Identification of retinal ischemia for Medical

Pronouncement through fundus image and

Machine Learning Techniques
1 2
[Link] Srinivas Pasupula
Dept of CSE(AIML) Department of CSE (AI&ML)
BVRIT HYDERABAD College Vignana Bharathi Institute of Technology
of Engineering for Women Hyderabad, Telengana
veeramachaneniindu@[Link] [Link]@[Link]

4
3
Gill Santosh Kumar Dr. Cheekatla Swapnapriya
[Link] Mechanical Engineering Dept of CSE
Vignan's institute of information technology (A) Vignan's institute of information technology (A)
Duvvada Visakhapatnam ,AP, INDIA Duvvada, Visakhapatnam, [Link]
gillsantoshkumar@[Link] swapnachsp@[Link]

6
5
Dr.S NagaMallik Raj [Link]
Dept of CSE Department of ECE
Vignan's institute of information technology (A) Miracle Educational Society,
Duvvada, Visakhapatnam, [Link] Vizianagaram AP, INDIA
mallikblue@[Link] Razh1976@[Link]

countries face limited resources for implementing

nationwide screening programs. Innovations such as
Abstract - Retinal ischemia is caused by diabetes which
ultra wide field imaging with scanning confocal
affects small arteries including retinopathy. It is a
condition marked by the deterioration of the retina's ophthalmology, handheld mobile devices, tele-
[Link] worse progression of Retinal ischemia may ophthalmology for remote assessments, and artificial
lead to the complete blindness. In this case, early detection intelligence for automated detection and classification of
of diabetes and retinal ischemia is more crucial to detect diabetic retinopathy are revolutionizing screening
the lesions affected and provide the proper guidance and approaches and enhancing cost-efficiency. Furthermore,
treatment to the patients suffering from diabetes. The emerging research indicates that retinal imaging may
identification of RI can be done manually and help identify individuals at risk for cardiovascular
automatically. In the manual process, analysis and disease or cognitive decline, potentially broadening the
explanation is given by ophthalmologists and opticians. It
scope of diabetic retinopathy screening beyond
is a time taking process and expensive. In case of
automation process, early detection of RI can be done by preventing vision-threatening conditions.
automated systems using Artificial intelligence and deep Diabetes is a fairly common condition all over the
learning over traditional methods. Most of the research world. It is the most prevalent cause of blindness in
work has done based on early detection of RI by taking two persons under the age of 50. It is a systemic illness that
major facets fundus images and RI detection techniques. affects up to 80% of patients for more than ten years.
Many studies agreed that an early diagnosis may save
Keywords: Retinal ischemia, Fundus Images, Retinal the lives of 90 percent of diabetes patients. Diabetes
Imaging, Optical Disk, Arteries, Exudates builds an individual's risk of retinal ischemia) [1]. Blood
is provided to all parts of the retina by little arteries,
1. Introduction which are powerless against high glucose levels. At the
point when a high measure of glucose or fructose
In recent years, personalized screening intervals that collects in the blood, the courses start to disintegrate due
consider various risk factors have been introduced,
demonstrating strong cost-effectiveness. However, many
to this deficient oxygen conveyance to cells. Any
obstacle in these supply routes causes genuine eye harm.
At this stage, numerous retinopathy indications happen,
including as HMs, MAs, EXs, and bury retinal miniature
vascular anomalies (IRMA). PRI creates when new
abnormal veins structure in various areas of the retina. A
confounded occurrence of RI might bring about vision
disability [4].RI is a degenerative disease, and
identifying it at a beginning phase is basic for
safeguarding a patient's vision; this requires customary
(a)
checking. A robotized evaluating technique for RI can
(b)
assist with lessening the probability of complete visual
Figure2. (a) Retina without diabetes (b) Retina infected
impairment because of RI while likewise diminishing
with diabetes
the strain on ophthalmologists. A PC helped indicative
(CAD) framework is being created for RI separating
II. LITERATURE SURVERY
request to recognize a neural portion of the eye with
possible RI from a typical retina [5–7].
Recent surveys are saying that almost 40 million of
An excessive amount of sugar in your blood can cause
people in India are affected by diabetes and due to this
blockage of the little veins that support the retina,
cause, most of the people are suffering from RI. 90% of
removing its blood supply over the long run.
the RI patients can be saved by early diagnosis. The
Subsequently, the eye makes an endeavor to frame fresh
prevention and identification of RI can be done
arteries. Nonetheless, these fresh arteries don't develop
manually or automatically. Manual process is a time
as expected and can promptly spill. There are two sorts
taking process. In this process ophthalmologists must
of retinal ischemia [8] Early Retinal ischemia. In this
have an expertise in the work. In present scenario due to
more continuous kind, known as non-proliferative
the advancement in the ophthalmology, AI is giving
retinal ischemia (NPRI), fresh arteries don't create
expertise to identifying and reducing the significant
(multiplying). The dividers of the veins in your retina
diseases including Retinopathy caused by diabetes In
decay when you have NPRI. Minuscule lumps rise up
[11] creator suggested the answer for reviewing RI and
out of the more modest vessel dividers, once in a while diabetic macular edema (DME) together. Here the
spilling liquid and blood into the retina. cerebrum (optic analyst utilizes picture level oversight to display the
nerve). Figure1 shows the side effects for various phases inside relationship cross illness consideration
of RI. organization to grade the inward connection between
these two infections [12].Here the author introduced
CNN which embeds deep learning approach. In this
semantic segmentation algorithm is used for
differentiating the retinal images as the normal or
infected one. This will help ophthalmologists and
opticians to grade images as mild NDPR, moderate
NDPR and severe NDPR [13].
Figure1. (a) Signs of NPRI (b) Sign of PRI In [14]the researcher proposes support vector
Figure1. Stage of RI machine model in order to extract the various lesions of
retinopathy to automatically categorize the group of
There are a few guide detection techniques to come NDPR by taking 400 fundus images from publicly
across the retinal ischemia symptoms. Those are Visual available datasets. This study utilizes image
acuity test, dilated eye exam, Tonometry, Fundus enhancement approach and used HSV algorithms and V
camera and so on. In visual acuity test, a process transform algorithm for the detection of retinal ischemia
followed is similar like we check like our sight number. by using 400 images from Kaggle dataset. In this work
The dilated eye exams will be another technique in [15], author suggested radial inverse force histogram
which the eye RIops will to add so that it will enlarge technique to increase the image retrieval efficiency and
the size of the inner layer of eyes and so that we can able this technique can also be used to identify the RI and
to get the inner details. The figure1 represents the diagnosis the lesions of RI. In [16] author proposed deep
normal retina and retinal ischemia with different lesions learning technique which supports CNN. Linear SVM is
used to train CNN. SVM differentiated the retinal
fundus images as RI or Non-RI images. Author uses Eye
PACS dataset for his experimentation.
 In [17] researcher introduced automated the APTOS 2019 Blindness Detection dataset and the
process through Computer Assisted Diagnostic (CAD) Diabetic Retinopathy Detection dataset, which provide
system. This will automatically detect and classify labeled retinal fundus images for tasks like classification
different types of NDPR by taking optical coherence and severity grading. Researchers can access these
tomography (OCT) images using segmentation methods. datasets for free by adhering to the terms specified by
In [18], author developed RI detection technique using the dataset providers, such as proper citation and non-
DCNN. Unlike the old method. Those layers are trained commercial use. Additionally, Kaggle hosts
to attract desired features by taking Kaggle dataset competitions and discussion forums, allowing users to
provided for retinal ischemia detection. After merging share insights, approaches, and pre-processed versions
all the attributes metadata the DCNNs and SVM of the data. By leveraging these resources, researchers
classifier are used to grade different stages of RI and are can ethically obtain high-quality data for retinopathy
labeled from zero to four. Along with this researcher analysis while ensuring compliance with licensing and
developed machine learning algorithms and app named attribution requirements.
as Deep retina is embedded into ophthalmoscope. By
using this normal people can also check their fundus
images themselves. In [19], presented a Deep CNN is a
challenging application is medical imaging. In order to
do this public dataset RI1 and MESSIDOR datasets
were used. This approach follows three steps. Fine
tuning of layers of CNN, fine tuning of trained CNN in a
layer wise and extract features then apply this to SVM
[Link] [20] a Multiple Instance Learning (MIL)
(a) (b)
technique was introduced for the detection of RI and its
lesions by using three different datasets named Kaggle,
Messidor and DIARETDB1 which are publicly available
were used. CNN was used for image attributes
extraction and for classification of different lesions
affected by retinal ischemia was done by global
aggression [Link][21], the author designed and
introduced mobile application for the screening of RI (c) (d)
patients in real time with the help of Deep neural
network(DNN).

III. METHODOLOGY IMPLEMENTED

Retinal ischemia is identified by the regions suffered

due to damage called lesions occurred on a neural
portion of the eye. The impacted lesions are referred as (e)
micro-aneurysms, hemorrhages and exudates. Figure3 Figure 3. Different stages of RI: (a) Normal
below gives the classification of RI based on the retinal (b) Mild RI (c) Moderate RI
different types of lesions. (d) Severe RI (e) Proliferative RI.

3.1. Data Acquisition Retinal ischemia is a disease characterized by the

process of worse development of retina which causes the
This section discusses about the datasets of fundus different types of organs suffered includes micro
photographs which are publicly available datasets and aneurysms, intra retinal edema, hemorrhages etc.
private datasets for finding the retinal ischemia. These
public datasets are also known as benchmark or standard
datasets which are used for researcher for their These lesions assume a significant part in finding and
experimentation. The utilization of these datasets is analysis of retinal ischemia. In the acknowledgment of
totally depending on the type of problem and RI, different methods have been utilizing to analyze the
methodology of research. Kaggle offers a robust retinal injuries. The wide examination identified with
platform for acquiring datasets related to retinopathy, Retinal ischemia sores identification and diverse
particularly for research and development in medical division strategies displayed in the impending areas as
imaging and AI applications. Notable datasets include subpart. Figure 4 beneath shows the distinctive division
 and location methods of RI dependent on the sort of finished utilizing Circular Hough Transform (CHT).
retinal sores prompts RI. Limited area based dynamic form model is utilized for
Detection technique the recognized OD division.
OD & OC of OD
Segmentation In segmentation algorithm that detects optic disc
Technique boundary vessel origin pixel location. In this method,
Detection
during first step of the process fundus images green
Technique of
OC planes were resized and reconstructed morphologically
then from the reconstructed image, bright regions are
Supervised and extracted which are near to the main arteries. In this
Blood unsupervised Skill proposed strategy, six district based components and
Vessel Gaussian blender model classifier are utilized for
extraction arranging OD area as more brilliant OD regions and
Skill Mathematical
morphology skill non-OD areas. The best up-and-comer region for the OD
Segmentation and was characterized dependent on the splendid probable
classification of OD with most extreme vessel total and strength other
retinal features
competitor regions are picked dependent on separation
Detection
Detection from the centroid. The demonstration of OD and OC is
arts of MAs Technique of MAs
shown in figure 5.
&HMs

Detection
Techniques of HMs

Detection arts
Pixel based,
of Exudate machine and
mathematical
morphological
techniques

Figure 4. Different methods for Identifying RI

3.1. Techniques of OD & OC Segmentation

Optic disc (OD) is treated as a preliminary stage of optic

cup(OC), macula, fovea and vessel detection in the
retinal images. In existing research, morphological and
geometric arts were used for OD detection and machine
learning and pixel-based arts were used for OC Figure 5. Segmentation result of OD
detection. Figure 4 gives the block diagram of OD and
OC segmentation.

Pre-proliferative RI is likewise called as Non-

proliferative retinal ischemia (NDPR). The RI at its
beginning phase it is called as NDPR. This NDPR is
grouped into gentle, moderate and extreme the
programmed organ harm discovery plot, which
comprises of four successive tasks they are vessel
extraction and optic disk evacuation, preprocessing,
lesion recognition and post handling. This plan
distinguishes the various kinds of injuries regardless of
their attributes. The technique for identification and
division of optic disk. Retinal fundus pictures will be
standardized in the preprocessing stages and the
difference of the picture is upgraded too. Expulsion of
veins is finished utilizing morphology tasks. Then, at
that point the identification and confinement of OD was
\ Table5. Detection and Segmentation algorithms

Database Segmentation Classifi

Author Result
choosen method er used
Large
OD Sobel filter and
Xu Li et al., Ac.= 93.37% ,
DERIVE with adaptive small
(2010) [39] Se = 85.58%
threshold arteries,
SVM,
Ac.= 95.27%,
D Martin et al., RIIVE, grey level
ANN Se.= 69. 45%,
(2011) [40] STARE features
Sp.= 98.19%
Mathematical Large
M MFraz et al., RIIVE, morphology, and thin Ac.= 95.52%
(2012) [41] STARE aggregation arteries, Sp. = 97.23
operators
Ac = 93.24%
Linear combo Blood
T.V Yuen et al., STARE, (STARE),
of response Vesel,
(2013) [42] RIIVE Ac= 94.07%
lines SVM
(RIIVE)
S E Rajan et al., Blood
RIIVE SIFT Ac= 92.1%
(2014) [43] Vessel
Ac= 95.05% ,
L Zhang et al., Blood
DERIVE KNN Se= 78.12%,
(2015) [44] Vessel
Sp= 96.62%
Arteries,
R G Ramana et Ac= 95.36% ,
PCA and Root
al., DERIVE Se = 70.79%,
Gabour filter guided
(2016) [45] Sp= 97.78%
tree,
Figure 6: Segmentation result of MA the distinction
DERIVE, between a
J P Dash et al.,
CHASED binary and an Arteries Ac= 95.9
Before segmentation process, preprocessing is done. In (2017) [46]
B1 enhanced
which, original retinal image is converted into enhanced picture
image. Preprocessing is very much necessary before M Hasan et al.,
Ac= 95.7,
segmentation. In preprocessing, noise and non-uniform HMM, OR HMM, OR Arteries Se= 81.0%
(2017) [47]
Sp= 97.0%
illumination will be removed. After preprocessing, we Iterative
get an enhanced image. This image is further applied to method,
Soorya M et al., Glauco
segmentation process. After applying to the threshold, (2018) [48]
VEIR identification of
ma
Ac=97.01%
binary map of arteries will be generated. This technique bend points as
OD countour
is called post processing, used to remove false pixels.
The techniques utilized for discovery and division of
retinal veins are ordered into two types.1) Supervised 3.3 Techniques of MAs &HMs Detection
strategies, 2) Rule based techniques. In regulated
technique, determine the data class first and afterward An automatic analysis for microaneurysm and detection
calculation is applied to a picture to discover the range. progression of RI is done in two ways. One approach
In this the predefined calculation is applied to unique follows the image analysis based on road map to get
pictures for arrangement of picture pixels. microaneurysms changes. Second approach discusses
the pathological features by means of statistical analysis
Some of the Algorithms for detection and segmentation and pattern classification techniques. It is also verified
of arteries is shown in table 1. the pathological risk factors which are leading to the
microaneurysms (MA) turnover. The miniature
aneurysms discovery and its order framework utilizing
channel banks for the RI and its injuries recognizable
proof. This has three phases; first is highlight extraction
dependent on applicant locales for miniature aneurysms
utilizing fundus photos. GMM and SVM classifiers were
utilized for characterization of separated elements.
Picture examination procedures for the early location of
RI is been presented. As per this examination, various
phases of RI and various injuries including
microaneurysms, exudates and hemorrhages were
recognized utilizing retinal pictures. New approaches
were developed an automation system to diagnosis the [1] T. Kauppi, V. Kalesnykiene, J.-[Link] et al., “The
DIARETDB1 retinal ischemia database and evaluation protocol,” in
hemorrhages, arteries and to categorize the retinal
Proceedings of the 18th British Machine Vision Conference(BMVC
ischemia into various classes including severe NDPR, ’07), pp. 15.1–15.10, September 2007.
moderate NDPR, mild NDPR, normal. To distinguish [2] D. Kayal and S. Banerjee, “A new dynamic thresholding based
veins, shading difference of veins of retina and shading technique for detection of hard exudates in digital retinal fundus
image,” in Proceedings of the 1st International Conferenceon Signal
based division was presented. To recognize
Processing and Integrated Networks (SPIN ’14), pp. 141–144,
hemorrhages, thickness investigation and jumping box February 2014.
methods were created. Finally, Random forest approach [3] C. Agurto, V. Murray, E. Barriga et al., “Multiscale AM-FM
is used to classify hemorrhages and retinal arteries [20]. methods for retinal ischemia lesion detection,” IEEE Transactionson
Medical Imaging, vol. 29, no. 2, pp. 502–512, 2010.
[4] G. MahenRIan, R. Dhanasekaran, and K. N. Narmadha Devi,
3.4 Techniques of exudates detection “Identification of exudates for Retinal ischemia based on
morphological process and PNN classifier,” in Proceedings ofthe 3rd
In this context, in [62] reported study on RI lesion International Conference on Communication and SignalProcessing
(ICCSP ’14), pp. 1117–1121, IEEE, Melmaruvathur, India, April
classification based on artificial neural networks (ANN).
2014.
The lesions including microaneurysms, exudates and [5] C. Agurto, V. Murray, H. Yu et al., “A multiscale optimization
arteries classification and also identifies the grade and stage approach to detect exudates in the macula,” IEEE Journal
of the diseases using retinal images. In [21] author ofBiomedical and Health Informatics, vol. 18, no. 4, pp. 1328–1336,
2014.
introduced a new methodology for identifying of retinal
[6] A. F. Aqeel and S. Ganesan, “Automated algorithm for retinal
lesions and diabetic retinopathy and its classification image exudates and RIusens detection, segmentation, and
automatically using fundus photographs. This report was measurement,” in Proceedings of the IEEE International
done based on extraction of feature attributes of Conferenceon Electro/Information Technology (EIT ’14), pp. 206–
215, Milwaukee,Wis, USA, June 2014.
bifurcation point, area of exudates, non-segmented
[7] P. M. Rokade and R. [Link], “Automatic detection of hard
texture and area of arteries. In [64] author proposed exudates in retinal images using haar wavelet transform,”
automatic RI detection method and to classify the International Journal of Application or Innovation in Engineering&
lesions caused by RI is categorized by ANN classifier. Management, vol. 4, no. 5, pp. 402–410, 2015.
[8] [Link]
Different lesions detected are microaneurysms,
retinopathy/symptoms-causes/syc-20371611.
exudates, homogeneity and arteries etc. using fundus [9] [Link],[Link],[Link],[Link],[Link],[Link],L.
images Peng, and D. R. Webster, ‘‘Grader variability and the importance of
reference standards for evaluating machine learning models for retinal
ischemia,’’ Ophthalmology, vol. 125, no. 8, pp. 1264–1272, Aug.2018
[10] P. Vashist, N. Gupta, S. Singh, and R. Saxena, ``Role of early
Conclusion screening for retinal ischemia in patients with diabetes mellitus: An
overview,'' Indian J. Community Med., Off. Publication Indian Assoc.
Detection of RI at early stages is to reduce the retinal Preventive Social Med., vol. 36, no. 4, p. 247, 2011.
[11] X. Li, X. Hu, L. Yu, L. Zhu, C. Fu and P. Heng, "CANet: Cross-
ischemia at early stage; it’s a major cause of vision loss.
Disease Attention Network for Joint Retinal ischemia and Diabetic
Late treatment of RI is a state that happens in individual Macular Edema Grading," in IEEE Transactions on Medical Imaging,
who have diabetes, it harms the retina, the light sensitive vol. 39, no. 5, pp. 1483-1493, May 2020, doi:
lining at the back of the eye. This paper examined tests 10.1109/TMI.2019.2951844.
[12] L. Qiao, Y. Zhu and H. Zhou, "Retinal ischemia Detection Using
done by creators for the discovery of retinal ischemia.
Prognosis of Microaneurysm and Early Diagnosis System for Non-
This work will be valuable for specialized people and Proliferative Retinal ischemia Based on Deep Learning Algorithms,"
scientists who need to utilize the continuous exploration in IEEE Access, vol. 8, pp. 104292-104302, 2020, doi:
around here. The field of picture preprocessing on 10.1109/ACCESS.2020.2993937.
[13] E. V. Carrera, A. González and R. Carrera, "Automated detection
clinical pictures needs significantly more consideration,
of retinal ischemia using SVM," 2017 IEEE XXIV International
because of its significance for exact discovery of Conference on Electronics, Electrical Engineering and Computing
diabetic elements in diabetic screening calculations. (INTERCON), Cusco, 2017, pp. 1-4, doi:
Advancing artificial intelligence algorithms to provide 10.1109/INTERCON.2017.8079692.
[14] Ö. Deperlıoğlu and U. Köse, "Diagnosis of Retinal ischemia by
more precise detection and grading of diabetic
Using Image Processing and Convolutional Neural Network," 2018
retinopathy, even in challenging cases. Integrating real- 2nd International Symposium on Multidisciplinary Studies and
time retinal monitoring tools into mobile devices or Innovative Technologies (ISMSIT), Ankara, 2018, pp. 1-5, doi:
wearable technology for continuous assessment of 10.1109/ISMSIT.2018.8567055.
[15] S. Kimpan, N. Maneerat and C. Kimpan, "Retinal ischemia image
disease progression. Incorporating technologies like
analysis using radial inverse force histograms," 2017 International
Optical Coherence Tomography (OCT) and fundus Conference on Intelligent Informatics and Biomedical Sciences
photography into a unified diagnostic system for (ICIIBMS), Okinawa, 2017, pp. 266-271.
detailed retinal analysis. [16] Shikhar Seth &BasantAgarwal (2018) A hybrid deep learning
model for detecting retinal ischemia, Journal of Statistics and
Management Systems, 21:4, 569-574.
References
[17] A. ElTanboly, M. Ghazaf, A. Khalil, A. Shalaby, A. Mahmoud,
A. Switala, M. El-Azab, S. Schaal, and A. El-Baz, ``An integrated
framework for automatic clinical assessment of retinal ischemia grade
using spectral domain OCT images,'' in Proc. IEEE 15th Int. Symp.
[Link].(ISBI), Apr. 2018, pp. 1431_1435.
[18] Y.-H. Li, N.-[Link], S.-[Link], and Y.-C. Chung, ``Computer-
assisted diagnosis for retinal ischemia based on fundus images using
deep convolutional neural network,'' Mobile Inf. Syst., vol. 2019, pp.
1_14, Jan. 2019.
[19] X. Li, T. Pang, B. Xiong, W. Liu, P. Liang and T. Wang,
"Convolutional neural networks based transfer learning for retinal
ischemia fundus image classification," 2017 10th International
Congress on Image and Signal Processing, BioMedical Engineering
and Informatics (CISP-BMEI), Shanghai, 2017, pp. 1-11, doi:
10.1109/CISP-BMEI.2017.8301998.
[20] L. Zhou, Y. Zhao, J. Yang, Q. Yu, and X. Xu, ``Deep multiple
instance learning for automatic detection of retinal ischemia in retinal
images,'' IET Image Process., vol. 12, no. 4, pp. 563_571, 2017.
[21] S. Suriyal, C. RIuzgalski and K. Gautam, "Mobile assisted retinal
ischemia detection using deep neural network," 2018 Global Medical
Engineering Physics Exchanges/Pan American Health Care Exchanges
(GMEPE/PAHCE), Porto, 2018, pp. 1-4, doi: 10.1109/GMEPE-
PAHCE.2018.8400760.
[22] V. Gulshan, L. Peng, M. Coram, M. C. Stumpe, D. Wu,A.
Narayanaswamy, S. Venugopalan, [Link], T. Madams, J.
CuaRIos,R. Kim, R. Raman, P. C. Nelson, J. L. Mega, and D. R.
Webster,``Development and validation of a deep learning algorithm
for detectionof retinal ischemia in retinal fundus photographs,'' JAMA,
vol. 316,no. 22, pp. 2402_2410, Dec. 2016.