Diabetic retinopathy detection using convolu-
tional neural network. Int. J. Adv. Res. Sci. Commun.
Technol, pages 69–75,.
Sasongko, M. (2017). Prevalence of diabetic retinopathy
and blindness in indonesian adults with type 2 dia-
betes. Am. J. Ophthalmol, 181:79–87,.
Sharif, N. M. (2020). Performance of image enhancement
methods for diabetic retinopathy based on retinal fun-
dus image. ISCAIE, pages 18–23,.
Toresa, D., Azrul, M., Shahril, E., Hazlyna, N., Abu, J., and
Amnur, H. (2021). Automated detection and counting
of hard exuyears for diabetic retinopathy by using wa-
tershed and double top-bottom hat filtering algorithm.
W.H.O. (2014). Global health estimates 2014 summary ta-
bles: deaths by cause, age and sex, by who region,
2000–2012. World Heal. Organ.
Zhuang, H. and Ettehadi, N. (2020). Classification of dia-
betic retinopathy via fundus photography: Utilization
of deep learning approaches to speed up disease de-
tection.
Diabetic Retinopathy Tracker System
191