EAI Endorsed Transactions

on Pervasive Health and Technology Research Article

Glaucoma Detection Using Explainable AI and Deep

Learning
Najeeba Afreen1,*, Rajanikanth Aluvalu1
1
Chaitanya Bharathi Institute of Technology, Hyderabad, India

Abstract

INTRODUCTION: Glaucoma is an incurable eye syndrome and the second leading reason of vision loss. A retinal scan is
usually used to detect it. Glaucoma poses a challenge to predict in its nascent stages because the side effects of glaucoma
are not recognized until the advanced stages of the disease are reached. Therefore, regular eye examinations are important
and recommended. Manual glaucoma screening methods are labour-intensive and time-consuming processes. However,
deep learning-based glaucoma detection methods reduce the need for manual work and improve accuracy and speed.
OBJECTIVES: conduct a literature analysis of latest technical publications using various AI, Machine learning, and Deep
learning methodologies for automated glaucoma detection.
RESULTS: There are 329 Scopus articles on glaucoma detection using retinal images. The quantitative review presented
state-of-art methods from different research publications and articles and the usage of a fundus image database for
qualitative and quantitative analysis. This paper presents the execution of Explainable AI for Glaucoma prediction
Analysis. Explainable AI (XAI) is artificial intelligence (AI) that allows humans to understand AI decisions and
predictions. This contrasts with the machine learning “black box” concept, where even the designer cannot explain why
the AI made certain decisions. XAI is committed to improving user performance. To provide reliable explanations for
Glaucoma forecasting from unhealthy and diseased photos, XAI primarily employs an Adaptive Neuro-fuzzy Inference
System (ANFIS).
CONCLUSION: This article proposes and compares the performance metrics of ANFIS & SNN fuzzy layers, VGG19,
AlexNet, ResNet, and MobileNet.

Keywords: ANFIS & SNN Fuzzy layer, VGG19, AlexNet, ResNet, MobileNet, Fundus image

Received on 29 December 2023, accepted on 28 March 2024, published on 05 April 2024

Copyright © 2024 N. Afreen et al., licensed to EAI. This is an open access article distributed under the terms of the CC BY-NC-SA
4.0, which permits copying, redistributing, remixing, transformation, and building upon the material in any medium so long as the
original work is properly cited.

doi: 10.4108/eetpht.10.5658

Corresponding author. Email: [email protected]

*

intensive and making firm conclusions even more

complicated. About 80 million people worldwide have
1. Introduction
glaucoma. About 12 million people in India have
Globally, glaucoma is thought to impact 76 million glaucoma. However, as in many developing countries,
individuals, and by 2040, this figure is projected to rise to approximately 50-80% of glaucoma cases in India go
112 million. [15]. The human eye is highly susceptible to undetected. Several diagnostic tests are performed in real
diseases such as cataracts, glaucoma, and myopia, and life, such as visual acuity, retina, and tonometry, but these
over time these diseases become more complicated, and are extremely exhaustive and challenging for the patient
the visual acuity of the human eye gradually deteriorates. [14]. Automated glaucoma detection and tracking saves
In fact, preliminary detection is essential to rule out time and money while improving clinical processes and
complete sightlessness, dissimilar to other ocular diseases, making the most use of scarce resources. Early treatment
there is no accepted reference standard for the diagnosis and diagnosis of glaucoma would be achieved by an
of glaucoma, making data comprehension is time- automated system, preventing individuals from

EAI Endorsed Transactions on

Pervasive Health and Technology
1 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

irreversibly losing their eyesight. An integrated system potentials have been used as a result of the relentless
can make use of the expertise of many ophthalmologists advancements of deep learning. Deep learning methods
to produce findings that are precise and reliable. Figure 1 were especially used to image data obtained by digital
depicts the study of people who had had glaucoma for photography, optical coherence tomography, and other
several years. means [44]. Unlike more conventional machine learning
Funding opportunities exist at the National Eye Institute methods, deep learning-based systems don't need human
(NEI) of the National Institutes of Health (NIH), the intervention at each stage of the process, from feature
Indian Council of Medical Research (ICMR), and the All- extraction through model tuning. The fundus pictures
India Institute of Medical Sciences (AIIMS) in India. used as input are displayed in Figure 2.
Additionally, Glaucoma Research Foundation and L.V.
Prasad Eye Institute are some of the non-profit
organizations that provide funding for glaucoma research.
A key approach to combating the rising glaucoma
problem should be raising awareness of the disease's risk
factors and symptoms. People over the age of 60 should
have their eyes checked for glaucoma every one to two
years, those between the ages of 40 and 60 should
undergo screening every three to five years, and those
under the age of 40 should see a doctor every two to four
years to aid in the early diagnosis of glaucoma.

estimation of cases over years

120
estimation of cases over years 111.8
100

80 79.6
76
70
64.3
60

40

20

0
2013 2016 2020 2023 2040

Figure 2: Retinal Fundus image

Figure 1: Prevalence of glaucoma over the years in a) Glaucomatous image
millions b) Normal image
Automated glaucoma decision-making has received
Deep learning is a subfield of machine learning that
attention from numerous studies to help ophthalmologists
enables artificial neural networks to self-learn via
identify glaucoma early. The advancement of various
supervised, semi-supervised, and unsupervised learning.
technologies for the quick and precise identification of
In recent years, Convolutional Neural Networks (CNNs)
glaucoma has been a subject of focus of research.
have witnessed explosive growth in popularity (CNN). A
Artificial intelligence methods have been successful in
system of deep learning that generates features
identifying Glaucoma using visual components such as
mechanically. When used for picture recognition, it can
fundus photos and optic coherence tomography (OCT)
yield reliable results, according to numerous research. As
scans. Researchers were able to assess or classify the
a result, ophthalmologists may find a deep learning-based
condition using artificial intelligence to predict its severity
CAD system to be highly helpful in making earlier
and course. In Ophthalmology, where conventional AI
diagnoses in a quicker and more effective manner [41]. In
solutions have been used for a while, a big number of

EAI Endorsed Transactions on

Pervasive Health and Technology
2 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

recent years, CNN, RCNN, VGG, Resnet, and other sometimes viewed as a black box since it offers no
commonly used deep learning approaches for glaucoma explanation. By demonstrating how their intrinsic
diagnosis have been used. In our proposed model we will qualities of explainability and expressiveness aid in the
be working on different deep learning techniques such as construction of moral, justifiable intelligent systems, they
mobileNet, AlexNet, VGG19, ResNet along with XAI investigated the advantages of adopting logic techniques
techniques. for explainable AI. By summarising existing strategies in
Explainable AI technology has been instrumental in the literature, they concentrate on a single problem linked
aiding healthcare professionals in accurately and precisely to application of argumentation theory in medicinal
diagnosing glaucoma. By combining advanced Machine informatics.
Learning models used with medical imaging and Cyras et al. [4] have provided a fascinating review of AF-
ophthalmologic data, this new AI-enabled glaucoma based explanations, which use strategies and tactics that
detection is able to provide results that are not just more come from the field of computational argumentation to
accurate, but whose diagnosis can also be explained to a explainable AI models. The term "model" is used
much greater degree than traditional methods of extremely broadly in this study to pertain to a wide range
diagnosis. Additionally, Explainable AI can help doctors of categories, including planning, LP tools, decision
explore the reasons behind a clinical decision, be it at the assistance, and recommender systems. The poll begins by
individual or larger population level. Its ability to shine a outlining the various kinds of argumentation explanations
light on various levels of complexity related to situations and categorizing them as either intrinsic or post-hoc
such as glaucoma will no doubt be essential for improving complete and posteriori approximate explanations have
future diagnostics in the healthcare field. been created from the posteriori explanations. What the
model describes, and the reasoning framework employed
to complete the task is at the center of the survey.
2. Related Study Kashif Siddiqui, Thomas E. Doyle [5] have talked about
Maede Zolanvari, Zebo Yang [1] highlighted the issue how trustworthy AI in medicine is. In their study, a
with AI, which makes it difficult for researchers and framework was created to analyse the credibility of AI in
industry experts to explain the choices made by medicine in relation to common shareholders in the
sophisticated AI algorithms because they (as AI users) are creation of medical devices. Through the assessment of a
unable to fully comprehend the factors that go into these collection of explainable AI (XAI) techniques, they
"black boxes'" decision-making. They emphasized the investigated AI models. The four components of trust—
significance of integrating Explainable AI (XAI) into an explainability, verifiability, fairness, and robustness—
AI-based IoT system. Using statistical theory (TRUST), have been defined.
which is used in numerical data like security systems and To anticipate aggregates and clarify the predictions with
network data for IoT, they proposed an XAI model that regard to alterations in the microbiome's component
offers transparency. prompted by dissimilarity in phenotypic features, Anna
Erico Tjoa [2] has discussed that AI and ML models have Paola Carrieri and Niina Haiminen [6] have devised an
demonstrated outstanding performance across a range of approach called logical artificial intelligence (XAI).
disciplines. But because there is a great need for Inferring microbial signatures affiliated with each trait
accountability and openness in the medical industry, it is and aptly predicting a variety of phenotypes from the leg
necessary to provide explanations for the decisions and skin microbiome, including age, skin moisture, and,
predictions made by machines in order to support their shockingly, menopausal and smoking status, are just a
accuracy. They collected data from studies on the few of the outcomes of their study. This demonstrates the
interpretability of machine learning algorithms or ability of explainable AI. They also considered using skin
computer algorithms in general, categorized them, and hydration model on a second, independent cohort to
then used the same categories for interpretability in the examine the model's efficacy in terms of prediction and its
medical area. In particular, the category aims to provide capacity for explanation.
physicians and practitioners with a viewpoint on the usage Koji Fujita, Toshiki Shibahara [7] have identified
of interpretable algorithms that are offered in a variety of misclassified malicious activities using XAI. They talked
formats. about how utilizing machine learning; it is nearly
Luciano Caroprese[3] has discussed the necessity for a challenging to obtain 100% classification accuracy
fully transparent and explicable artificial intelligence because these classifiers would always produce false
(XAI), as opposed to artificial intelligence, which is positives negatives. Human confirmation is essential to

EAI Endorsed Transactions on

Pervasive Health and Technology
3 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

ensure that these false positives and false negatives are image into the polar direction framework to further
not actually happening. A characterization verifier, an enhance the division outcome.
artificial intelligence model has been developed for the Eduardo Pinos-Velez, María Flores-Rivera [12] have
purpose of the study, compares the classification results of implemented the ancillary tools for identifying and
a classifier with the clarification findings of any analysing human eye medical pictures to help in the
associated questions asked. provisional diagnosis of glaucoma. They focused on
Abeyagunasekera, S. H. P., Yuvin Perera [8] have biomedical image processing to identify the
proposed a strategy to detect COVID-19 patients from characteristics regarded most significant within pictures
photographs. “They've adopted a unified concept based on collected from the back of the eye in order to diagnose
the use of Intelligent Fake Sharp (XAI) devices to develop glaucoma, a disorder that predominantly affects the
the sensitivities of the assumptions made by physical statistics of the cup and the optical disc.
Convolutional Neural Networks (CNNs) with medical Dhaval Vaghjiani, Sajib Saha [13] They have created
images. This system makes it possible for several methods novel visualisation techniques that make it possible to
to be used in order to provide justifications for the comprehend the innate image elements that bestow to
assumptions made by Blackbox models. These include glaucoma identification at various CNN levels in addition
LISA, also known as Close by Interpretable Model Cynic to CNN-based methods for glaucoma detection. They also
Explanations (LIME), integrated slants, Anchors, and built a collection of interpretable notions to help them
Shapley Added substance Explanations (SHAP), which is better comprehend the contributing picture characteristics
based on the Shapley values.” involved in the illness detection process. A VGG16 model
Bryce Murray, Derek T. Anderson [9] have investigated a was used on the clipped fundus pictures. They have
human-machine interaction process. For the ChI, they utilised the visualisation approach suggested by Zeiler et
analyzed older low-level XAI techniques, which we then al. to better understand which picture aspects are
categorised based on how and where they fit into the most impacting CNN's predictions for glaucoma detection.
recent XAI jargon. Two case studies—one synthetic and Md.Tariqul Islam, Sheik Asif Imran, Asif Arefeen [14]
the other an unresolved real-world issue—were the have assessed convolutional neural networks' (CNN)
subjects of their study. The case studies highlighted the performance and discussed a novel strategy for
initial set of suggested actions in light of both events that identifying eight different forms of ocular diseases. They
humans can influence and act upon and those that we have put up a fresh, practical, and clever replacement for
cannot. existing techniques to detect early-stage optical disorders
Erik Cambria [10] have delved at XAI strategies that use using indicators of ocular pathology based on neural
natural language to convey what has been done, what is networks. They developed a method that can accurately
happening at the now, what will happen in the future, and identify eight disorders, including glaucoma. Their illness
provide the information upon which these happenings are detection technology has undergone real-time testing on a
founded (via text or dialogue systems). They examined 70 variety of images with various sorts of ocular disorders
XAI publications with natural language explanations that and is therefore supported by evidence of its viability in
appeared between 2006 and 2021 in major journals and real-world applications. Their study will significantly
conferences. With the use of three distinct layers— influence ophthalmology and the field of medical science.
definition of context, explanation construction, and Kaveri A. Thakoor, Sharath C. [15] models using
message generation—they developed a framework for convolutional neural networks (CNNs) to find glaucoma
examining the whole learning process, from the black-box in optical coherence tomography (OCT) images have been
model to the final user. They also compiled a list of each developed, and CNNs have been tested using thought
layer's properties and compared it to the main XAI sanctioning vectors (TCAVs) to identify which image
literature techniques. concepts are used by CNNs to produce predictions.
Huazhu Fu, Jun Cheng [11] have discussed that the Researchers also compared the TCAV outcomes with the
current glaucoma diagnostic techniques divide the optic eye obsessions of specialists to identify regular dynamic
disc (OD) and optic cup (OC) and rely on manually features used by both artificial intelligence and human
created visual features from fundus photographs. The trained professionals.
authors have produced an M-Net deep learning system Sadikul Alim Toki, Sohanoor Rahman [16] have
that simultaneously addresses the OD and OC division attempted to use deep neural networks to automatically
gives on a single platform with several names. They also classify photos of healthy and sick retinoid fundus. The
used a polar change, which converts the original fundus major objective of their study was to develop a system

EAI Endorsed Transactions on

Pervasive Health and Technology
4 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

that would combine CNN and deep learning to recognise, Gauri Ramanathan, Diya Chakrabarti [22] have created
extract, and evaluate disease-specific features in retinal and presented a technique to quickly allow patient
fundus pictures. This tool will facilitate early disease detection of glaucoma, cataract, and retinal disorders
detection, enabling patients to maintain good vision while using machine learning models. To complete the
preventing significant vision loss and blindness. They recommended assignment and achieve high-performance
showed how effectively CNN can separate exudates in efficiency, they carefully investigated and observed the
colour fundus images. many already-existing works. They compared the
Andres Diaz-Pinto, Adrián Colomer [17] have utilised a accuracy of various ML models.
poorly disposed model to develop a retinal image Parag Jibhakate, Shreshtha Gole [23] have talked about
synthesiser and a semi-managed learning approach for employing the VGG-16 and ResNet-50 machine learning
automated glaucoma evaluation on both a large unlabeled models to diagnose glaucoma early. These two transfer
data set and a small glaucoma-named data set. learning algorithms were compared to one another. They
Mhd Hasan Sarhan, M. Ali Nasseri [18] have talked about looked into how convolutional network depth affected
numerous machine-learning methods for analysing accuracy in big environments for picture recognition.
ophthalmic data. They looked at machine learning To aid in the diagnosis of glaucoma, Javier Civit-Masot
methods that had been put forth in the previous four years [24] developed a system that analyses photographs of the
for the diagnosis of eye problems. Anyone with inside of the eye (the fundus) using a trained and tested
glaucoma, AMD, or diabetic retinopathy may find the neural network. By using machine learning, they found
results of this poll very interesting. both the optic disc and cup. then transferred their
Anita Manassakorn, Supatana Auethavekiat[19] have knowledge to a pre-trained CNN. To increase final
assessed CNN's capability to identify glaucoma in OCTA detection and differentiate glaucoma-positive cases, the
images. They created GlauNet, a specially designed CNN findings of both methods were integrated.
model. And suggested GlauNet was capable of Chethan M, Chandrashekar Dasari [25] have developed a
distinguishing between eyes with and without glaucoma. machine learning model for glaucoma diagnosis with
They came to the conclusion that glauNet could be trained reliable estimations. By gathering datasets, performing
with a limited dataset and was resistant to image artefacts analysis in numerous instances, and reviewing the
and various camera projection settings. If training is done distinctions between clinical diagnosis and algorithmic
on actual data that has more variation, this robustness can features, they took a step and gave advice to the doctor
be increased even more. about which type of glaucoma develops in the human eye
Juan Carrillo, Lola Bautista [20] have developed models as well as other problems.
to identify glaucoma from photographs of the eye's Anuradha Pandey, Mrs. Pooja Patre [26] have discussed
fundus. Even though there have been a number of works how inaccurate the traditional image processing method
on automatic detection based on colour fundus pictures, was. “They extracted features like CDR and RDR from
they claimed that the key challenge was to give a precise images using image processing techniques, and then
estimation of the CDR. developed a computational tool classification was carried out using machine learning
for automatically detecting glaucoma in fundus models from the neural network, support vector machine,
photographs of the eye and proposed a novel method for decision tree, and K nearest neighbours. These models
cup segmentation that exhibits increased accuracy when were contrasted with deep learning models.
compared to existing techniques. Diping Song, Bin Fu, Fei Li [27] have suggested using
Alexandru Lavric, Adrian I. Petrariu [21] have described OCT and VF data together with a Deep Relation
in detail the creation of a sophisticated deep learning Transformer (DRT) to do glaucoma diagnosis. To
system that, when presented with pictures of the fundus of investigate implicit pairwise links between OCT and VF
the eye, can reliably identify glaucoma. They nurtured the information in both global and regional contexts, a novel
GlauNet deep learning system to improve early-stage ID deep reasoning mechanism is proposed. In order to
performance. Many AI glaucoma detection architectures improve the representation with complementary
were demonstrated, each one utilising ML and deep information for each modal, a carefully thought-out deep
learning. They found that their ingenious strategy was transformer method is built using the pairwise relations.”
about 99.05% accurate. According to the researchers, this Three sequential modules, based on reasoning and
is the highest degree of accuracy yet achieved in transformer methods, are created to extract and gather
identifying glaucoma using deep learning CNN crucial data for glaucoma diagnosis.
computations without motion learning.

EAI Endorsed Transactions on

Pervasive Health and Technology
5 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

Silvia Ovreiu, Elena-Anca Paraschiv [28] have noticed images. Both the entire retinal fundus image and a
that CNN performs well in the early diagnosis of a specific area of the image were examined.
number of diseases; they have lately been utilised in the Barros, D.M.S., Moura, J.C.C., Freire, C.R. et al [33]
ophthalmological area for the diagnosis of a number of have outlined the goal of assessing the most recent
eye disorders, including glaucoma. As a result, they algorithms put out by various parties and outlining the
suggested a novel technique for glaucoma diagnosis that crucial elements in the creation of an automated diagnosis
makes use of 201-layer DenseNet neural networks that are system. Additionally, they evaluated the potential of
densely coupled. They gave an overview of some new machine learning algorithms to automate the detection of
glaucoma detection work utilising deep learning glaucoma and other abnormal eye illnesses in their
algorithms and talked about the advantages and earliest stages.
disadvantages of the more advanced methods currently L. Li et al [34] have explored how effective the current
available. methods are at removing significant amounts of
Zhang et al. [29] have developed an AI order structure duplication from fundus pictures for the purpose of
that classifies glaucomatous and healthy eyes using OCT glaucoma identification, that may lower the dependency
data from the macula and the plate. Researchers in and accuracy of glaucoma detection. To overcome this
Guangzhou used a 19-layer CNN architecture that took drawback, they proposed the AG-CNN, a metric-based
into account issues such visual ID dropout, data convolutional neural networks (CNNs) for glaucoma
enhancement, flat reflection, asymmetrical rotation, and detection. To be more precise, we first create a sizable
random deletion. Despite the absence of an automated attention-based glaucoma (LAG) database.
glaucoma diagnosis, this method has demonstrated 96% “Shubham Joshi [35] has presented a computer-aided
accuracy when used in conjunction with image processing design framework to help in detecting glaucoma at an
techniques. suggesting that it may be utilised to identify early stage and in screening for and treating the disease.
early glaucoma. The major objective was to evaluate the picture inspection
Rakhmetulayeva Sabina, Syrymbet Zarina [30] have model for its efficacy in the early detection and diagnosis
suggested a strategy for merging and contrasting all of glaucoma and in the evaluation of other visual
outcome data for glaucoma definition criteria to help a illnesses. Great images have been used in conjunction
larger investigation to enhance feature discovery. The with advanced machine learning algorithms for glaucoma
method, specifically the convolutional neural network diagnosis.”
deep learning algorithm, was suggested as a means of Ratuja Shinde [36] have proposed developing a CAD
evaluating machine learning's potential. Using picture system that can automatically detect glaucoma. They
cropping and a CNN model, we looked for signs of validated input images using the lightweight Le-Net
glaucoma and identified the illness at its observable architecture and segmented using the U-Net CNN model.
stages. By use of the cutting-edge augmented spot method, they
Anshul Thakur [31] have built models that use convex have extracted the Region-of-Interest (ROI).
representation to identify clinically significant patterns of Ajitha S [37] has demonstrated the capability of the deep
glaucomatous vision loss and can detect glaucoma years learning model to find glaucoma from fundus images and
before symptoms appear. To detect the patterns of suggests that proposed system can assist ophthalmologists
glaucomatous vision loss, they conducted a thorough in a quick, accurate, and reliable diagnosis of glaucoma.
archetypal analysis and then projected visual fields over has proposed a model for detecting glaucoma using deep
the patterns. The picture produced by projections was learning techniques such as Softmax classifier, SVM
more reliable in identifying glaucomatous visual loss. classifier.
Hansi Gunasinghe, James McKelvie [32] have discovered Ana-Maria Tefan [38] Taking into account advancements
that current research only takes ROI or the full set of in deep learning, transfer learning, and artificial
picture attributes into account. To develop an enlarged intelligence's capacity to process retinal images,
feature set, they blended features from the ONH with researchers have examined the pros and cons of the most
features from the entire retinal fundus image. Using effective glaucoma ID approaches. We looked into AI
standard PC vision layout coordination and an analysis of systems based on highlight extraction, deep learning, and
existing pre-prepared deep learning models as component recursive learning to determine the proper sequence and
extractors, they determined which models performed the locate glaucoma in retinal images.
best for the task of glaucoma localization in retinal fundus Faizan Abdullah [39] has described the difficulties that
glaucoma brings in terms of image processing and

EAI Endorsed Transactions on

Pervasive Health and Technology
6 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

machine learning, which will be able to spot gaps in results obtained, the proposed keen framework for the
existing studies. They gave a thorough discussion of the primary task is trustworthy, shows high accuracy,
many methods currently in use for using machine learning produces stable expectancies, and necessitates fewer
to identify and classify glaucoma from fundus pictures. preparatory bounds.
They gave an overview of computational methods for Omer Deperlioglu [44] has put out a hybrid approach that
diagnosing glaucoma illness. They discussed cutting-edge combines deep learning and image processing with
supervised and unsupervised learning techniques and Explainable Artificial Intelligence (XAI) to ensure
came to the conclusion that, in terms of analytical reliable diagnosis of glaucoma utilising enhanced
performance, supervised learning techniques coloured fundus image data. An explicable convolutional
outperformed unsupervised techniques. neural network used the augmented picture data for the
Lauren J. Coan,[40] a thorough evaluation of the pros and diagnosis (CNN). The Class Activation Mapping (CAM)
cons of using artificial intelligence-enhanced glaucoma method used by the XAI enabled heat map-based
localization frameworks that generate and use segmented justifications for the CNN's visual interpretation. The
fundus images. They highlighted the two approaches, public datasets such as Drishti-GS, ORIGA, and HRF
provided a comprehensive overview of the two fundus were used to test the hybrid solution's performance over
imaging systems utilised in state-of-the-art artificial the course of twenty classification tries.” The ORIGA-
intelligence-powered glaucoma location strategies, and Light dataset has the greatest mean values when taking
analysed the development of AI-powered glaucoma the performance evaluation into account. Except for the
discovery procedures while praising research gaps. The XAI contributions, Efforts were also made to create a
review was made more accessible to ophthalmologists by straightforward solution that employs deep learning in
including definitions of important artificial intelligence conjunction with more conventional image processing
terms relevant to AI-enabled glaucoma detection. methods.
Mamta Juneja,[41] recent research has suggested that “Diping Song [45] has claimed that by evaluating data
even if trailblazing architectures like Xception, Inception, from several tests, including peripapillary optical
ResNet, DenseNet, VGG, etc., produce satisfactory results coherence tomography and visual field (VF) testing, deep
on clinical classification tasks, they are not yet ready for learning has had remarkable success in the diagnosis of
deployment. Hence, an Xception-based architectural glaucoma. However, due to a number of limitations,
recommendation with 85 layers was implemented, and its implementing these created models to clinical practise is
efficacy as a screening tool was confirmed by testing on still difficult. To improve the glaucoma indicative
real-time data. They introduced the CoG-NET model, performance of the first OCT technique, they established
which uses 85 layers of deep neural networks and relies a cross-modular refining system to communicate data
heavily on separable convolution, to boost classification across the OCT&VF and OCT teams.” In order to direct
accuracy. element extraction from OCT inputs, they created a one-
Bingyan Liu [42] has released a new, self-contained of-a-kind AFR module for model preparation. Several of
model predicated on flawed thinking about how to more this early research used a matched dataset to test out the
precisely and broadly complete the tasks of separating the proposed strategy.
optic plate and cup and testing for glaucoma. We use an “Kim et al. [46] proposed a deep learning-based CAD
effective division and ordering network and the latest system for treating glaucoma. Gradient activation maps
developments in solo area variation on the division similar to CNN were used to locate and classify the input
organization's result space to solve the area shift problem. images. With fine-tuned weights and a 32-batch learning
More precise and reliable glaucoma screening predictions batch size, the final layer of the VGG model was modified
may be made through the usage of grouping and division for training. Finally, with an accuracy of 91% after 80
organisations to carry out glaucoma screening activities. iterations, training was conducted on 50% of the dataset
de Zarza` [43] has unveiled a sophisticated glaucoma and testing on the remaining 50%.”
detection system that operates automatically. “The Bala Prabhakar [47] argued that the increased likelihood
method is built on a three-step process based on forms of of conventional techniques' misdiagnosis by currently
EfficientNet, a proposed group of models discovered used regular systems for discovering solutions makes
using NAS that achieves convincing accuracy on more advanced methods, such as the use of manmade
Imagenet, producing consistent results that outperform the reasoning, possible (simulated intelligence). The authors
benchmark techniques, and applying Move Gaining from updated their readers on the numerous AI strategies
Imagenet to the specific task at hand. According to the utilised in glaucoma diagnosis and evaluation, with a

EAI Endorsed Transactions on

Pervasive Health and Technology
7 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

particular emphasis on how artificial intelligence (ML) optic disc. For automatic disc localization, a order-based
may be used to expose software as a medical device semi-automatic ground truth generating technique was
(SaMD) in correct diagnosis or early illness detection. also used. Although this method's efficiency was
Tae Joon Jun [48] has disseminated an Adaptive determined to be higher, it was still unable to shorten the
Positioning Convolutional Neural Network (TRk-CNN) computation time.
that performs very well when classes of images to be Juneja M et al. [53] formed the Disk Cup Glaucoma
identified share a strong level of similarity. They Network (DC-Gnet) using extensive knowledge of
separated the glaucoma image dataset they used to glaucoma detection in retinal fundus photographs. After
evaluate TRk-CNN into three groups: normal eyes, eyes first isolating the blue, green, and red channels from the
with a glaucoma suspicion, and eyes with glaucoma. data images, pre-handling was used to block out the
Yue Wu [49] has claimed that the conventional abnormalities. Afterwards, using the refined technique,
approaches of assessing disease activity are effective at the optic plate and cup were separated. Using the low-
capturing the structural and functional changes in passpass channel, we were able to lessen the visibility of
glaucomatous eyes. To identify glaucoma early and more the images' high-recurrence components. The connected
precisely, a novel technology is still required. Clinical shroud of the optic plate and optic cup was used to
diagnostics for glaucoma, such as optical coherence ascertain crucial aspects like the Second rate Unmatched
tomography, visual field, and intraocular pressure testing, Nasal Common (ISNT), the Circle Mischief Likelihood
were discussed in detail. “New technologies and existing Scale (DDLS), and the CDR.
clinical tests for detecting glaucoma disease activity were Henry Shen-Lih Chen [54] “The thickness of the retinal
summarised, including the visual field (VF) test, nerve fibre layer (RNFL) may be evaluated and
intraocular pressure (IOP) test, and optical coherence abnormalities in the RNFL can be recognized using
tomography (OCT).” spectral domain optical coherence tomography (SD-OCT)
Ying Xue [50] has claimed that a single feature is images.” After gathering this data, an algorithm for early
insufficient to explain the course of glaucoma alone and glaucoma diagnosis using colour fundus pictures is
that more accurate and thorough diagnostic techniques developed using deep learning (DL). They developed and
must be developed immediately. In order to categorise tested a DL method for regenerating RNFL thickness
glaucoma into four severity categories, in light of distribution pictures utilising thickness information
intraocular pressure (IOP), corneal fluorescein extracted from fundus images in a manner akin to OCT
angiography (CFP), and visual field testing, they measurement.
suggested a multi-highlight profound learning (MFDL) Andres Diaz-Pinto [55] has claimed that the handmade
strategy (VF). They developed a coarse-to-fine approach features based on segmentation used by existing
to determining glaucoma severity that involves screening, algorithms for automatic glaucoma evaluation rely on
localization, and characterisation. fundus pictures. The analysts automated the glaucoma
Supiksha Jain [51] has confirmed that automated assessment procedure by using fundus photos and five
glaucoma diagnosis relies heavily on the extraction of the distinct ImageNet-arranged models (VGG16, VGG19,
cup from retinal fundus pictures. They developed an InceptionV3, ResNet50, and Xception). Photographs from
efficient method for detecting glaucoma by employing a an unanticipated data collection were used to test the
General Ill-disposed Organization (GAN) trained on the method. they discovered that the fine-tuned model
Rider Manta-Beam Rummaging Enhancement (Rider performs competitively.
MRFO), in which the optic plate division process is Shaswat Singh [56] has proposed a segmentation
carried out by Fluffy Nearby Data C-Means grouping technique utilising Convolutional Neural Networks
(FLICM bunching), and the vein discovery process is also (CNN) was proposed. An optic cup and plate separation
carried out by starting. To improve the detection of was used because to the sophisticated convolutional
glaucomatous pictures, several important features were neural network layout of the Glaucoma Organization (G-
extracted. Net). Accuracy in division between the optic plate and
Bajwa M. N et al. [52] developed a significant cup was 95.8% and 93.0%, respectively. for the presented
improvement in diagonising glaucoma from retinal fundus technique, which functioned with two networks operating
images. At first, the optic plate and the fundus picture simultaneously.
were isolated and constrained using Areas with WangMin Liao [57] has put forth a unique clinically
Convolutional Brain Organization (RCNN). After that, interpretable ConvNet design for correct glaucoma
Deep CNN was used to classify glaucoma in the retrieved diagnosis as well as for more opace interpretation by

EAI Endorsed Transactions on

Pervasive Health and Technology
8 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

emphasising the different regions that the network created an online GUI platform that allows users to find
recognises. They obtained a highly accurate clinically the eye defect.
interpretable diagnosis result. According to their Guangzhou An [64] has created a classification model for
experimental findings, the suggested EAMNet can glaucomatous optic discs using objective machine
segment the optic disc and make an accurate diagnosis of learning, and it provides the classificatory criteria to help
glaucoma (0.88 AUC) (0.9 Adisc and 0.278 Edisc). with clinical glaucoma therapy. Each eye's data was
Fengze Wu [58] has suggested a method for extracting the retrieved along with the patients' background information,
matching VCDR, finding, including a method for yielding a total of 91 parameters. The request models
mechanically separating OD and OC in fundus pictures. (GBDT) were developed using computerised reasoning
Our approach relied on Detectron2's Mask R-CNN to classifiers like the brain organisation (NN), guileless
perform object identification and instance segmentation. Bayes (NB), support vector machines (SVM), calculated
To differentiate between OD and OC, they developed two choice trees (ADT), and an examination investigation was
distinct models. They measured the success of the conducted.
initiative utilising DSC for OD/OC segmentation and M.Shanmuga Eswari [65] has created a technique that
MAE for VCDR. uses sophisticated machine learning to detect glaucoma in
N S Jeya Shyla [59] has proposed a new method in this diabetic individuals. They used a Bayesian optimization
work to increase classification accuracy with less time support vector machine to predict glaucoma with 96.6%
features. The ELM classifier is used to extract many accuracy using a local real-time dataset of diabetic
features for high-precision glaucoma detection. Numerous individuals.
variables, including statistical features, HoG, and DWT, Narmatha Venugopal [66] has used the PH-CNN model to
are thought to have an impact on ELM training. create an automated model for classifying Glaucoma
Tomaz Ribeiro VianaBisneto [60] has created a GAN- images. This model works through the stages of
based, ordered-variety-file-based automated glaucoma classification, feature extraction and reduction. Discrete
detection method. We used a GAN to segment the OD Wavelet Transform is used in the initial stage of the
from the rest of the retinal picture. As the general feature extraction process (DWT).
characteristics of concrete objects serve as a unifying Wheyming Tina Song [67] has developed an effective,
principle for coordinating pixels, the semantic dependable, Moreover, automated methods for addressing
segmentation produced by GANs yields optimal markings the challenges of manual glaucoma recognised proof are
on various images. pricey and time-consuming, and present mechanised
Deepak Parashar[61] has said that early glaucoma glaucoma location procedures lack either better execution
detection is crucial for preventing permanent visual loss. or any quantitative heartiness testing approaches. They
A brand-new FAWT-based technique for the automatic demonstrated the potential to replace the currently used
classification of glaucoma phases was also provided. manual diagnosis of glaucoma, and they provided a
FAWT deconstructed components have yielded a variety framework with clear mathematical notations and
of entropy and FD properties. They discovered that comprehensible graphics to make it easy for others to
retrieved features from FAWT are helpful for detecting confirm our findings.
glaucoma. LDA has also been used to rank the features Ali Serener [68] employed ResNet and GoogLeNet deep
and reduce their dimensionality. neural network classification methods after performing
Yasmeen George [62] has implemented a full-blown DL image pre-processing (including RGB channel histogram
system for detecting glaucoma and calculating VFI. The equalisation for each image). Approximately one million
model is prepared on 3D volumes using data from three images from the ImageNet collection were used to
different sources: the initial 3D-OCT block, data initialise the model, which was then refined using new
generated during training using 3D graduate CAM imagery. The histogram evening out approach was used as
heatmaps, and data from the training dataset itself. By part of the pre-handling process, which included isolating
utilising the abundant main data inborn in the high-goal the blue, red, and green components of each image before
3D OCT forms under the supervision of the graduate applying it separately to each channel.
CAM consideration map, the model demonstrates superior Yunzhe Sun [69] has proposed a GAN based Domain
performance when compared to gauge models and Adaptation Method for glaucoma detection. They offer
element-based ML approaches. comprehensive approach to address the issue of a
Shamia D [63] has suggested a deep CNN network for the significant homogeneous domain change. They said that
early diagnosis of glaucoma and other eye disorders. They looking at the optic plate/cup and its surroundings was the

EAI Endorsed Transactions on

Pervasive Health and Technology
9 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

key to tracking down glaucoma. To bridge the gap 589 healthy subjects and 837 glaucoma sufferers, they
between the original and final spreads, they employed a increased classification performance using an 18-layer
crafty recreating accident that not only uses raw data but CNN (the accuracy was 98.13% in this study).
also maintains the original space photographs' visual
integrity. Experiments performed on different open-and-
2.1. Comparison of Different Models for
shut name datasets reveal how the method has the
Diagnosis of Glaucoma
potential to enhance the future accuracy of glaucoma
diagnosis. In this study, there are different collected papers related to
Sreng et al. [70] suggested a two-stage, automated Glaucoma detection on different ML, and DL models.
glaucoma architecture that uses Multiple Deep Based on how these works were effective in the detection
Convolution Brain Networks (MDCNN) in place of of glaucoma an analyzed study of their statistics, merits,
encoders to ease the burden on ophthalmologists and the and demerits are mentioned. Explainable AI has been
DeepLabv3+ optic disc region. another hot research topic in recent years and there will be
Rui Fan [71] We looked into the transferability and further analysis of it in our research paper. Compared
interpretability of the state AI approach ViT for using with ML models Dl models work well in analysing
fundus pictures in glaucoma diagnosis. Numerous glaucoma at a beforehand stage.
experiments have shown that ViT has a high-performance To facilitate readers checking the literature,
level on external test sets made up of fundus photos from Table 1 shows summarization and analysis, which lists the
people of various racial and ethnic origins, including approaches for detection of glaucoma at an early stage, its
those of Chinese, Japanese, Spanish, African, and parameters, advantages, and disadvantages.
European heritage.
Raghavendra et al. [72] It was suggested that retinal
fundus images be run via a Convolutional Neural
Network (CNN) model to ensure accurate glaucoma
CAD. With a total of 1426 images, split evenly between
Table 1
Table of Collected Papers

S.no Author’s Model Parameters Merits Demerits Dataset

Name used used
sensitivity Specificity accuracy
Raghavendra 18 layer 98% 98.30% 98.13 results Starting Medical
et al.[72] CNN demonstr from College
ate the scratch to
robustne train a CNN
ss of the is a difficult
1 system task. They
need a
bunch of
data that
has been
labelled.
Dhaval VGG16 92% 94% 93% High For ACRIM
Vaghjiani[13] accuracy VGG16, the A,
and robustness Drishti-
simple against GS,
2
sequenti artefacts sjchoi86
al nature was not HRF and
examined. DRIONS
-DB
Mamta CoG-NET 95% 99% 95.3% indicatin Computatio RIM-
Juneja[41] g nal time is One,
3 enhance more Drishti,
d REFUG
learning E and

EAI Endorsed Transactions on

Pervasive Health and Technology
10 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

and ACRIM
improve A.
d
predictio
n.
Anita GlauNet 88.9% 89.6% 87.05% GlauNet The quality OCTA
Manassakorn[ had and public
19] 93.5% diversity of dataset
negative training
predictiv data used to
e value. build the
model may
4
impact on
how
accurate the
GlauNet
algorithm
is.

Javier Civit- ResNet50 91% 89% 90% ResNet On the test RIM-
Masot[24] architect set for the OneV3,
ure is Fundus DRISHT
best Image I
suited as Dataset for dataset
a feature Glaucoma
extractor Analysis
for and
5
glaucom Research,
a ResNet-50
identific does not
ation generalise
from well.
retinal
fundus
pictures
Javier Civit- MobileNe 89% 82% 86% The Accuracy RIM-
Masot[24] tV2 network density is OneV3,
6 is lighter higher DRISHT
and is compared to I dataset
feasible VGG16
Javier Civit- Xception 93% 85% 89% Compare RIM-
Masot[24] d to the One V3,
other DRISHT
architect I dataset
ures, the
Xception
7 architect
ure
offers
superior
trade-off
between
the

EAI Endorsed Transactions on

Pervasive Health and Technology
11 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

quantity
of
paramete
rs and
the
acquired
AUC.
Andres Inception 92% 87% 90% This more ACRIM
Diaz-Pinto[55] V3 sturdy resources A
substitut and database,
e for an processing HRF,
automati power are Drishti-
c needed GS1,
glaucom RIM
8
a
screenin
g
equipme
nt is
available
.
Anuradha SVM 92% 98% 97% online
9 Pandey[26] accessibl
e dataset
Anuradha LDA 90% 93% 97% online
10 Pandey[26] accessibl
e dataset
Anuradha Decision 92% 98% 87% online
11 Pandey[26] Tree accessibl
e dataset
Anuradha KNN 98% 99% 98% online
12 Pandey[26] accessibl
e dataset
Anuradha Neural - - 97% online
13 Pandey[26] Network accessibl
e dataset
Silvia DenseNet 0.941 0.1 97% In It requires ACRIM
Ovreiu[28] contrast significant A
to computatio dataset.
conventi nal
onal resources
convolut and time for
ional training
14 neural
networks
,
DenseNe
t has
dense
connecti
ons

EAI Endorsed Transactions on

Pervasive Health and Technology
12 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

between
its
layers,
allowing
for
effective
informati
on flow
and less
overfittin
g.
Shaswat G-Net - - 95% G-Net If the DRISHT
Singh[56] enables training I-GS
it to data is dataset,
record biased or Aravind
intricate inaccurate, eye
connecti the model hospital
ons may make
between incorrect
15 the predictions.
images
of the
fundus
and the
existence
of
glaucom
a.
Wangmin EAMNet - - 88% The To train, a ORIGA
Liao[57] precise lot of data dataset
area of and
activity computer
that power are
processe needed.
s to the
16 diagnosi
s of
glaucom
a is
provided
by
EAMNet
.
Tomaz Ribeiro GAN 30.1 69.7 69.7 With the RIM-
Viana use of ONE and
Bisneto[60] retinal Drishti-
pictures, GS
17
GAN
seeks to
carry out
OD

EAI Endorsed Transactions on

Pervasive Health and Technology
13 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

detection
, feature
extractio
n, and
pattern
classifica
tion in
order to
offer a
quick,
precise,
and
automati
c
glaucom
a
diagnosi
s.
Yunzhe DAGD 75.00 94.44 89.58 In the Drishti-
Sun[69] task of GS
domain dataset,
adaptatio PrivateD
n for ataset1
glaucom and
18 a PrivateD
diagnosi ataset2,
s, it REFUG
performs E
exceptio challenge
nally dataset
well.
N S Jeya Histogram 89.9 95.8 98.5 An ELM drawbacks DR
Shyla[59] of classifier of DWTs HAGIS
oriented is used are Shift database
gradients( to extract invariance, was
HoG), many constant develope
Discrete features time- d.
19
Wavelet for high- frequency
transform precision coverage,
(DWT), glaucom and poor
a frequency
detection resolution
.
Deepak Flexible 93.20 96.67 93.40% FAWT is ORIGA-
Parashar[61] Analytic a light,
Wavelet develope Drishti,
Transform d variant HRF and
20
of DWT, RIM-
(FAWT)
a ONE.
transfor
m that is

EAI Endorsed Transactions on

Pervasive Health and Technology
14 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

highly
helpful
for
analysin
g
medical
images.
Liu Li[34] AG-CNN 84.8 87.43 85.2 In terms The model's LAG
of performanc database
generalis e depends
ation on the
capabiliti quality and
es, the variety of
AG- the training
CNN data, and
method limited
performs training
21
admirabl data may
y, far result in
better poor
compare performanc
d to e or biased
other results.
cutting-
edge
techniqu
es.
Narmatha PH-CNN 97.10 90.63 95.04 PH-CNN Poor- Private
Venugopal[66] model quality database
gives images
more might affect
accuracy the
22
and performanc
robustne e of model
ss and lead to
inaccurate
results.
Narmatha GoogLeN 80% 90.96% 85.48 Outperfo Has low Private
Venugopal[66] et rmed accuracy database
various
models
23
in terms
of
specificit
y
Rui Fan[72] DeiT(Data 85% 81% 90% To Usually, OHTS
-efficient increase training test set
image the requires a
24
Transform generaliz lot of data.
er)[72] ability of
AI

EAI Endorsed Transactions on

Pervasive Health and Technology
15 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

models
in
healthcar
e
applicati
ons,
DeiT
may be
especiall
y
helpful.

3. Summary of Research Findings In other words, trust remains meaningless in the

absence of clinical implications. Due to these factors, AI
Qualitative comparisons for the methods across categories and ML systems must be both interpretable and explicable
and necessary explanations are discussed here for Table 2. in order to follow clinical protocols while analysing data
Moreover, the common techniques used today for and making decisions to improve patient care. These
glaucoma detection have been discussed. models can enhance our capacity to find novel biological
Increasingly, machine learning is being phenomena.
employed in medical applications like computer-aided
diagnostics. Many researchers have utilized machine
learning to categorize glaucoma and healthy eyes because 4. Feature Extraction
it can find links between input characteristics and labels.
A step towards developing a standardized, clinically
Even while machine learning has its own benefits, it was
accepted collection of OCT or retinal characteristics for
not always the most effective method for detecting
glaucoma diagnosis is identifying properties that are
glaucoma because of how complex ML models are and
similar to humans and machines for reliable glaucoma
how many iterations and hyperparameters they require.
identification. Understanding the characteristics (or
They are essentially determined by how well they perform
"concepts"") that the top-performing CNN uses to
on the dataset under study, which does not persuade
distinguish between glaucomatous and healthy pictures
physicians, medical professionals, or patients. They
will help researchers develop their own algorithms. In
frequently just affect the computer industry [73].
order to classify or segment pictures using machine
As the technology grew and much unlabelled
learning algorithms, features must be extracted that
data started generating, the researchers moved towards a
accurately describe the input data. Features are either
more sophisticated approach like Deep learning models,
created by hand, that is, manually specified, such as by an
after trying their best in ML models. [73][74] The most
algorithm developer, or they are retrieved using a
popular deep learning models such as CNN have been
predetermined filter bank, or they are eventually learned
discussed as they gave a best-in-class performance in
during training to carry out suitable task-specific
glaucoma detection. Medical diagnostics have benefited
processing of the incoming data.
from the use of DL algorithms for the screening of
Convolutional Neural Networks (CNNs) emerged as
oculopathy in retinol fundus images, the [78] diagnosis of
the preferred feature extraction method in medical image
diabetic retinopathy, and the prediction of glaucoma
analysis and it is widely used for ophthalmic image
subtypes. Despite this, these approaches are frequently
analysis [18]. In this review, two primary picture
criticized for giving no chance to comprehend the
modalities that are frequently utilized to look into
processes involved in classification determinations. A
ophthalmic-related disorders are taken into account. (1)
well-liked deep learning model, CNN is well-known for
Fundus color images that display the macula, optic disc,
its useful results in medical applications. [6] Since it has
and retinal vasculature, as well as different abnormalities
numerous optimized parameters to demonstrate a
associated with ophthalmic disorders. (2) High-resolution
sufficient success level for the target challenges, XAI
cross-sectional scans of the retinal layers in the posterior
components are still required.
segment or anterior segment structure are obtained using

EAI Endorsed Transactions on

Pervasive Health and Technology
16 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

optical coherence tomography (OCT) imaging. When we layer, which is commonly implemented as a Kohonen
talk about OCT images in [11] the author used layers of layer. These fuzzy layers in ANFIS and SNN aid to deal
CNN to classify the images of OCT where each layer is with the ambiguity and uncertainty in the input data,
followed by RLU and the maximum pooling layer. In [15] strengthening and improving the models' ability to address
the author utilized Concept Activation Vectors [CAVs] in practical issues. Based on retinal fundus pictures, ANFIS
order to determine the rate of change of class prediction and SNN can be utilized to create predictions in reference
as a function of pixel intensity at a specific pixel location. to glaucoma detection.
When it comes to using fundus images it is based on both The most compelling contributions of this research are
classification and segmentation, in [17] they cropped all highlighted:
the fundus images around the optic disc using CNN based • Use Explainable AI methods for identifying
method to find the most probable pixels in the optic disc of glaucoma using retinal fundus images [1].
region. In [11] the authors have performed automation • Evaluating the existing algorithms ANFIS
segmentation of OD and OC from fundus images using a and SNN and Further compare these models
multi-label deep network. In with deep learning models.
[12][20][40][41][42][44][48][50][71] the authors • Propose an approach using XAI and DL for
emphasize on the main feature for the detection of Glaucoma detection.
glaucoma is CDR i.e., the ratio between the cup and the
disc and the thickness of optic nerve. Figure 2 provides the basic methodology of the suggested
system. It explains the process as it takes fundus images
as input, different image pre-processing methods such as
5. Preprocessing Methods rescaling, flipping, and resizing are applied and the main
proposed model ANFIS and SNN fuzzy layer is applied
After the input picture has been obtained, pre-processing
on the images with other DL methods and finally, the
is used to remove extra noise from the image. When
images are tested to find glaucoma positive and negative
assessing medical photos, the pre-processing step is
images.
crucial. To improve picture contrast, the input image goes
through pre-processing. Pre-processing's main objective is
to smooth the input picture. Future Advancement
An in-depth study has been conducted over the past ten
The authors of [11] employed polar transformation to years whose main focus has been the utilization of
convert the fundus picture into a polar coordinate system,
which provides the benefits of spatial restriction, computational approaches for the purpose of treatment
equivalent augmentation, and balancing cup proportion and diagnosis diseases related to eyes. Numerous
while improving segmentation performance. The pre- techniques try to combine feature selection with
processing methods of images mostly involves around classification using classifiers like SVM, Random Forest
resizing the images as in [44][50], in [16] the authors used Gaussian approach, etc., CNN methods like AG-CNN,
min-max normalization in an attempt to reduce the GAN, etc. In terms of detection accuracy, these
background noise of the image, resized the images at a
approaches have attained a certain level of entitlement.
particular height, width and count and further they have
used Prewitt operator to find the edges of the images. In With the combination of fine-tuning applied to the XAI
[41][42][48][54][71] the author has pre-processed the models, a model that is both considerably more accurate
initial image using cropping and augmentation. and efficient than the current model can be built in the
future. The accuracy can be enhanced with a more
6. Proposed Model efficient and accurate dataset. Future research may
explore utilizing our technique on a bigger and more
The proposed XAI methods include Spiking Neural diverse set of images.
Network (SNN) and Adaptive Neuro-Fuzzy Inference
System (ANFIS) both employing fuzzy logic in their
models to arrive at conclusions based on input data. A
fuzzy layer is employed in ANFIS to draw conclusions
about the input data and the model's rules. In order to
transfer input data to a lower-dimensional representation
and make judgments based on the characteristics that
were learned, the SNN additionally employs a fuzzy

EAI Endorsed Transactions on

Pervasive Health and Technology
17 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

[3] Caroprese, L., Vocaturo, E., & Zumpano, E. (2022).

Argumentation approaches for explainable AI in medical
informatics. Intelligent Systems with Applications, 16,
200109.
[4] Čyras, K., Rago, A., Albini, E., Baroni, P., & Toni, F.
(2021). Argumentative XAI: a survey. arXiv preprint
arXiv:2105.11266.
[5] Siddiqui, K., & Doyle, T. E. (2022, September). Trust
Metrics for Medical Deep Learning Using Explainable-AI
Ensemble for Time Series Classification. In 2022 IEEE
Canadian Conference on Electrical and Computer
Engineering (CCECE) (pp. 370-377). IEEE.
[6] Carrieri, A. P., Haiminen, N., Maudsley-Barton, S.,
Gardiner, L. J., Murphy, B., Mayes, A. E., ... & Pyzer-
Knapp, E. O. (2021). Explainable AI reveals changes in
skin microbiome composition linked to phenotypic
differences. Scientific reports, 11(1), 1-18.
[7] Fujita, K., Shibahara, T., Chiba, D., Akiyama, M., &
Uchida, M. (2022, May). Objection!: Identifying
Misclassified Malicious Activities with XAI. In ICC 2022-
IEEE International Conference on Communications (pp.
2065-2070). IEEE.
[8] Abeyagunasekera, S. H. P., Perera, Y., Chamara, K.,
Kaushalya, U., Sumathipala, P., & Senaweera, O. (2022,
April). LISA: Enhance the explainability of medical
images unifying current XAI techniques. In 2022 IEEE 7th
International conference for Convergence in Technology
(I2CT) (pp. 1-9). IEEE.
[9] Murray, B., Anderson, D. T., & Havens, T. C. (2021, July).
Fig 3: Basic methodology Actionable XAI for the Fuzzy Integral. In 2021 IEEE
International Conference on Fuzzy Systems (FUZZ-
IEEE) (pp. 1-8). IEEE.
Conclusions [10] Cambria, E., Malandri, L., Mercorio, F., Mezzanzanica,
M., & Nobani, N. (2023). A survey on XAI and natural
This article presented detailed investigation of various language explanations. Information Processing &
glaucoma detection methods. Early diagnosis of glaucoma Management, 60(1), 103111.
[11] Fu, H., Cheng, J., Xu, Y., Wong, D. W. K., Liu, J., & Cao,
is essential as it is among the most dangerous eye X. (2018). Joint optic disc and cup segmentation based on
conditions that can asymptomatically result in multi-label deep network and polar transformation. IEEE
sightlessness, is necessary to help patients avoid total loss transactions on medical imaging, 37(7), 1597-1605.
of vision. Deep learning models have been shown to be [12] Pinos-Velez, E., Flores-Rivera, M., Ipanque-Alama, W.,
Herrera-Alvarez, D., Chacon, C., & Serpa-Andrade, L.
helpful for facilitating this early diagnosis. In this study, (2018, October). Implementation of support tools for the
various methods for diagnosing glaucoma are examined presumptive diagnosis of Glaucoma through identification
and mean accuracy, specificity, and sensitivity are and processing of medical images of the human eye.
In 2018 IEEE International Systems Engineering
considered. The approaches' benefits and drawbacks are Symposium (ISSE) (pp. 1-5). IEEE.
examined. XAI models are a step forward toward [13] Vaghjiani, D., Saha, S., Connan, Y., Frost, S., &
glaucoma detection and are expected to contribute more Kanagasingam, Y. (2020, November). Visualizing and
to medical fields. The suggested method might help understanding inherent image features in CNN-based
glaucoma detection. In 2020 Digital Image Computing:
doctors identify glaucoma early. Techniques and Applications (DICTA) (pp. 1-3). IEEE.
[14] Islam, M. T., Imran, S. A., Arefeen, A., Hasan, M., &
Shahnaz, C. (2019, November). Source and camera
References independent ophthalmic disease recognition from fundus
image using neural network. In 2019 IEEE International
Conference on Signal Processing, Information,
[1] Zolanvari, M., Yang, Z., Khan, K., Jain, R., & Meskin, N. Communication & Systems (SPICSCON) (pp. 59-63).
(2021). Trust xai: Model-agnostic explanations for ai with IEEE.
a case study on iiot security. IEEE Internet of Things [15] Thakoor, K. A., Koorathota, S. C., Hood, D. C., & Sajda,
Journal. P. (2020). Robust and interpretable convolutional neural
[2] Tjoa, E., & Guan, C. (2020). A survey on explainable networks to detect glaucoma in optical coherence
artificial intelligence (xai): Toward medical xai. IEEE tomography images. IEEE Transactions on Biomedical
transactions on neural networks and learning Engineering, 68(8), 2456-2466.
systems, 32(11), 4793-4813. [16] Toki, S. A., Rahman, S., Fahim, S. M. B., Al Mostakim,
A., & Rhaman, M. K. (2022, May). RetinalNet-500: A

EAI Endorsed Transactions on

Pervasive Health and Technology
18 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

newly developed CNN Model for Eye Disease Detection. coherence tomography and color fundus images. Journal of
In 2022 2nd International Mobile, Intelligent, and healthcare engineering, 2019.
Ubiquitous Computing Conference (MIUCC) (pp. 459- [30] Sabina, R., & Zarina, S. (2022, April). Convolutional
463). IEEE. Neural Network Analysis of Fundus for Glaucoma
[17] Diaz-Pinto, A., Colomer, A., Naranjo, V., Morales, S., Xu, Diagnosis. In 2022 International Conference on Smart
Y., & Frangi, A. F. (2019). Retinal image synthesis and Information Systems and Technologies (SIST) (pp. 1-6).
semi-supervised learning for glaucoma assessment. IEEE IEEE.
transactions on medical imaging, 38(9), 2211-2218. [31] Thakur, A., Goldbaum, M., & Yousefi, S. (2020). Convex
[18] Sarhan, M. H., Nasseri, M. A., Zapp, D., Maier, M., representations using deep archetypal analysis for
Lohmann, C. P., Navab, N., & Eslami, A. (2020). Machine predicting glaucoma. IEEE Journal of Translational
learning techniques for ophthalmic data processing: a Engineering in Health and Medicine, 8, 1-7.
review. IEEE Journal of Biomedical and Health [32] Gunasinghe, H., McKelvie, J., Koay, A., & Mayo, M.
Informatics, 24(12), 3338-3350. (2021, April). Comparison Of Pretrained Feature
[19] Manassakorn, A., Auethavekiat, S., Sa-Ing, V., Extractors For Glaucoma Detection. In 2021 IEEE 18th
Chansangpetch, S., Ratanawongphaibul, K., Uramphorn, International Symposium on Biomedical Imaging
N., & Tantisevi, V. (2022). GlauNet: Glaucoma Diagnosis (ISBI) (pp. 390-394). IEEE.
for OCTA Imaging Using a New CNN Architecture. IEEE [33] Barros, D., Moura, J. C., Freire, C. R., Taleb, A. C.,
Access, 10, 95613-95622. Valentim, R. A., & Morais, P. S. (2020). Machine learning
[20] Carrillo, J., Bautista, L., Villamizar, J., Rueda, J., & applied to retinal image processing for glaucoma detection:
Sanchez, M. (2019, April). Glaucoma detection using review and perspective. Biomedical engineering
fundus images of the eye. In 2019 XXII Symposium on online, 19(1), 1-21.
Image, Signal Processing and Artificial Vision [34] Li, L., Xu, M., Liu, H., Li, Y., Wang, X., Jiang, L., ... &
(STSIVA) (pp. 1-4). IEEE. Wang, N. (2019). A large-scale database and a CNN model
[21] Lavric, A., Petrariu, A. I., Havriliuc, S., & Coca, E. (2021, for attention-based glaucoma detection. IEEE transactions
November). Glaucoma Detection by Artificial Intelligence: on medical imaging, 39(2), 413-424.
GlauNet A Deep Learning Framework. In 2021 [35] Joshi, S., Partibane, B., Hatamleh, W. A., Tarazi, H.,
International Conference on e-Health and Bioengineering Yadav, C. S., & Krah, D. (2022). Glaucoma Detection
(EHB) (pp. 1-4). IEEE. Using Image Processing and Supervised Learning for
[22] Ramanathan, G., Chakrabarti, D., Patil, A., Rishipathak, S., Classification. Journal of Healthcare Engineering, 2022.
& Kharche, S. (2021, October). Eye Disease Detection [36] Shinde, R. (2021). Glaucoma detection in retinal fundus
Using Machine Learning. In 2021 2nd Global Conference images using U-Net and supervised machine learning
for Advancement in Technology (GCAT) (pp. 1-5). IEEE. algorithms. Intelligence-Based Medicine, 5, 100038.
[23] Jibhakate, P., Gole, S., Yeskar, P., Rangwani, N., Vyas, A., [37] Ajitha, S., Akkara, J. D., & Judy, M. V. (2021).
& Dhote, K. (2022, July). Early Glaucoma Detection Identification of glaucoma from fundus images using deep
Using Machine Learning Algorithms of VGG-16 and learning techniques. Indian Journal of
Resnet-50. In 2022 IEEE Region 10 Symposium Ophthalmology, 69(10), 2702.
(TENSYMP) (pp. 1-5). IEEE. [38] Ştefan, A. M., Paraschiv, E. A., Ovreiu, S., & Ovreiu, E.
[24] Civit-Masot, J., Domínguez-Morales, M. J., Vicente-Díaz, (2020, October). A review of glaucoma detection from
S., & Civit, A. (2020). Dual machine-learning system to digital fundus images using machine learning techniques.
aid glaucoma diagnosis using disc and cup feature In 2020 International conference on e-health and
extraction. IEEE Access, 8, 127519-127529. bioengineering (EHB) (pp. 1-4). IEEE.
[25] Chethan, M., Dasari, C., Uttarkar, G. V., & Sachin, D. N. [39] Abdullah, F., Imtiaz, R., Madni, H. A., Khan, H. A., Khan,
(2019, January). Diagnosis of Glaucoma using Machine T. M., Khan, M. A., & Naqvi, S. S. (2021). A review on
Learning-A Survey. In 2019 Third International glaucoma disease detection using computerized
Conference on Inventive Systems and Control (ICISC) (pp. techniques. IEEE Access, 9, 37311-37333.
210-214). IEEE. [40] Coan, L., Williams, B., Venkatesh, M. K. A., Upadhyaya,
[26] Pandey, A., Patre, P., & Minj, J. (2020, October). S., Al Kafri, A., Czanner, S., ... & Czanner, G. (2022).
Detection of Glaucoma Disease using Image Processing, Automatic detection of glaucoma via fundus imaging and
Soft Computing and Deep Learning Approaches. In 2020 artificial intelligence: A review. Survey of ophthalmology.
Fourth International Conference on I-SMAC (IoT in [41] Juneja, M., Thakur, S., Uniyal, A., Wani, A., Thakur, N.,
Social, Mobile, Analytics and Cloud)(I-SMAC) (pp. 1-7). & Jindal, P. (2022). Deep learning-based classification
IEEE. network for glaucoma in retinal images. Computers and
[27] Song, D., Fu, B., Li, F., Xiong, J., He, J., Zhang, X., & Electrical Engineering, 101, 108009.
Qiao, Y. (2021). Deep relation transformer for diagnosing [42] Liu, B., Pan, D., Shuai, Z., & Song, H. (2022). ECSD-Net:
glaucoma with optical coherence tomography and visual A joint optic disc and cup segmentation and glaucoma
field function. IEEE Transactions on Medical classification network based on unsupervised domain
Imaging, 40(9), 2392-2402. adaptation. Computer Methods and Programs in
[28] Ovreiu, S., Paraschiv, E. A., & Ovreiu, E. (2021, July). Biomedicine, 213, 106530.
Deep Learning & Digital Fundus Images: Glaucoma [43] de Zarzà, I., de Curtò, J., & Calafate, C. T. (2022).
Detection using DenseNet. In 2021 13th International Detection of glaucoma using three-stage training with
Conference on Electronics, Computers and Artificial EfficientNet. Intelligent Systems with Applications, 16,
Intelligence (ECAI) (pp. 1-4). IEEE. 200140.
[29] An, G., Omodaka, K., Hashimoto, K., Tsuda, S., Shiga, Y., [44] Deperlioglu, O., Kose, U., Gupta, D., Khanna, A.,
Takada, N., ... & Nakazawa, T. (2019). Glaucoma Giampaolo, F., & Fortino, G. (2022). Explainable
diagnosis with machine learning based on optical framework for Glaucoma diagnosis by image processing
and convolutional neural network synergy: analysis with

EAI Endorsed Transactions on

Pervasive Health and Technology
19 | Volume 10 | 2024 |
N. Afreen and R. Aluvalu

doctor evaluation. Future Generation Computer Symposium on Multidisciplinary Studies and Innovative
Systems, 129, 152-169. Technologies (ISMSIT) (pp. 567-570). IEEE.
[45] Song, D., Li, F., Li, C., Xiong, J., He, J., Zhang, X., & [59] Shyla, N. J., & Emmanuel, W. S. (2021, February).
Qiao, Y. (2022). Asynchronous feature regularization and Automated Classification of Glaucoma Using DWT and
cross-modal distillation for OCT based glaucoma HOG Features with Extreme Learning Machine. In 2021
diagnosis. Computers in Biology and Medicine, 151, Third International Conference on Intelligent
106283. Communication Technologies and Virtual Mobile
[46] Kim, M., Janssens, O., Park, H. M., Zuallaert, J., Van Networks (ICICV) (pp. 725-730). IEEE.
Hoecke, S., & De Neve, W. (2018). Web applicable [60] Bisneto, T. R. V., de Carvalho Filho, A. O., & Magalhães,
computer-aided diagnosis of glaucoma using deep D. M. V. (2020). Generative adversarial network and
learning. arXiv preprint arXiv:1812.02405. texture features applied to automatic glaucoma
[47] Prabhakar, B., Singh, R. K., & Yadav, K. S. (2021). detection. Applied Soft Computing, 90, 106165.
Artificial intelligence (AI) impacting diagnosis of [61] Parashar, D., & Agrawal, D. K. (2020). Automated
glaucoma and understanding the regulatory aspects of AI- classification of glaucoma stages using flexible analytic
based software as medical device. Computerized Medical wavelet transform from retinal fundus images. IEEE
Imaging and Graphics, 87, 101818. Sensors Journal, 20(21), 12885-12894.
[48] Jun, T. J., Eom, Y., Kim, D., Kim, C., Park, J. H., Nguyen, [62] George, Y., Antony, B. J., Ishikawa, H., Wollstein, G.,
H. M., ... & Kim, D. (2021). TRk-CNN: transferable Schuman, J. S., & Garnavi, R. (2020). Attention-guided
ranking-CNN for image classification of glaucoma, 3D-CNN framework for glaucoma detection and
glaucoma suspect, and normal eyes. Expert Systems with structural-functional association using volumetric
Applications, 182, 115211. images. IEEE Journal of Biomedical and Health
[49] Wu, Y., Szymanska, M., Hu, Y., Fazal, M. I., Jiang, N., Informatics, 24(12), 3421-3430.
Yetisen, A. K., & Cordeiro, M. F. (2022). Measures of [63] Shamia, D., Prince, S., & Bini, D. (2022, April). An Online
disease activity in glaucoma. Biosensors and Platform for Early Eye Disease Detection using Deep
Bioelectronics, 196, 113700. Convolutional Neural Networks. In 2022 6th International
[50] Xue, Y., Zhu, J., Huang, X., Xu, X., Li, X., Zheng, Y., ... Conference on Devices, Circuits and Systems
& Si, K. (2022). A multi-feature deep learning system to (ICDCS) (pp. 388-392). IEEE.
enhance glaucoma severity diagnosis with high accuracy [64] An, G., Omodaka, K., Tsuda, S., Shiga, Y., Takada, N.,
and fast speed. Journal of Biomedical Informatics, 136, Kikawa, T., ... & Akiba, M. (2018). Comparison of
104233. machine-learning classification models for glaucoma
[51] Jain, S., Indora, S., & Atal, D. K. (2022). Rider Manta ray management. Journal of healthcare engineering, 2018.
foraging optimization-based generative adversarial [65] Eswari, M. S., & Balamurali, S. (2021, March). An
network and CNN feature for detecting intelligent machine learning support system for glaucoma
glaucoma. Biomedical Signal Processing and Control, 73, prediction among diabetic patients. In 2021 International
103425. Conference on Advance Computing and Innovative
[52] Bajwa, M. N., Malik, M. I., Siddiqui, S. A., Dengel, A., Technologies in Engineering (ICACITE) (pp. 447-449).
Shafait, F., Neumeier, W., & Ahmed, S. (2019). Two-stage IEEE.
framework for optic disc localization and glaucoma [66] Venugopal, N., & Mari, K. (2019, November). An
classification in retinal fundus images using deep Automated glaucoma image classification model using
learning. BMC medical informatics and decision perceptual hash-based convolutional neural network.
making, 19(1), 1-16. In 2019 International Conference on Smart Systems and
[53] M. Juneja, S. Thakur, A. Wani, A. Uniyal, N. Thakur, P. Inventive Technology (ICSSIT) (pp. 185-190). IEEE.
Jindal, DC-Gnet for detection of glaucoma in retinal [67] Song, W. T., Lai, C., & Su, Y. Z. (2021). A Statistical
fundus imaging, Mach. Vis. Appl. 31 (2020) 1–14. Robust Glaucoma Detection Framework Combining
[54] Chen, H. S. L., Chen, G. A., Syu, J. Y., Chuang, L. H., Su, Retinex, CNN, and DOE Using Fundus Images. IEEE
W. W., Wu, W. C., ... & Kang, E. Y. C. (2022). Early Access, 9, 103772-103783.
Glaucoma Detection by Using Style Transfer to Predict [68] Serener, A., & Serte, S. (2019, October). Transfer learning
Retinal Nerve Fiber Layer Thickness Distribution on the for early and advanced glaucoma detection with
Fundus Photograph. Ophthalmology Science, 2(3), convolutional neural networks. In 2019 Medical
100180. technologies congress (TIPTEKNO) (pp. 1-4). IEEE.
[55] Diaz-Pinto, A., Morales, S., Naranjo, V., Köhler, T., [69] Sun, Y., Yang, G., Ding, D., Cheng, G., Xu, J., & Li, X.
Mossi, J. M., & Navea, A. (2019). CNNs for automatic (2020, July). A GAN-based domain adaptation method for
glaucoma assessment using fundus images: an extensive glaucoma diagnosis. In 2020 International Joint
validation. Biomedical engineering online, 18(1), 1-19. Conference on Neural Networks (IJCNN) (pp. 1-8). IEEE.
[56] M. Juneja, S. Singh, N. Agarwal et al., “Automated [70] Sreng, S., Maneerat, N., Hamamoto, K., & Win, K. Y.
detection of Glaucoma using deep learning convolution (2020). Deep learning for optic disc segmentation and
network (G-net)”. Multimed Tools Appl 79, 15531–15553, glaucoma diagnosis on retinal images. Applied
2020. Sciences, 10(14), 4916.
[57] Liao, W., Zou, B., Zhao, R., Chen, Y., He, Z., & Zhou, M. [71] Fan, R., Alipour, K., Bowd, C., Christopher, M., Brye, N.,
(2019). Clinical interpretable deep learning model for Proudfoot, J. A., ... & Zangwill, L. M. (2023). Detecting
glaucoma diagnosis. IEEE journal of biomedical and Glaucoma from Fundus Photographs Using Deep Learning
health informatics, 24(5), 1405-1412. without Convolutions: Transformer for Improved
[58] Wu, F., Chiariglione, M., & Gao, X. R. (2022, October). Generalization. Ophthalmology Science, 3(1), 100233.
Automated Optic Disc and Cup Segmentation for [72] Raghavendra, U., Fujita, H., Bhandary, S. V., Gudigar, A.,
Glaucoma Detection from Fundus Images Using the Tan, J. H., & Acharya, U. R. (2018). Deep convolution
Detectron2's Mask R-CNN. In 2022 International

EAI Endorsed Transactions on

Pervasive Health and Technology
20 | Volume 10 | 2024 |
 Glaucoma Detection Using Explainable AI and Deep Learning

neural network for accurate diagnosis of glaucoma using

digital fundus images. Information Sciences, 441, 41-49.
[73] Kamal, M. S., Dey, N., Chowdhury, L., Hasan, S. I., &
Santosh, K. C. (2022). Explainable AI for glaucoma
prediction analysis to understand risk factors in treatment
planning. IEEE Transactions on Instrumentation and
Measurement, 71, 1-9.
[74] Priyanka, V., and V. Uma Maheswari. "Automated
Glaucoma Detection Using Cup to Disk Ratio and Grey
Level Co-occurrence Matrix." In Machine Learning and
Information Processing: Proceedings of ICMLIP 2020, pp.
425-434. Springer Singapore, 2021.

EAI Endorsed Transactions on

Pervasive Health and Technology
21 | Volume 10 | 2024 |