A Generalized and Explainable Deep Learning Framework for Ocular Disease Detection using Retinal Images

A Generalized and Explainable Deep Learning Framework for Ocular Disease Detection using Retinal Images

A GENERALIZED AND EXPLAINABLE DEEP LEARNING FRAMEWORK FOR OCULAR DISEASE DETECTION USING RETINAL IMAGES

A thesis Submitted for the partial fulfillment of the requirement for the degree of

Master of Technology

in

COMPUTER SCIENCE & ENGINEERING

By

NITASHA DASH

(Registration No-240743303/Computer Science and Engineering)

Supervised By

Dr. Umakanta Dash

(GUIDE)

GITA AUTONOMOUS COLLEGE, BHUBANESWAR

04-2026

Department of Computer Science & Engineering

GITA Autonomous College, Bhubaneswar

Ref no: Date: .

Certificate

This is to certify that the thesis titled A Generalized and Explainable Framework for Ocular Disease Detection Using Retinal Images is an authentic work carried out by Ms. Nitasha Dash, Regd no- 2407433003/Computer Science & Engineering for the award of the Masters degree at GITA under my guidance. The matter embodied in this thesis work has not been submitted earlier for the award of any degree to the best of my knowledge and belief.

Prof. (Dr.) Tarini Prasad Panigrahy Prof (Dr.) Umakanta Dash Professor & Head of the Dept. (Guide)

External Examiner

[From Expert Panel, BPUT, Odisha]

Department of Computer Science & Engineering

GITA Autonomous College, Bhubaneswar

ACKNOWLEDGEMENT

I express and gratitude to Prof. (Dr) Umakanta Dash, project supervisor for his guidance and constant support.

I also take this opportunity to thank Prof. (Dr.) Tarini Prasad Panigrahy, head of Department, Computer Science & Engineering, for his constant support and timely advice.

Lastly, words run to express my gratitude to all the faculties of the CSE Dept. And friends for their support and co-operation, constructive criticism, and valuable suggestion during preparation of this project report.

Thanking All.

(Full signature of the student)

NITASHA DASH

Registration No-2407433003

List of Acronyms

AI Artificial Intelligence DL Deep Learning

ML Machine Learning

CNN Convolutional Neural Network ViT Vision Transformer

XAI Explainable Artificial Intelligence LBP Local Binary Pattern

ODIR Ocular Disease Intelligent Recognition AMD Age-related Macular Degeneration

DR Diabetic Retinopathy GL Glaucoma

CA Cataract

OCT Optical Coherence Tomography

OCTA Optical Coherence Tomography Angiography CFP Colour Fundus Photography

ARIA Retinal Image Analysis AUC Area Under Curve

IoU Intersection over Union ReLU Rectified Linear Unit

DCNN Deep Convolutional Neural Network

List of key words

Retinal Fundus Images Deep Learning

Convolutional Neural Network (CNN) Vision Transformer (ViT)

VGG-19 / VGG-16

ResNet50 EfficientNet-B7 Xception

Local Binary Pattern (LBP)

Explainable Artificial Intelligence (XAI) Medical Image Analysis

Binary Classification ODIR Dataset

Computer-Aided Diagnosis

List of tables

TABLE NUMBER PAGE NO.

Table 5.1: Model Training Configuration 24

Table 6.1: Performance Metrics of Models 30

Table 6.2: Comparative Analysis of Models 33

List of Figures

FIGURE NUMBER PAGE NO.

Figure 4.1: System Architecture Diagram 16

Figure 5.1: Model Training Workflow 23

Figure 6.1: Output Images (Normal, Glaucoma, DR, Cataract) 31

Figure 6.2: ROC Curve for Multiclass Classification 32

ABSTRACT

In order to prevent vision loss and improve patient outcomes, early detection of ocular diseases using retinal fundus images is essential. However, traditional manual diagnosis by clinicians is frequently costly, time-consuming, and prone to human error, especially when handling large volumes of medical data. As a result, the creation of automated computer-aided diagnostic systems is now crucial to contemporary healthcare. The Ocular illness Intelligent Recognition (ODIR) dataset, which includes retinal fundus images representing various ocular illness categories, is used in this thesis to propose a generalized and explainable deep learning architecture for ocular disease identification utilizing retinal images. The multiclass classification problem is reduced to a binary classification task to improve model stability and performance, and a data balancing method is used to address the problem of class imbalance. Imbalance, a data balancing technique is used, and to improve model performance and stability, the multiclass classification issue is reduced to a binary classification task. In order to extract significant features and precisely identify ocular disorders, the suggested framework incorporates cutting-edge deep learning architectures as VGG-19, ResNet50, and Efficient NetB7, Xception. Additionally, to enhance the extraction of textural features from retinal pictures, Local Binary Pattern (LBP) approaches are applied. In order to help clinicians better comprehend the diagnostic process and boost confidence in automated systems, the system also incorporates explainable artificial intelligence techniques that provide visual interpretations of model predictions. According to experimental results, the suggested framework outperforms conventional approaches in terms of accuracy, recall, specificity, and F1-score. This suggests that the developed model offers an effective, dependable, and comprehensible solution for early detection of ocular diseases and can be a useful clinical decision support tool in practical healthcare applications.

Keyword: Ocular Disease Detection, Retinal Fundus Images, Deep Learning, Convolutional Neural Network (CNN), Vision Transformer (ViT), VGG-19, ResNet50, Local Binary Pattern (LBP), Explainable Artificial Intelligence (XAI), Medical Image Analysis, Binary Classification, ODIR Dataset, Computer-Aided Diagnosis, Efficient NetB7, Xception

TABLE OF CONTENTS
Chapter	A enda Description	Pa e No.
CHAPTER-1		1
1	1.1 INTRODUCTION	2-3
	1.2 BACKGROUND AND MOTIVATION	4
CHAPTER-2		5
2	2.1 LITERATURE REVIEW	6
	2.1.1 General	6-7
	2.1.2 LITERATURE ANALYSIS OF AI-BASED OCULAR DISEASE DETECTION	8-12
CHAPTER-3		13
3	3.1 PREVIOUS WORK	14-19
	3.1.1 Gap Analysis	20
CHAPTER-4		21
4	4.1 RESTNET: RESIDUAL NETWORK	22-24
	4.1.1 RESTNET50	24-26
	4.2 VGG-16: VISUAL GEONIETRY GROUP	27
	4.3 EFFICIENTNET-B7	28-29
	4.4 XCEPTION	29-31
	4.5 PROPOSED MODEL	31
	4.5.1 Retinal image as input	32
	4.5.2 Processing Image	32
	4.5.3 Data Enhancement	32
	4.5.4 Extracting Feahue	32
	4.5.5 Layers of full connectivity	33
	4.5.6 Soft:Max Classifier	33
	4.5.7 Disease Prediction	33
	4.5.8 The Explainable AI Module	33
	4.5.9 Evaluation of Perfonnance	33
CHAPTER-5		34
5	5.1 NIETHEDOLOGY	35
	5.1.1 Worktlow	35
	5.1.2 Ove1view	36
	5.1.3 General System Design	36
	5.1.4 Data Collection	36-37
	5.1.5 Image processing	38-39
	5.1.6 Training Model	39-40
	5.1.7 Model Evaluation	41
CHAPTER-6		42
6	6.1 RESULT	42-43
	6.1.1 NIETRICS EVALUATION	44-45
	6.1.2 Comparison with other models	45-46
	6.1.3 Confusion Matrix	46-47
	6.1.4 OUTPUT	47-49
CHAPTER-?		50
7	7.1 co CLUSIO AND FUTURE WORK	51
	7.1.1 Conclusion	51
	7.1.2 Furore Work	51
CHAPTER-8		52
8	8.1 REFERNCES	53-55

1. INTRODUCTION

   Since it enables people to view and engage with the environment, vision is one of the most crucial senses. The human eye is necessary for recognizing things, understanding spatial relationships, and performing everyday activities such as reading, driving, talking, and making decisions. Any visual impairment can have a significant effect on a person's productivity, quality of life, and independence. Visual impairment and blindness are major public health concerns worldwide, impacting millions of people of all ages. According to the World Health Organization (WHO), at least a billion of the nearly 2.2 billion people worldwide who are blind or visually impaired may have been prevented or are still untreated due to inadequate access to early diagnosis and healthcare services.

   The field of medical imaging and diagnostics has undergone a major transformation due to the fast development of artificial intelligence, notably deep learning. The field of ophthalmology has significantly gained from these advancements in technology, thanks to the availability of high-resolution retinal imaging methods. Retinal images are essential for understanding many eye conditions, including age-related macular degeneration, diabetic retinopathy, and glaucoma. Since early identification of these diseases is critical for preventing vision loss, conventional diagnostic techniques are frequently time-consuming, particularly in resource- constrained environments, and necessitate the expertise of skilled ophthalmologists.

   Notably, convolutional neural networks (CNNs) and other deep learning models have shown outstanding results in image categorization and illness diagnosis. These models are capable of autonomously learning complicated features from retinal images in the context of detecting ocular disorders, which allows for precise and quick diagnosis. Most current methods, however, are created for particular diseases and cannot be applied generally to a variety of eye disorders. Due to this restriction, their use is limited in actual clinical settings where patients may have a variety of or overlapping conditions.

   The absence of interpretability or explainability is yet another significant barrier to the implementation of deep learning techniques in healthcare. Several excellent models function as "black boxes," making predictions without providing specific logic. This lack of

   transparency might undermine trust among doctors and impede implementation in real life, which could harm medical decision-making. Because of this, there is an increasing demand for explainable artificial intelligence (XAI) methodologies that can emphasize key aspects in retinal images and provide insightful information about model choices.

   The advancement of automated diagnostic systems depends on a comprehensive framework that can identify several eye illnesses while preserving great accuracy and understandability. A framework like this should be able to account for differences in image quality, patient demographics, and disease presentation. Furthermore, incorporating explainability tools into the model might help clinicians comprehend and validate predictions, which would increase clinical consistency and acceptability.

   Using retinal images, this thesis offers a clear and comprehensive deep learning architecture for identifying eye diseases. By integrating reliable feature extraction with understandable decision-making procedures, the framework seeks to facilitate the identification of several eye disorders within a single system. The suggested method attempts to close the gap between model performance and clinical usability by utilizing cutting-edge deep learning architectures and explainability methodologies.

   Ultimately, the creation of such a framework has the potential to facilitate earlier diagnosis, lessen the strain on healthcare providers, and enhance patient outcomes. It can be very helpful in underserved areas where access to expert eye care is restricted. The integration of generalized learning and explainable outputs represents a major advancement in the development of reliable and scalable AI-driven healthcare solutions for the identification of ocular diseases.

2. BACKGROUND AND MOTIVATION

Eye diseases are one of the leading causes of blindness and visual impairment worldwide, affecting the quality of life of millions of people. Because disorders like age-related macular degeneration, diabetic retinopathy, cataracts, and glaucoma can manifest silently without apparent symptoms, early diagnosis is essential for effective therapy and prevention of permanent vision loss. Manual examination of retinal photographs by ophthalmologists, on the other hand, is often labor- intensive, time-consuming, and dependent on clinical expertise. Furthermore, the growing number of patients and the scarcity of skilled experts make it challenging to make quick and accurate diagnoses, particularly in rural and resource-constrained healthcare environments. The rapid advancement of artificial intelligence and deep learning has automated medical image analysis a feasible approach to address these challenges. Particularly, deep learning models, such as transformer-based architectures and convolutional neural networks (CNNs), have demonstrated outstanding performance in picture classification, segmentation, and disease identification. This study is motivated by the need for a trustworthy, widely applicable, and understandable automated system for identifying eye diseases using retinal fundus images. Since healthcare practitioners prioritize accuracy over transparency in their decision-making processes, they may become less trusting of many modern paradigms. Furthermore, variations in disease characteristics, lighting, and picture quality may affect model performance, emphasizing the need for a framework capable of generalizing well across a range of clinical settings. Addressing class imbalance and improving feature representation are also critical for ensuring precise and consistent diagnostic findings. This thesis aims to create and implement a generalized and explainable deep learning framework that integrates cutting-edge architectures such Convolutional Neural Networks and Vision Transformers in order to precisely identify ocular dis-eases. explainable artificial intelligence techniques and texture-based feature extraction methods like Local Binary Pattern (LBP) are integrated to increase the model's performance, interpretability, and clinical dependability. By providing an efficient and automated solution for the early detection of ocular disorders, the proposed system has the potential to help medical professionals in their decision-making, reduce the burden of diagnosis, and ultimately improve patient care and prevent avoidable vision loss.

1. LITERATURE REVIEW

   1. General

      The increasing incidence of eye disorders like cataracts, glaucoma, age-related macular degeneration, and diabetic retinopathy has become a major public health issue around the world, making a significant contribution to visual impairment and blindness. To avoid irreversible vision loss and improve patient outcomes, it is critical to identify and properly diagnose these illnesses early. Fundus photography and Optical Coherence Tomography (OCT) are examples of retinal imaging methods that are essential for identifying structural and vascular problems in the eye. However, traditional diagnostic techniques heavily rely on manual examination by ophthalmologists, which can be time-consuming, subjective, and prone to human mistakes, particularly in places where specialized medical treatment is scarce. Because of this, there is a growing need for dependable, automated, and efficient diagnostic systems that can help medical professionals in detecting and categorizing eye illnesses.

      Because of its capacity to automatically learn intricate patterns from massive datasets, deep learninga branch of artificial intelligence (AI)has become a potent tool for medical image analysis in recent years. In image categorization, segmentation, and disease prediction tasks, convolutional neural networks (CNNs) and their sophisticated architectures, such as Rest Net, VGG Net, and Xception, have shown outstanding performance. These models are able to analyses retinal images with great precision and identify minute features such as microaneurysms, hemorrhages, exudates, and optic nerve damage, allowing for the early diagnosis of eye illnesses. Moreover, by integrating transfer learning approaches, pre-trained models can be customized for particular medical imaging uses, which increases computational efficiency and lessens the need for massive labelled datasets. As a result, systems based on deep learning have become a crucial part of contemporary computer-aided diagnosis in ophthalmology.

      Many contemporary systems are made to identify a single disease and are unable to generalize across many eye illnesses, even if deep learning models have been very successful in retinal disease diagnosis. Patients may simultaneously have many eye conditions in real-world clinical contexts, necessitating diagnostic systems that can manage multi-class classification issues. This restriction

      is overcome by a generalized deep learning framework that uses a single model architecture to identify and categorize a variety of eye disorders.

      By lowering system complexity, increasing scalability, and enhancing clinical usability, such frameworks accomplish this by integrating comprehensive diagnostic support into a single platform. Moreover, generalized models can be trained using a variety of datasets, making them more robust and able to adapt to various populations and imaging settings.

      The lack of transparency in the decision-making processes of these models is another major obstacle to the use of deep learning in healthcare. Due to the complexity of their internal processes, deep learning algorithms are frequently regarded as "black box" systems, which makes it harder for clinicians to have complete confidence in automated predictions. As a result of this problem, Explainable Artificial Intelligence (XAI) has been created, which seeks to offer straightforward and understandable explanations for model choices.

      Methods such as Gradient-weighted Class Activation Mapping (Grad-CAM), Local Interpretable Model-agnostic Explanations (LIME), and Shapley Additive Explanations (SHAP) create visual heatmaps and feature importance ratings that emphasize the particular areas of retinal images that are responsible for disease predictions. By improving the overall dependability of clinical decision- making, these explanations aid practitioners in validating model results and gaining faith in artificial intelligence systems.

      By facilitating quicker, more precise, and more clear disease identification, the integration of general and explainable deep learning methodologies into retinal picture analysis has the potential to transform ophthalmic diagnostics. In places with limited access to ophthalmologists, such as rural and underserved communities, these systems can help healthcare professionals screen huge numbers of patients. Additionally, automated retinal illness detection systems can help with early intervention approaches by lowering the diagnostic burden and increasing the consistency of clinical evaluations.

      The creation of a universal and interpretable deep learning architecture for identifying ocular diseases represents a potential way to improve the efficiency, dependability, and accessibility of contemporary healthcare systems as artificial intelligence advances.

   2. LITERATURE ANALYSIS OF AI-BASED OCULAR DISEASE DETECTION

Deep learning techniques for Ocular Disease Detection

By facilitating automated disease detection and classification with great accuracy and efficiency, deep learning methods have revolutionized the field of medical imaging. Convolutional Neural Networks (CNNs) are frequently employed in ophthalmology for retinal picture analysis because of their capacity to locate spatial features and complicated patterns in medical pictures. Researchers have created a number of CNN-based models for identifying ocular disorders, including age-related macular degeneration, glaucoma, and diabetic retinopathy. These models often include several steps, such as image preprocessing, feature extraction, classification, and performance evaluation.

To improve image quality and ensure consistent input data for deep learning models, image preprocessing methods like normalization, contrast enhancement, and noise reduction are used. This results in improved system performance and reliability overall. The capacity of deep learning models to automatically extract significant features from retinal images without manual intervention is one of their most important benefits. Traditional machine learning methods depended too much on manually created features, which restricted their scalability, adaptability, and general usefulness.

On the other hand, deep learning models learn intricate feature representations directly from enormous datasets, allowing them to spot subtle anomalies that may be difficult for the human eye to see. For instance, deep learning algorithms can precisely identify microaneurysms, hemorrhages, and exudates in retinal pictures, all of which are thought to be early signs of diabetic retinopathy. In a similar way, CNN-based models can detect structural alterations in the retinal nerve fibre layer and optic nerve head, which are frequently linked to glaucoma and other conditions affecting the optic nerve.

Recent studies have also concentrated on the use of sophisticated deep learning architectures, such as Residual Networks (Rest Net), Visual Geometry Group Networks (VGGNet), and Inception networks, for identifying and categorizing retinal illnesses. These architectures include cutting- edge design elements like skip connections, deep feature hierarchies, and multi-scale feature

extraction mechanisms, which greatly enhance model performance an lower the chance of overfitting.

Deep learning models based on these architectures have been shown in studies to be capable of achieving high diagnostic accuracy, sensitivity, and specificity in identifying a wide range of ocular diseases. Additionally, researchers have been able to use pre-trained models that were trained on massive datasets like ImageNet and fine-tune them for particular retinal image analysis applications thanks to the use of transfer learning methods. This strategy greatly shortens training duration, increases model generalizability, and improves the real-world clinical application of diagnostic systems based on deep learning.

Generalized deep learning framework for multi- disease classification

The creation of generalized deep learning frameworks that can identify a variety of eye ailments using a single model is a major field of study in medical image analysis. Since conventional deep learning models are frequently created to identify a single illness, their usefulness in clinical settings where patients may have a variety of concurrent ailments is restricted.

Generalized frameworks use multi-class classification approaches to address this constraint and enable the detection of numerous illnesses in a single system. Automated diagnostic systems become more appropriate for use in real-world healthcare settings due to this method, which increases operational efficiency, lowers computational complexity, and improves scalability.

The majority of generalized deep learning frameworks make use of shared feature extraction layers, followed by several classification layers that distinguish between various disease categories. These frameworks are made to handle a wide variety of datasets, including images of various retinal illnesses, enabling the models to learn generalized representations of retinal abnormalities.

These models can accurately differentiate between normal and abnormal retinal conditions by learning shared characteristics across several disease categories. Additionally, researchers have looked at the use of multi-task learning approaches, which allow a single model to carry out many connected tasks at once, such as lesion segmentation and disease categorization.

By utilizing shared information across tasks, boosting prediction accuracy, and lowering the chance of overfitting, multi-task learning enhances model performance. The capacity of generalized frameworks to adjust to various datasets and imaging settings is another significant feature.

Methods for transfer learning and domain adaptation are frequently employed to improve the generalization performance of deep learning models. These techniques enable models to learn from previously trained datasets and use that knowledge to new datasets with little retraining.

Consequently, compared to disease-specific models, generalized deep learning frameworks show enhanced robustness, dependability, and adaptability. Their capacity to manage a variety of clinical situations makes them especially well-suited for real-world clinical implementation in contemporary healthcare systems, as well as for extensive screening programs.

Explainable Artificial Intelligence (XAI) in retinal image analysis

Explainable Artificial Intelligence (XAI) has become an indispensable element of contemporary deep learning systems in healthcare, particularly in those that include medical diagnosis and clinical decision-making. Despite the fact that deep learning models have shown remarkable performance in disease detection and categorization, it is sometimes challenging to comprehend how predictions are made due to their complicated designs.

Healthcare professionals and medical experts have expressed doubts about the dependability, safety, and reliability of AI-based diagnostic systems due to this opacity. As a result, researchers have concentrated on creating explainability methodologies that offer valuable insights into the decision-making processes of deep learning models. These methods are intended to improve the transparency and interpretability of artificial intelligence systems, which would help people accept them into clinical practice.

Gradient-weighted Class Activation Mapping (Grad-CAM) is one of the most popular explainability techniques used in medical image analysis. It produces visual heatmaps that emphasize the areas of an image that have the greatest impact on a model's prediction. Using these

heatmaps, physicians can confirm if the model is concentrating on clinically significant anatomical features, such lesions, aberrant blood vessels, or injured retinal tissue.

Local Interpretable Model-agnostic Explanations (LIME), which approximates complex models using simpler and more interpretable models to explain model predictions, is another widely utilized method. Likewise, Shapley Additive Explanations (SHAP) give feature importance scores that measure the contribution of each input feature to the ultimate prediction. These explainability approaches are essential for increasing the interpretability of deep learning models employed in medical imaging.

Integrating XAI methods into deep learning frameworks has greatly increased the transparency, dependability, and reliability of diagnostic systems based on artificial intelligence. Explainable models promote ethical decision-making and aid regulatory approval for the use of AI technologies in healthcare settings, as well as boosting practitioner confidence.

Additionally, explainability helps clinical decision-making by offering visual proof that may be used to support diagnoses, track the course of the illness, and eventually enhance therapy plans. Because of this, Explainable Artificial Intelligence is now an essential element of contemporary medical imaging systems, especially in those used for high-risk clinical choices that necessitate a high degree of accountability and dependability.

Techniques for Ocular Disease Detection

Despite the great advancements in explainable artificial intelligence and deep learning, there are still many challenges in creating scalable and reliable diagnostic systems for detecting eye illnesses. The restricted amount of high-quality annotated medical datasets necessary to properly train deep learning models is one of the main challenges.

The accuracy and generalizability of deep learning models may be adversely impacted by variations in the quality, resolution, and acquisition settings of medical images obtained from various healthcare facilities. In addition, medical datasets frequently suffer from class imbalance, with some disease categories being underrepresented when compared to others. Due to this

imbalance, automated diagnostic systems may have lower performance in identifying uncommon diseases, which would impact their overall reliability, as well as skewed model predictions.

The excessive computing cost of training deep learning models is another major barrier. High- performance computing resources, such as graphics processing units (GPUs), are often necessary for training big neural networks in order to achieve efficient processing and quicker model convergence. The need for sophisticated computing infrastructure may restrict the availability and adoption of AI-based diagnostic systems in low-resource healthcare settings, especially in low- income nations.

Researchers are now concentrating on creating lightweight, energy-efficient deep learning architectures that may be implemented on mobile devices, embedded systems, and portable diagnostic platforms in order to address this problem for real-time disease detection and monitoring. Additionally, it is still essential to guarantee the reliability, fairness, and ethical use of artificial intelligence applications in healthcare.

To guarantee consistent performance across diverse populations, age categories, and demographic circumstances, deep learning models must be extensively validated using varied and representative datasets. Additionally, the use of AI technologies in medical practice must be governed by the establishment of uniform clinical standards and regulatory frameworks that ensure its safe and ethical implementation.

Future research is anticipated to concentrate on incorporating artificial intelligence systems into regular clinical processes, increasing the interpretability of models, and raising the bar for data quality. These improvements will promote the creation of individualized and preventative healthcare strategies, which will eventually lead to better patient results and more efficient contemporary healthcare systems.

1. PREVIOUS WORK

   J. Kadam A, et al. (2023) [1]: The paper introduced an early prediction model that utilizes retinal imaging and artificial intelligence to address the increasing global burden of cardiovascular diseases (CVDs). The study emphasizes the human retina as a non-invasive indicator of vascular health capable of identifying biomarkers associated with cardiovascular risk. In this work, high- quality retinal fundus images are acquired and pre-processed to enhance vascular visibility and minimize noise. Advanced deep learning techniques, particularly convolutional neural networks (CNNs), are employed to automatically detect significant retinal features such as vessel diameter, tortuosity, and microvascular abnormalities. To further improve prediction accuracy, the system integrates clinical information, including patient demographics and relevant health indicators. The findings demonstrate that AI-driven retinal analysis can provide reliable diagnostic outcomes, enabling early detection and timely medical intervention. Overall, the research highlights the effectiveness of combining retinal imaging with artificial intelligence as a non-invasive, cost- effective, and accessible approach for the early prediction of cardiovascular diseases in modern healthcare systems.

   Ghenciu L, et al. (2024) [2]: It emphasize the increasing significance of early identification of cardiovascular diseases (CVDs) using advanced retinal imaging techniques. The study focuses on Oculomics, which explores the relationship between retinal microvascular features and overall systemic vascular health. Techniques such as retinal fundus imaging and optical coherence tomography/angiography (OCT/OCTA) are employed to identify critical retinal biomarkers including vessel calibre, tortuosity, and branching patternsthat can indicate cardiovascular risk at an early stage. The authors further highlight the importance of artificial intelligence (AI) in processing the large volumes of retinal data generated during routine eye examinations. Their findings demonstrate that AI-based models can effectively predict cardiovascular risk factors, cardiovascular events, and metabolic disorders with strong performance metrics, achieving AUC values between 0.71 and 0.87, sensitivity ranging from 71% to 89%, and specificity between 40% and 70%. These results suggest that, in certain cases, AI-driven retinal analysis may outperform traditional diagnostic methods. Overall, the research underscores the potential of retinal imaging as a non-invasive, scalable, and cost-effective approach for personalized healthcare and early disease prevention. However, the authors also emphasize the need for further research to

   standardize imaging protocols and validate these biomarkers across diverse populations before widespread clinical implementation.

   Parmar, et al. (2024) [3]: conducted a comprehensive study discussing the revolutionary role of Artificial Intelligence (AI) in ophthalmology, particularly focusing on the diagnosis and treatment of retinal diseases. The study highlights the effectiveness of key AI concepts such as Machine Learning (ML) and Deep Learning (DL) in improving screening efficiency, enabling early disease detection, and enhancing patient outcomes. The authors explored the application of AI across a wide range of retinal disorders, including diabetic retinopathy, age-related macular degeneration, retinopathy of prematurity, retinal vein occlusion, hypertensive retinopathy, and various inherited retinal diseases. Furthermore, the review examined modern AI models, their performance metrics, and their clinical relevance in real-world healthcare settings. It also addressed critical challenges associated with AI implementation, such as the black-box phenomenon, data bias, and limitations in comprehensive patient evaluation. Overall, the study emphasizes that Artificial Intelligence should function collaboratively with healthcare professionals to support clinical decision-making and improve healthcare delivery, rather than replace human expertise.

   Aziz S, et al. (2023) [4]: It emphasizes the vital role that the healthcare industry plays in fostering economic growth, job creation, and overall global well-being. The authors highlight the increasing significance of advanced technologies, particularly Machine Learning (ML) and Artificial Intelligence (AI), in addressing persistent healthcare challenges, especially in the diagnosis of diabetic retinopathy and hypertensive retinopathy. To systematically analyse research trends in AI- based retinopathy diagnosis, the literature review utilizes a combination of Scientometric analysis and the conceptual tree of science paradigm. Scientometric analysis is employed to identify leading countries, institutions, and researchers contributing to innovation in automated retinal disease detection. Meanwhile, the evolutionary perspective traces the progression of foundational research and technological advancements over time. The study further highlights that AI-based diagnostic systems significantly enhance detection accuracy, enable early diagnosis, and improve patient treatment outcomes. Additionally, the authors note that the COVID-19 pandemic accelerated the adoption of remote healthcare technologies, thereby strengthening the role of AI in digital health and telemedicine solutions. Overall, the review underscores the continued

   importance of leveraging AI and ML to transform healthcare delivery, particularly in the diagnosis of retinal diseases.

   V. Lai, et al. (2025) [5]: They carried out research to develop a sophisticated Artificial Intelligence (AI)-based approach for identifying referable age-related macular degeneration (AMD), including intermediate and advanced stages, as well as neovascular AMD.The study also focused on the automatic segmentation of choroidal neovascularization (CNV) using colour fundus retinal images. Furthermore, the research investigated the relationship between retinal disease and brain health by utilizing AI-based Retinal Image Analysis (ARIA) to estimate brain health risk scores, such as white matter hyperintensities and depression. The dataset used in this study comprised 1,480 retinal images collected from Zhongshan Hospital of Fudan University, which were utilized for model training and validation through a 10-fold cross-validation procedure. To further evaluate the robustness and generalizability of the proposed model, two additional validation sub-datasets, each consisting of 238 images, were employed. The researchers applied fluorescein angiography- based labels to train the InceptionResNetV2 deep learning architecture for detecting referable and neovascular AMD. In addition, a transfer learning-based ResNet50 U-Net model was implemented to perform the segmentation of choroidal neovascularization (CNV). The experimental results demonstrated excellent diagnostic performance of the proposed system. During the cross- validation process, the model achieved sensitivities of 97.4% and 98.1%, specificities of 96.8% and 96.1%, and accuracies of 97.0% and 96.4% in detecting referable AMD and neovascular AMD, respectively. In external validation, the system achieved accuracies of 92.9% and 93.7%, with an Area Under the Curve (AUC) value of 0.967, indicating strong classification capability. For the segmentation of choroidal neovascularization (CNV), the model achieved a global accuracy of 93.03%, a mean accuracy of 91.83%, and mean Intersection over Union (IoU) of 68.7%, demonstrating reliable and consistent segmentation performance. Overall, the study concluded that the proposed AI-based retinal image analysis framework is a highly efficient and cost-effective screening tool for detecting referable and neovascular AMD in both retinal and wide- field images. The system shows significant potential for implementation in clinical and resource- constrained healthcare environments, where it can facilitate early disease detection, support timely referral decisions, and enhance patient care outcomes.

   Moreover, the integration of brain health risk assessment further highlights the broader clinical importance of retinal imaging in predicting neurological and systemic health conditions.

   Sobhi N, et al. (2025) [6]: It examined the increasing use of artificial intelligence (AI) in analyzing retinal images for the detection and management of complications associated with diabetes mellitus (DM). The study highlights that diabetes significantly increases the risk of vascular diseases and that retinal vascular imaging serves as an important biomarker for assessing both microvascular and macrovascular health. The authors noted that AI-enabled systems, initially developed for large-scale screening of diabetic retinopathy (DR), are now being extended to detect additional complications, including neuropathy, nephropathy, and atherosclerotic cardiovascular disease, as well as to predict future cardiovascular risks. The review was based on an extensive literature search across major scientific databases such as PubMed, Scopus, and Web of Science, focusing on studies related to diabetes, retinal imaging, and AI methodologies. The findings indicate that AI-driven retinal image analysis can support comprehensive patient care by enabling early diagnosis, risk assessment, and prognosis of systemic complications associated with diabetes. Furthermore, the study addresses key implementation challenges, including data privacy, ethical considerations, equitable access to healthcare, and model explainability. Overall, the authors conclude that AI-assisted retinal imaging has the potential to become a crucial component of modern personalized medicine by providing non-invasive, efficient, and predictive evaluation of diabetes-related health outcomes.

   Wang Y, et al. (2025) [7]: studied the function of artificial intelligence (AI) in forecasting the risk of cardiovascular disease (CVD). Early diagnosis and prevention are crucial for reducing the burden of illness, according to the study, which highlights that CVD is still the top killer worldwide. The authors systematically searched PubMed, the Web of Science Core Collection, and other major electronic databases, and chose 43 pertinent articles based on established inclusion and exclusion criteria. According to the review, using AI to analyses retinal imaging techniques like optical coherence tomography (OCT), optical coherence tomography angiography (OCTA), and colour fundus photography (CFP) shows great promise for non-invasive CVD risk assessment. The results indicate that using AI technology in conjunction with regular eye exams can facilitate early identification and wide-scale screening of cardiovascular risk. Additionally, the study shows that multimodal methods that combine several imaging sources have demonstrated superior

   performance, even if single-modality AI models have attained a high level of diagnostic accuracy. However, the authors also point out a number of drawbacks, such as reliance on datasets from a single centre and a restricted ability to generalize across communities.

   Jin K, et al. (2025) [8]: They examined the evolving role of artificial intelligence (AI), particularly advanced deep learning models, in transforming medical practice across various specialties, including ophthalmology. The study emphasizes that the eye, due to its unique microvascular and neural architecture, is closely connected to overall systemic health, making ocular imaging an effective tool for detecting systemic diseases. The authors highlight that multimodal ocular image analysis using AI can serve as a valuable alternative or complementary screening method, particularly in resource-constrained settings. The review discusses current AI applications that utilize retinal and other ocular imaging techniques to predict a wide range of systemic diseases, including cardiovascular disorders, dementia, chronic kidney disease, and anemia. Furthermore, the study identifies existing challenges, such as data limitations, model reliability, and difficulties in clinical integration, while also outlining future research directions to improve the accuracy, accessibility, and clinical utility of AI-driven ocular diagnostics. Overall, the authors conclude that AI-assisted ocular imaging holds significant potential for the early detection and monitoring of systemic diseases, thereby supporting the advancement of preventive and personalized healthcare.

   Garg N, et al. (2022) [9]: They conducted a comprehensive review identifying cardiovascular disease (CVD) as one of the leading causes of mortality and morbidity worldwide, thereby emphasizing the urgent need for innovative approaches for early risk identification and diagnosis. The authors examined recent advancements in Artificial Intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL) techniques, for predicting cardiovascular risk using diverse healthcare data sources. One of the primary objectives of the study was to investigate the application of retinal fundus imaging as a non-invasive and easily accessible method for cardiovascular risk assessment. The study demonstrated that retinal blood vessels share anatomical and physiological similarities with systemic vasculature, making retinal images valuable indicators of cardiovascular health. AI algorithms can automatically extract vascular biomarkers from fundus images, including vessel diameter, tortuosity, and branching patterns, thereby enabling the early detection of cardiovascular abnormalities. Furthermore, the authors highlighted significant advancements in AI-driven cardiovascular disease prediction systems, such as improved

   diagnostic accuracy, automated screening processes, and enhanced clinical decision support. However, the study also identified several challenges that limit the widespread adoption of AI- based diagnostic systems. These challenges include the limited availability of large annotated datasets, issues related to model interpretability, concerns regarding data privacy and security, and the need for extensive clinical validation before integration into routine healthcare practice. In conclusion, Garg N., et al. emphasized that integrating AI-based retinal image analysis into regular healthcare screening programs has significant potential to improve early cardiovascular risk prediction and support preventive healthcare strategies. The authors recommended that future research should focus on the development of interpretable AI models, integration of multimodal healthcare data, and real-world clinical implementation to enhance the reliability, scalability, and clinical acceptance of AI-assisted cardiovascular diagnostics.

   Xu J, et al. (2023) [10]: It summarized recent developments in the use of retinal imaging for Artificial Intelligence (AI)-assisted screening and prediction of systemic diseases. The study highlighted the growing application of AI technologies in analysing retinal images to identify and predict a wide range of systemic conditions, including cardiovascular disease, systemic lupus erythematosus, renal disease, and neurodegenerative disorders. The authors emphasized that AI- based retinal image analysis significantly improves screening accuracy and prediction efficiency by automatically detecting subtle retinal biomarkers associated with systemic health conditions. By reviewing multiple existing studies, the research demonstrated the effectiveness of advanced Machine Learning (ML) and Deep Learning (DL) models in enabling early diagnosis and risk assessment through non-invasive retinal imaging techniques. Furthermore, the study identified several existing challenges in the field, including the need for larger and more diverse datasets, improved model interpretability, standardized evaluation frameworks, and comprehensive clinical validation. Addressing these challenges is essential to ensure the safe and reliable deployment of AI systems in real-world healthcare environments. In conclusion, Xu, J., et al. advocated for collaborative efforts among clinicians, researchers, and policymakers to promote the integration of AI-enhanced retinal imaging into routine clinical practice. The authors recommended that future research should focus on improving algorithm transparency, expanding multimodal data integration, and developing regulatory guidelines to support the effective management and prevention of systemic diseases using AI technologies.

   1. Gap Analysis

Although retinal image analysis has made significant progress in artificial intelligence, there are still many challenges in developing trustworthy diagnostic systems for eye illnesses. Most current research focuses on individual diseases rather than numerous diseases. A large, deep learning model is necessary to accurately identify a variety of ocular diseases. The limitations include the lack of model generalization across diverse datasets, the opacity of predictions, and the scarcity of resources in low-resource environments. To enhance the accuracy and interpretability of identifying several ocular illnesses, the study intends to create a broad, understandable deep learning framework that will ultimately improve clinical decision-making and healthcare implementation.

1. RESTNET: RESIDUAL NETWORK

   The Residual Network (Rest Net), a deep convolutional neural network architecture developed by Kaiming He and his colleagues at Microsoft Research in 2015, was introduced during the ImageNet Large Scale Visual Recognition Challenge 2015. Because Rest Net addressed the problem of performance degradation caused by greater network depth while also enabling the successful training of extremely deep neural networks, its introduction marked a significant advancement in deep learning. Prior to Rest Net, increasing the number of layers in a neural network frequently resulted in decreased accuracy due to issues such as vanishing gradients and optimization difficulties.

   The primary novel concept of Rest Net is residual learning, which employs skip or shortcut connections that allow data to pass between one or more network layers. Instead of learning a direct connection between input and output, the network learns the residual function, which reflects the difference between the desired output and the input. These skip connections, which aid in maintaining the gradient flow during backpropagation, enable the efficient training of networks with hundreds of layers. Consequently, Rest Net outperforms traditional deep neural network architectures in terms of accuracy, convergence speed, and generalization ability.

   Each residual block in a Rest Net architecture includes batch normalization, convolutional layers, activation functions such as the Rectified Linear Unit (ReLU), and shortcut connections. These residual blocks serve as the network's foundation and make it easier to retrieve important features from images. Because of its strong feature learning capabilities, Rest Net is widely employed in a variety of computer vision applications, including retinal illness identification, medical image analysis, object detection, and image classification. Particularly in the field of medical imaging, ResNet is useful for identifying complex patterns in retinal images, which helps with the early detection and automated classification of diseases.

   A variety of ResNet variations have been developed in order to accommodate changes in network depth, computational complexity, and performance. These variations share the same residual learning notion but vary mostly in their number of layers and overall design. The most well-known

   variations of Rest Net: ResNet-18, ResNet-34, ResNet-50, ResNet-101, and ResNet-152 are each made to handle varying degrees of computational demand and dataset complexity.

   The 18-layer ResNet-18 is the smallest and most computationally efficient model. It uses simple residual blocks and is suitable for use cases with restricted computer resources or when working with smaller datasets. This version keeps consistency in image categorization assignments while providing reliable performance with reduced memory usage and faster training.

   ResNet-34 is a more sophisticated version of ResNet-18 that has 34 layers, which increases its capacity to extract features. It is ideal for medium-sized datasets and applications needing moderate performance since it strikes a balance between computational efficiency and model accuracy.

   One of the most common variations of ResNet is ResNet-50, which has 50 layers. It employs a bottleneck architecture that maintains high accuracy while reducing computational cost. Because of its outstanding feature extraction capabilities and efficient performance, ResNet-50 is often used in medical image analysis and transfer learning, such as retinal illness detection systems. For many deep learning applications, it is the preferred choice because it strikes a good balance between model complexity and classification accuracy.

   For more complicated image recognition applications, ResNet-101 is a 101-layer deeper architecture. It provides greater accuracy and better feature representation, but it also demands more processing power and a longer training period. This version is often used with massive datasets and in complex research projects where performance is more important than computational efficiency.

   With 152 layers, ResNet-152 is one of the deepest versions of the ResNet family. It can capture incredibly intricate image features with remarkable accuracy. However, it also requires significant computational resources and memory, making it suitable for large datasets and high-performance computing environments.

   Due to its capacity to efficiently train very deep networks, ResNet has established itself as one of the most important deep learning architectures in modern computer vision. Its variations provide the flexibility to select the appropriate model depending on the dataset size, computational

   resources, and application needs. ResNet-based models are extensively employed in retinal disease detection and medical image analysis because they facilitate feature extraction, support transfer learning, and improve the reliability of automated diagnostic systems.

   1. RESTNET50

      Overview

      The ResNet-50 deep convolutional neural network architecture is a member of the Residual Network (Rest Net) family, which was originally introduced by Kaiming He and his Microsoft Research team at the ImageNet Large Scale Visual Recognition Challenge 2015. The ResNet architecture was designed to address the drawbacks of traditional deep neural networks, particularly the vanishing gradient and degradation problems that occur as networks become very deep.

      The ResNet-50 model, for instance, makes use of residual learning through shortcut or skip connections, which enables the network to learn complex features while maintaining efficient gradient flow throughout the training process. There are 50 layers in the network. This architecture allows deep neural networks to achieve higher accuracy and better convergence than earlier convolutional neural network designs.

      In the field of retinal image-based eye disease identification, ResNet-50 has emerged as one of the most well-liked deep learning models because of its ability to extract hierarchical and discriminatory features from medical images. The model is often employed as a pre-trained network in transfer learning, where it uses knowledge acquired from large datasets and applies it to particular medical datasets.

      Due to its capability, ResNet-50 is particularly well-suited for developing interpretable and generalized deep learning frameworks for the automated identification of eye diseases like age- related macular degeneration, glaucoma, cataracts, and diabetic retinopathy. It is ideal for real- world healthcare applications because of its well-balanced accuracy, computational efficiency, and scalability.

      Architecture

      The ResNet-50 architecture makes extensive use of residual blocks, which are made up of several convolutional layers and form the foundation of the network. Each residual block consists of convolutional layers, batch normalization layers, activation functions like the Rectified Linear Unit (ReLU), and shortcut connections that allow the input to bypass certain layers. These shortcut connections, which include several layers, aid in the network's efficient training and gradient flow maintenance throughout backpropagation. The input in the ResNet-50 architecture is organized hierarchically. Prior to arriving at the final categorization layer, the image goes through a number of convolution and feature extraction phases.

      The typical architecture of ResNet-50 begins with an input layer that receives retinal images, followed by an initial convolution layer and a max-pooling layer that reduces the spatial size of the image while preserving crucial information. The network then passes through four main residual block phases, each of which extracts features at different levels of abstraction. As the network becomes deeper, the feature maps become more complex, capturing minute details in retinal images such as retinal anomalies, blood vessel structures, and lesions. The extracted features are then processed by a fully connected layer and a global average pooling layer, and the result is passed into a SoftMax classifier that determines the final category of eye ailment.

      In understandable deep learning frameworks, the ResNet-50 architecture may be combined with visualization techniques such as feature maps and attention mechanisms to enhance model interpretability. These approaches improve transparency and build confidence in automated diagnostic systems by helping clinicians understand which regions of the retinal image contribute to the model's decision.

      Advantages and Application

      Utilizing deep learning approaches, ResNet-50 provides several advantages for medical image analysis and the diagnosis of eye diseases. One of its key benefits is the capacity to train extremely deep neural networks without experiencing the vanishing gradient issue, which is made possible by the use of residual connections that preserve a steady gradient flow throughout the training process. The model is able to learn intricate visual patterns and minute details from retinal images

      thanks to this ability, which leads to more accurate categorization and dependable diagnostic results. Furthermore, ResNet-50 facilitates transfer learning, which enables pre-trained models to be customized to medical datasets with sparse labelled data, lowering the training time and computational expense. The scalability and flexibility of ResNet-50, which enables it to be incorporated into generalized deep learning architectures for the simultaneous detection of several eye diseases, is another major benefit.

      The architecture is also compatible with explainable artificial intelligence methodologies, such as visualization approaches and attention maps, that increase the interpretability of model predictions and improve transparency in clinical decision-making. A variety of medical imaging applications have made extensive use of ResNet-50. Retinal disorders, such as diabetic retinopathy, glaucoma, hypertensive retinopathy, macular degeneration, and retinal detachment, are often detected and classified using it, especially in the field of ophthalmology. The model is used outside of ophthalmology in other fields like radiology image analysis, tumour identification, and general medical diagnosis, proving its value as a strong backbone model for healthcare applications based on deep learning.

      Due to its residual learning mechanism, which allows for deeper networks to achieve higher performance without increasing training complexity or sacrificing accuracy, ResNet-50 is a significant improvement over prior convolutional neural network architecture like Alex Net and VGGNet. Furthermore, ResNet-50's bottleneck design minimizes the number of parameters while preserving significant computational efficiency, enabling the model to handle enormous datasets more effectively. When employing retinal images to identify eye illnesses, ResNet-50 offers better detection of minor symptoms like microaneurysms, haemorrhages, and optic nerve injury than conventional machine learning and shallow neural network methods. By integrating explainable artificial intelligence approaches with ResNet-50, the dependability, transparency, and clinical relevance of deep learning systems have been improved, making them ideal for practical medical and diagnostic use.

2. VGG-16: VISUAL GEOMETRY GROUP

   It is a profound convolutional neural network design created by the University of Oxford's Visual Geometry Group. The model makes use of tiny 3×3 convolution filters to retrieve useful features from images and is made up of 16 layers, 13 of which are convolutional layers and 3 of which are fully connected. VGG-16 is simple to implement and works well for image categorization due to its straightforward and consistent design.

   For the purpose of detecting ocular disorders using retinal images, VGG-16 is frequently used as a pretrained model for transfer learning, which allows for the identification of retinal anomalies such as lesions, haemorrhages, and optic disc alterations. Its deep architecture facilitates trustworthy feature extraction and enhances classification accuracy, making it ideal for medical image analysis and automated diagnostic systems.

   Due to its consistent performance and well-organized architecture, VGG-16 also serves as a solid baseline model in deep learning frameworks for healthcare applications.

   Fig – 4.2 VGG-16: VISUAL GEOMETRY GROUP

3. EFFICIENTNET-B7

   Researchers at Google created the sophisticated deep learning architecture known as EfficientNet- B7, which employs a compound scaling technique to concurrently maximize the network's depth, width, and input resolution. This model is a member of the Efficient Net family and is designed to perform at a high level while maintaining computational efficiency and classification accuracy. EfficientNet-B7 uses sophisticated optimization methods along with Mobile Inverted Bottleneck Convolution (MBConv) layers to improve feature extraction and minimize computational complexity in image classification tasks.

   EfficientNet-B7 is especially useful in the field of retinal image-based ocular disease diagnosis for analyzing high-resolution retinal images and identifying minute pathological characteristics, such as micro-aneurysms, exudates, and retinal vascular abnormalities. For the timely identification and categorization of eye illnesses, these capabilities are critical.

   EfficientNet-B7 is considered a powerful and efficient model for generalized deep learning frameworks in medical image analysis and automated healthcare diagnostic systems because of its ability to achieve high accuracy while efficiently utilizing computational resources.

   Fig- 4.3 EFFICIENTNET-B7

   In the field of ocular disease detection using retinal images, EfficientNet-B7 is particularly effective for analyzing high-resolution retinal images and detecting subtle pathological features such as microaneurysms, exudates, and retinal vascular abnormalities. These capabilities are essential for the early diagnosis and classification of eye diseases. Due to its ability to deliver high accuracy while efficiently utilizing computational resources, EfficientNet-B7 is considered a suitable and powerful model for generalized deep learning frameworks in medical image analysis and automated healthcare diagnostic systems.

4. XCEPTION

   XCEPTION The deep convolutional neural network architecture known as Xception (Extreme Inception) was created by François Chollet at Google. The model, which is an extension of the Inception architecture, employs depth wise separable convolutions rather than standard convolution operations in order to improve computational efficiency and feature extraction. The three main components of the Xception architecturethe Entry Flow, Middle Flow, and Exit Flowwork together to progressively extract complex features from images. Xception is often used in retinal image analysis to diagnose eye disorders because of its high accuracy and effective learning skills. Its design supports understandable artificial intelligence methodologies, which makes it perfect for developing transparent and reliable diagnostic systems for healthcare uses.

   The Xception architecture's Entry Flow is in charge of the preliminary feature extraction from the input image. It progressively minimizes spatial dimensions while capturing important low-level features, such as edges, textures, and fundamental anatomical structures, by processing raw retinal images through a series of convolutional and pooling layers. At this point, the input is transformed into a smaller, more informative form in preparation for further processing.

   The core component of the Xception architecture is the Middle Flow, which uses depth wise separable convolutions repeatedly across several identical modules. Due to this repetitive pattern, the network is able to learn complicated and abstract feature representations while remaining computationally efficient. This stage is essential for detecting subtle abnormalities, such as

   microvascular changes, lesions, and uneven patterns, in retinal image analysis that might point to different eye illnesses.

   The last step of the Xception model is the Exit Flow, which uses the high-level features collected from prior layers and prepares them for categorization. Before feeding the feature maps into a SoftMax classifier, this step further lowers their dimension and sends them through global pooling layers or fully connected layers. Based on the learned characteristics, the classifier then calculates the likelihood of various categories of eye illness. The usage of depth wise separable convolutions, which greatly lowers the number of parameters and computing cost when compared to conventional convolutional networks, is one of Xception's main advantages. The model's efficiency makes it appropriate for use in real-time diagnostics and large-scale medical image datasets since it can be trained more quickly without sacrificing accuracy.

   Fig – 4.4 XCEPTION

   In addition to its efficiency, Xception allows for integration with explainable artificial intelligence methods like Grad-CAM and feature visualization approaches. These methods help to emphasize the key areas in retinal images that have an impact on the model's predictions, increasing transparency and fostering confidence among healthcare professionals. Consequently, Xception is a robust and dependable deep learning model for creating cutting-edge, understandable, and adaptable diagnostic systems in today's healthcare environment.

5. PROPOSED MODEL

   Fig – 4.5 PROPOSED MODEL

   1. Retinal image as input

      The system starts by gathering retinal photographs, which act as its main input. These pictures are usually taken with fundus cameras or other retinal imaging tools and provide vital visual data on the retina's architecture, such as blood vessels, the optic disc, and any possible anomalies. The diagnostic model's overall performance depends greatly on the quality and variety of these pictures.

   2. Processing Image

      The input images are subjected to preprocessing procedures at this point, including noise reduction, picture improvement, resizing, and normalization. Normalization standardizes pixel values for stable training, while resizing guarantees consistent image dimensions for deep learning models. Enhancement techniques improve image clarity, making crucial retinal features more visible and simpler for the model to learn, while noise reduction gets rid of undesirable distortions.

   3. Data Enhancement

      To artificially expand the size and diversity of the dataset, data augmentation is used. The same picture may have several variations made using techniques like rotation, flipping, and scaling. This helps prevent overfitting, improves model generalization, and ensures that the model can handle variations in image orientation and scale encountered in real-world situations.

   4. Extracting Features

      At this point, valuable features are automatically extracted from the retinal images using deep learning models like Xception, EfficientNet-B7, VGG-16, or ResNet-50. These models can detect patterns like lesions, anomalies, and the structure of blood arteries. The process of extracting features converts raw image data into high-level representations that are essential for correctly classifying diseases.

   5. Layers of full connectivity

      The retrieved characteristics are then fed via fully connected layers, which function as a classifier by learning intricate interactions between features. By combining and processing the features, these layers assist the model in differentiating between various disease categories and getting them ready for the last choice.

   6. SoftMax Classifier

      The SoftMax classifier transforms the output of the fully connected layers into probability values for each potential disease class. It gives each class a probability score, making sure that the total of all probabilities is one. The predicted illness is the class with the greatest likelihood of occurring.

   7. Disease Prediction

      The model predicts the ultimate illness using the SoftMax output. This step determines if the retinal image represents a normal condition or a particular eye ailment like diabetic retinopathy, glaucoma, or macular degeneration.

   8. The Explainable AI Module

      To increase interpretability, the system includes an explainable AI module like Grad-CAM visualization. It draws attention to the key areas of the retinal image that have an impact on the model's prediction. By allowing clinicians to comprehend the rationale behind the choice, this contributes to enhancing trust and transparency in the diagnostic process.

   9. Evaluation of Performance

Lastly, the model's performance is assessed using metrics like accuracy, precision, recall, F1-score, and confusion matrix. These indicators evaluate the model's ability to forecast various disease classes and aid in determining its strengths and shortcomings, thereby assuring its reliability prior to actual implementation.

1. METHEDOLOGY

   1. Workflow

      Fig- 5.1.1 Workflow

      STEPS:

      1. Retinal Fundus Dataset Collection

      2. Image Enhancement

      3. Image Segmentation

      4. Image Augmentation

      5. Processed Retinal Dataset

      6. Transfer Learning Model Training

      7. Model Evaluation

      8. Model Comparison and Result Generation

      9. Best Model Selection

      10. Proposed DCNN Model Development 11.Classification Result

   2. Overview

      The retinal fundus image dataset contains images of both healthy and pathological conditions, such as cataracts, diabetic retinopathy, and glaucoma. Image quality is enhanced and important components are segmented during preprocessing. Increased variety results from data augmentation. For feature extraction and classification, transfer learning is employed using models such as ResNet50 and EfficientNet-B7. The optimum model for classifying retinal images is chosen as EfficientNet-B7.

   3. General System Design

      The system is organized as a modular pipeline for the automatic detection of retinal illnesses. First, the retinal fundus dataset is entered, then it is pre-processed using image enhancement and segmentation methods to improve its quality and extract key features. The dataset is then increased, and the models are made more generalizable by using augmentation methods. Several pre-trained transfer learning models, such as ResNet50, VGG-16, Xception, and EfficientNet-B7, are trained using this processed dataset. Each model is evaluated using standard classification measures, which allows for a systematic comparison. Using the highest-performing EfficientNet-B7 model as a basis, a specialized DCNN is then constructed to provide the best categorization of the condition, producing trustworthy and precise results for healthy retinal images of CA, DR, or GL. This approach provides excellent accuracy in the identification of eye diseases, as well as efficiency and scalability.

   4. Data Collection

      The creation of an effective ocular disease categorization system, particularly one that employs deep learning methodologies, depends on data collection. The datasets must be of high quality and diversity in order to ensure that the model is accurate, reliable, and applicable in a wide range of real-world situations.

      The Ocular Disease Classification Dataset, which was used for this research, came from Rob flow Universe( https://universe.roboflow.com/fyp-oq1go/ocular-disease-classification)

      This dataset contains 888 images of the retinal fundus, which are divided into six categories: Diabetes, Hypertension, Glaucoma, Cataracts, Age-Related Macular Degeneration (AMD), and Normal. The dataset is meant to be used for image categorization in the field of detecting systemic and eye diseases.

      The dataset is divided into an 80:20 ratio, with 710 training images and 178 testing images. With this arrangement, the model can learn and precisely evaluate its performance on previously unseen images. For successful classification, the thorough retinal fundus images aid in detecting early indicators of systemic and eye illnesses.

      Fig- 5.1.4 Data Collection

   5. Image processing

      Image Enhancement

      Before retinal fundus pictures are used in deep learning model training, the proposed system places a lot of importance on image processing as a way to improve their quality and usability. The three main stages of image segmentation, image augmentation, and image enhancement all play distinct roles in the overall performance of the classification system.Image enhancement is the first step towards improving the aesthetic appeal of unprocessed retinal fundus images and making crucial elements more apparent. Medical images are often plagued by inadequate lighting, noise, and poor contrast, all of which can negatively impact the correctness of models.

      Some of the strategies for improvement are as follows:

      Brightness Enhancement: This changes the image's intensity levels to make darker areas more apparent.

      Increased Contrast: Improves the contrast between bright and dark areas, which makes it easier to see structures such as blood arteries.

      Noise Reduction: Eliminates unwanted graininess or distortions from the picture by employing filtering techniques, which produce clearer and softer images.

      Image Augmentation

      The direct application of image improvement follows dataset collection in the figure provided. This improves the reliability of feature extraction by ensuring that all images are uniform and clearer before being processed in later stages. Using image augmentation, the dataset's size and diversity are artificially increased without adding any new images. This is particularly critical in medical imaging, where data may be scarce. Augmentation reduces the likelihood of overfitting and improves the model's ability to extrapolate its results to new data.

      Rotation: To mimic the numerous eye orientations when capturing a picture, images are rotated at different angles. Images are flipped horizontally or vertically to create mirror images.

      Zooming: Replicates different distances from the retina by zooming in or out of the image.

      Translation: Slightly shifts the image in various directions to create positional changes.

      The diagram employs augmentation following improvement to guarantee that the enhanced photographs are utilized to produce a variety of realistic versions. The training models are then fed these improved images, increasing their accuracy and resilience.

      Image Segmentation

      The objective of picture segmentation is to find and separate essential regions within the retinal image. By segmenting the image, the model can focus on clinically relevant characteristics rather than analysing the entire image. The primary areas of segmentation include:

      Blood Vessels: The process of isolating vital vascular structures that aid in the diagnosis of illnesses. The optic disc is essential. It is essential to locate the optic disc region in order to diagnose glaucoma.

      Lesions: Uncovering uncommon features, such as bleeding, microaneurysms, or exudates.

   6. Training Model

      During the training model phase, retinal fundus images are classified into categories such as CA, DR, GL, and Normal using transfer learning with pre-trained convolutional neural networks (CNNs). Instead of being built from scratch, the current models utilized in medical imaging are changed from those that have been trained on large datasets like ImageNet.

      ResNet50: In order to solve the problem of the vanishing gradient, ResNet50, a deep CNN with 50 layers, makes use of residual connections (skip connections). In the model, it works as follows: Retinal pictures are processed by convolutional layers. Residual blocks permit a direct gradient flow between layers. It gains hierarchical features like borders, textures, and patterns of disease. The last fully connected layer is modified to support 4-class classification (CA, DR, GL, Normal). Purpose of use: Works well on complex structures and provides effective training regardless of the number of layers.

      VGG-16: The 16 layers of the VGG-16 CNN are made up of small 3×3 convolution filters that are arranged in a deep stack. This is how the model functions: A sequence of pooling and convolution layers are used to process images. It produces precise spatial features such as lesions and vascular architecture. Categorization is done by fully connected layers.

      Purpose for use: It has a simple and consistent design and serves as an excellent baseline for comparison.

      Xception: In terms of effectiveness and performance, Xception surpasses ordinary CNNs by utilizing depth wise separable convolutions. The model's mode of operation: The convolution is divided into depth wise convolution for spatial filtering and pointwise convolution for feature combination. It accurately depicts complex retinal patterns while maintaining precision and reducing computational cost.

      Purpose for its use: It provides improved feature extraction with fewer parameters and delivers outstanding performance in medical imaging tasks.

      EfficientNet-B7: The complex CNN known as EfficientNet-B7 strikes a balance between depth, width, and resolution using compound scaling.

      The model's mechanism: It uses an optimized architecture that balances network depth (layers), width (channels), and image resolution. It is capable of identifying numerous minute details in retinal images and achieves higher accuracy with fewer parameters compared to traditional models. Purpose for its use: EfficientNet-B7 aims for high accuracy in image classification while being computationally efficient. It uses a compound scaling approach to learn complex patterns. It is useful in medical image analysis for detecting details like microaneurysms. EfficientNet-B7 is effective with fewer parameters, making it suitable for real-time diagnostics and large datasets. It supports transfer learning for faster training and better model generalization, making it ideal for accurate and scalable disease categorization systems.

   7. Model Evaluation

The retinal fundus dataset contains images of glaucoma, diabetic retinopathy, and cataracts, including both healthy and unhealthy cases. By segmenting key retinal components, reducing noise, and adjusting brightness and contrast, preprocessing improves image quality. Data augmentation techniques such as rotation, flipping, scaling, and translation enhance model performance by increasing dataset diversity and improving model robustness.

Using pre-trained models such as ResNet50, VGG-16, Xception, and EfficientNet-B7, transfer learning is applied to extract features and perform classification. The models are evaluated using metrics such as accuracy, precision, recall, F1 score, and specificity. EfficientNet-B7 serves as the foundation for the deep convolutional neural network used to classify retinal images.

Furthermore, techniques like cross-validation are used during the evaluation process to improve robustness and minimize bias in performance assessment. Visualization methods such as ROC (Receiver Operating Characteristic) curves and AUC (Area Under the Curve) help analyze the trade-off between sensitivity and specificity at different threshold values. These analyses further explore the diagnostic capabilities of the model. In the model comparison and result generation phase, results from all models are systematically compared, considering the advantages and limitations of each architecture. EfficientNet-B7 outperforms the other models due to its advanced scaling strategy and superior feature representation capability. As a result, it is selected as the best- performing model for the final classification task. This comprehensive evaluation process ensures that the chosen model is reliable, accurate, and suitable for real-world clinical applications, ultimately enhancing the effectiveness of automated retinal disease diagnosis.

1. RESULT

   1. METRICS EVALUATION

      Distribution of Training/Testing (80:20):

      The dataset used to identify diseases was split into an 80:20 ratio, with 80% used to train the model and 20% set aside for testing. The model is trained on a large enough dataset thanks to this split, and an independent section is kept aside for fair testing. The model's capacity to generalize to unseen retinal images is evaluated using the testing set, which provides a trustworthy indication of its actual performance.

      Accuracy:

      The suggested model for identifying diseases shows excellent categorization performance across all categories. The Normal class had the greatest accuracy, at 95.98%, among the classes that were assessed, demonstrating the model's outstanding ability to differentiate between images of healthy retinas. Additionally, the model produced remarkable accuracy for Diabetic Retinopathy (90.98%) and Glaucoma (92.76%), whereas Cataract had 89.95% accuracy. The model is able to accurately categorize retinal diseases and non-diseases, as seen by these findings.

      Sensitivity(Recall):

      The model's capacity to accurately identify positive disease cases is assessed through sensitivity. With 95% of cases correctly identified, Glaucoma had the highest sensitivity, indicating that it has exceptional detection ability for this disease. Normal (85%) and Cataract (80.76%) were next in line with Diabetic Retinopathy, which had a high sensitivity of 88%. Because of its comparatively lower sensitivity, cataract may result in false negatives by failing to identify some genuine instances, which is a crucial factor to take into account when making a medical diagnosis.

      Specificity:

      The model's capacity to accurately recognize negative events is measured by its specificity. With a specificity of 90.67%, cataract had the best performance in accurately identifying non-cataract instances. Diabetic Retinopathy had 85% specificity, while Glaucoma (87.95%) and Normal (89%) also had high specificity. According to these findings, the model maintains a comparatively low rate of false positives across various disease categories.

      DED	ACC	SEN	SPE	PRE
      CA	89.95	80.76	90.67	78
      DR	90.98	88	85	84
      GL	92.76	95	87.95	92
      NOR	95.98	85	89	80

      TABLE-1 MODEL PERFORMANCE METRICS

      FIG-6.1 GRAPHICAL REPRESENTATION OF PERFORMANCE METRICS

      120

      100

      80

      60

      40

      20

      0

      ACC

      SEN

      SPE

      PRE

      CA DR GL NOR

   2. Comparison with other models

      The table compares the performance of four deep learning modelsResNet50, VGG16, Xception, and EfficientNetB7across four classes: Cataract (CA), Diabetic Retinopathy (DR), Glaucoma (GL), and Normal. For every illness category, EfficientNetB7 consistently outperforms all other models in terms of accuracy, sensitivity, specificity, and precision. With the greatest accuracy values of 89.95% (CA), 90.98% (DR), 92.76% (GL), and 95.98% (Normal), it demonstrates its exceptional ability to classify retinal disorders. Xception exhibits variability, with a relatively high sensitivity in some cases but a lower overall accuracy, while VGG16 and ResNet50 show moderate performance. Overall, the findings indicate that EfficientNetB7 is the most efficient model for disease identification in this study, since it offers the most consistent and well-rounded performance across all measures.

      DED	MODEL	ACC%	SEN%	SPE%	PRE%
      CA	RestNet50	64.57	69	59	62
      	VGG 16	67.89	74	66	74
      	Xception	63.76	78.95	57.65	82
      	EfficientNetB7	89.95	80.76	90.67	78
      DR	RestNet50	63.05	73	70	68
      	VGG 16	69.67	76	66	69
      	Xception	57.85	85	58	52
      	Efficient NetB7	90.98	88	85	84
      GL	RestNet50	66.57	85.98	59.65	82
      	VGG 16	69.67	83	64.78	86
      	Xception	85.89	80	84	08
      	Efficient NetB7	92.76	95	87.95	92
      NORMAL	RestNet50	64.65	68	63	61
      	VGG 16	67.59	74	65	67
      	Xception	63.78	77.49	67.76	81
      	Efficient NetB7	95.98	85	89	80

      TABLE -2 COMPARISION WITH OTHER MODEL

   3. Confusion Matrix

      A comparative performance analysis of four deep learning modelsResNet50, VGG16, Xception, and EfficientNetB7across four categories: Cataract (CA), Diabetic Retinopathy (DR), Glaucoma (GL), and Normalis presented in the table. In terms of accuracy, sensitivity, specificity, and precision across all disease categories, EfficientNetB7 consistently outperforms other models. It demonstrates its superior ability in classifying retinal diseases by attaining the best accuracy rates of 89.95% (CA), 90.98% (DR), 92.76% (GL), and 95.98% (Normal). Xception's performance varies, exhibiting comparatively high sensitivity in some instances but lower overall accuracy, while VGG16 and ResNet50 perform somewhat well. In summary, the findings show that EfficientNetB7 is the most efficient model for identifying diseases in this study because it delivers the most consistent and balanced performance across all metrics.

   4. OUTPUT

NORMAL GLUCOMA

DIABETIC RETINOPATHY CATARACT

The output images represent the classification results of the proposed retinal disease detection model across four categories: Normal, Glaucoma, Diabetic Retinopathy, and Cataract. Each image illustrates distinct visual characteristics associated with the respective condition. The Normal image shows a clear retinal structure with well-defined blood vessels and optic disc, indicating a healthy eye. The Glaucoma image highlights change around the optic nerve region, which may indicate increased intraocular pressure. The Diabetic Retinopathy image exhibits visible lesions and vascular abnormalities, reflecting damage caused by diabetes. In contrast, the Cataract image appears blurred and hazy, representing lens opacity that reduces image clarity. These outputs demonstrate the models ability to effectively differentiate between various retinal conditions based on their visual features.

FIG- 6.1.4 ROC CURVE

The figure illustrates the ROC (Receiver Operating Characteristic) curves for multiclass classification of retinal diseases, including Cataract (CA), Diabetic Retinopathy (DR), Glaucoma (GL), and Normal. The ROC curve plots the True Positive Rate (Sensitivity) against the False Positive Rate, providing insight into the models classification performance.

All four curves lie significantly above the diagonal reference line (random classifier), indicating that the model performs well in distinguishing between classes. Among them, the Cataract (CA) and Glaucoma (GL) curves are closest to the top-left corner, reflecting superior performance with high true positive rates and low false positive rates. Their corresponding accuracies are approximately 0.94 (CA) and 0.93 (GL), demonstrating strong classification capability.

The Normal class also shows good performance with an accuracy of 0.90, while Diabetic Retinopathy (DR) has slightly lower performance with an accuracy of 0.88, as indicated by its curve being slightly farther from the top- left corner compared to others.

Overall, the ROC curves confirm that the model achieves high discriminative power across all classes, with particularly strong performance for Cataract and Glaucoma detection

1. CONCLUSION AND FUTURE WORK

   1. Conclusion

      The proposed generalized and explainable deep learning framework uses retinal images to demonstrate a viable approach to automated ocular disease diagnosis. By integrating image preprocessing, segmentation, augmentation, and transfer learning techniques, the system achieves superior classification performance. The use of multiple pre-trained models enables comparative analysis, aiding in the selection of the most effective architecture. In particular, EfficientNet-B7 demonstrates excellent performance in feature extraction and accuracy. Evaluation metrics are incorporated to ensure a comprehensive assessment of model performance. Additionally, explainability is emphasized, as it is essential for clinical acceptance. Although the system has certain limitations, it shows strong potential in assisting ophthalmologists. It supports the early detection and prevention of vision-related diseases. Due to its flexibility and scalability, the framework can be applied to a wide range of medical imaging applications. Overall, it represents a significant step toward AI-driven healthcare solutions.

   2. Future Work

Future studies employing larger and more diverse retinal datasets will help enhance model generalization. Clinical applications can be made easier to interpret by incorporating advanced explainable AI techniques. Model optimization methods can also be applied to reduce computational complexity, enabling real-time deployment. Integration with cloud-based and mobile systems can further improve accessibility, particularly in remote areas. Additionally, future research can explore the detection of a wider range of systemic and ocular diseases, thereby extending the scope of the framework.

1. REFERNCES

   The references used in this study are collected from ( https://universe.roboflow.com/fyp- oq1go/ocular-disease-classification)

   1. Kadam, A. J., Gaikwad, D. P., & Gaikwad, N. C. (2025). Integrating retinal imaging and AI for early cardiovascular disease prediction. Acero, 7(11).

   2. Ghenciu, L. A., Dima, M., Stoicescu, E. R., Iacob, R., Boru, C., & Henegan, O. A. (2024). Retinal imaging-based Oculomics: Artificial intelligence as a tool in the diagnosis of cardiovascular and metabolic diseases. Biomedicines, 12(9), 2150.

      /li>

   3. Parmar, U. P. S., Surico, P. L., Singh, R. B., Romano, F., Salati, C., Spadea, L., … & Zeppieri,

      M. (2024). Artificial intelligence (AI) for early diagnosis of retinal diseases. Medicina, 60(4), 527.

   4. Urina-Triana, M. A., Pinares-Melo, M. A., Mantilla-Morren, M., Butt-Aziz, S., Galeano-Muñoz, L., Naz, S., & Ariza-Clopas, P. P. (2024). Machine learning and AI approaches for analyzing diabetic and hypertensive retinopathy in ocular images: A literature review. IEEE Access, 12, 5459054607.

   5. Shi, C., Lee, J., Shi, D., Wang, G., Yuan, F., Lai, T. Y., … & Zee, B. C. Y. (2025). AI-based

      retinal image analysis for the detection of choroidal neovascular age-related macular degeneration (AMD) and its association with brain health. Brain Sciences, 15(11), 1249.

   6. Sobhi, N., Sadeghi-Bazargani, Y., Mirzaei, M., Abdollahi, M., Jafarzadeh, A., Pedram Mehr, S.,

      … & Acharya, U. R. (2025). Artificial intelligence for early detection of diabetes mellitus compactions via retinal imaging. Journal of Diabetes & Metabolic Disorders, 24(1), 104.

   7. Wang, Y., Yang, W., & Li, Y. (2025). Research advances on artificial intelligence-assisted diagnosis and risk assessment in cardiovascular disease using retinal imaging. Frontiers in Cardiovascular Medicine, 12,.

   8. Li, H., Cao, J., Grzybowski, A., Jin, K., Lou, L., & Ye, J. (2023). Diagnosing systemic disorders with AI algorithms based on ocular images. In Healthcare (Vol. 11, No. 12, p. 1739). MDPI.

   9. Kumar, A., & Singh, D. (2025). Diagnosis and prediction of cardiovascular risk in retinal imaging using artificial intelligence. In Artificial Intelligence in Modern Healthcare System (pp. 133159). Singapore: Springer Nature Singapore.

   10. Fang, P., Wu, Y., He, Y., Li, H., Guan, Z., Wang, X., & Shen, J. (2025). Research progress on AI-assisted screening and prediction of systemic diseases based on retinal images. The Visual Computer, 41(12), 95099537.

   11. Jackson, V. E., Wu, Y., Bonelli, R., Owen, J. P., Scott, L. W., Farashi, S., … & Egan, C. (2025). Multi-omics spatial effects on high-resolution AI-derived retinal thickness. Nature Communications.

   12. Yao, J., Hong, A. S. Y., Fukkatsu, K., & Ting, D. S. W. (2025). Artificial intelligence Oculomics for systemic health and longevity medicine: 2025 and beyond. Current Opinion in Ophthalmology.

   13. Rahman, F., Rahman, A., Talha, M., Irshad, N. U. N., & Imran, S. B. (2025). AI-powered Oculomics for early diagnosis of neurodegenerative diseases through retinal microvasculature analysis. Annals of Medicine and Surgery.

   14. Zhu, Z., Wang, Y., Qi, Z., Hu, W., Zhang, X., Wagner, S. K., & Wong, T. Y. (2025). Oculomics: Current concepts and evidence. Progress in Retinal and Eye Research.

   15. An, S., Teo, K., McConnell, M. V., Marshall, J., Galloway, C., & Squirrell, D. (2025). AI explainability in Oculomics: Role in establishing trust and future challenges. Progress in Retinal and Eye Research.

   16. Akpinar, M. H., Sengur, A., Faust, O., Tong, L., Molinari, F., & Acharya, U. R. (2024). Artificial intelligence in retinal screening using OCT images: A review of the last decade (2013 2023). Computer Methods and Programs in Biomedicine.

   17. Arnould, L., Meriaudeau, F., Genecia, C., Germanene, C., Delcourt, C., Kawasaki, R., & Cheung, C. Y. (2023). Using artificial intelligence to analyze the retinal vascular network: The future of cardiovascular risk assessment based on culomics? Ophthalmology and Therapy.

   18. Ting, D. S. W., Pasquale, L. R., Peng, L., Campbell, J. P., Lee, A. Y., Raman, R., … & Wong,

       T. Y. (2023). Artificial intelligence and deep learning in ophthalmology. British Journal of Ophthalmology.

   19. Gulshan, V., Peng, L., Coram, M., Stumpe, M. C., Wu, D., Narayanaswamy, A., & Webster,

       D. R. (2023). Development and validation of a deep learning algorithm for detection of diabetic retinopathy. JAMA.

   20. Poplin, R., Varadarajan, A. V., Blumer, K., Liu, Y., McConnell, M. V., Corrado, G. S., … & Webster, D. R. (2023). Prediction of cardiovascular risk factors from retinal fundus photographs using deep learning. Nature Biomedical Engineering.

   21. Keel, S., Lee, P. Y., Scheetz, J., Li, Z., Kotowicz, M. A., MacIsaac, R. J., & He, M. (2023). Development and validation of a deep learning system for detecting glaucomatous optic neuropathy. Ophthalmology.

   22. De Fauw, J., Ledsam, J. R., Romera-Paredes, B., Nikolov, S., Tomasev, N., Blackwell, S., … & Ronneberger, O. (2023). Clinically applicable deep learning for diagnosis and referral in retinal disease. Nature Medicine.

   23. Kermany, D. S., Goldbaum, M., Cai, W., Valentim, C. C., Liang, H., Baxter, S. L., … & Zhang,

K. (2023). Identifying medical diagnoses and treatable diseases using deep learning on retinal images. Cell.

Page 2 of 61 – Integrity Overview Submission ID trn:oid:::3618:135413136

19% Overall Similarity

The combined total of all matches, including overlapping sources, for each database.

Filtered from the Report

Bibliography Cited Text

Abstract

Match Groups

67 Not Cited or Quoted 19%

Matches with neither in-text citation nor quotation marks

0 Missing Quotations 0%

Matches that are still very similar to source material

0 Missing Citation 0%

Matches that have quotation marks, but no in-text citation

0 Cited and Quoted 0%

Matches with in-text citation present, but no quotation marks

Top Sources

13% Internet sources

16% Publications

0% Submitted works (Student Papers)

Integrity Flags

Our system's algorithms look deeply at a document for any inconsistencies that would set it apart from a normal submission. If we notice something strange, we 1ag it for you to review.

A Flag is not necessarily an indicator of a problem. However, we'd recommend you focus your attention there for further review.

0 Integrity Flags for Review

No suspicious text manipulations found.

Match Groups

Page 3 of 61 – Integrity Overview Submission ID trn:oid:::3618:135413136

Top Sources

67 Not Cited or Quoted 19%

Matches with neither in-text citation nor quotation marks

0 Missing Quotations 0%

Matches that are still very similar to source material

0 Missing Citation 0%

Matches that have quotation marks, but no in-text citation

0 Cited and Quoted 0%

Matches with in-text citation present, but no quotation marks

13% Internet sources

16% Publications

0% Submitted works (Student Papers)

Top Sources

The sources with the highest number of matches within the submission. Overlapping sources will not be displayed.

1

Internet

Publication

2

www.mdpi.com 1%

3

Publication

"Advanced Computing and Intelligent Technologies", Springer Science and Busin. 1%

4

Publication

"Proceedings of International Conference on Information Technology and Intellig. <1%

5

Internet

Suman Lata Tripathi, Om Prakash Kumar, Allwin Devaraj Stalin, Tanweer Ali. "Inn. <1%

6

Publication

mdpi-res.com <1%

7

Internet

Manoj Kumar, Tanweer Ali, Jaume Anguera, Suman Lata Trpathi. "Emerging Tech. <1%

8

Publication

www.medrxiv.org <1%

Internet

9

Tasneem Ahmed, Shrish Bajpai, Mohammad Faisal, Suman Lata Tripathi. "Advanc. <1%

Internet

10

www.frontiersin.org <1%

assets-eu.researchsquare.com <1%

Page 4 of 61 – Integrity Overview Submission ID trn:oid:::3618:135413136

11

Internet

12

Publication

www.expresscomputer.in <1%

13

Publication

Su, Juntao. "Explainable and Robust AI for Medical Applications.", The George Wa. <1%

14

Internet

Jaskaran Singh, Meenu Gupta, Rakesh Kumar. "Smart Technologies and Intelligen. <1%

15

Internet

listens.online <1%

16

Internet

squareboxseo.com <1%

17

Internet

www.coursehero.com <1%

18

Internet

www.ijisae.org <1%

19

Internet

taylor-amarel.com <1%

20

Internet

www.spectroanalysis.com <1%

21

Internet

arxiv.org <1%

22

Internet

files01.core.ac.uk <1%

23

Internet

iacis.org <1%

24

Internet

ir.juit.ac.in:8080 <1%

repository.up.ac.za <1%

Page 5 of 61 – Integrity Overview Submission ID trn:oid:::3618:135413136

25

Internet

26

Internet

thejas.com.pk <1%

27

Internet

toptenreviewed.com <1%

28

Internet

www.emjreviews.com <1%

29

Internet

www.jneurosci.org <1%

30

Publication

www.nature.com <1%

31

Internet

Ashish Kumar, Divya Singh. "Artificial Intelligence in Modern Healthcare System",. <1%

32

Internet

en.biltek.org <1%

33

Internet

repository.uel.ac.uk <1%

34

Publication

www.maxapress.com <1%

35

Publication

"Artificial Intelligence in Ophthalmology", Springer Science and Business Media L. <1%

Publication

36

J. Anitha Menon, Benita Veronica, Ben Milbourn, Mithra S. Gnana Sanga. "Sustain. <1%

37

Internet

Syed Nisar Hussain Bukhari, Kingsley A. Ogudo. "AD-ViTNet: Modeling Global Stru. <1%

38

Internet

daneshyari.com <1%

hal.science <1%

Page 6 of 61 – Integrity Overview Submission ID trn:oid:::3618:135413136

39

Internet

40

Internet

ijctet.org <1%

41

Internet

justagriculture.in <1%

42

Publication

www.saspublishers.com <1%

43

Publication

"Recent Trends in Image Processing and Pattern Recognition", Springer Science a. <1%

44

Publication

Bartlett, Benjamin A.. "Taking Responsibility for AI Systems in Healthcare.", The U. <1%

45

Publication

Lalit Mohan Goyal, Tanzila Saba, Amjad Rehman, Souad Larabi-Marie-Sainte. "Arti. <1%

46

Publication

Nilanjan Dey, Bitan Misra, Sayan Chakraborty. "Smart Medical Imaging for Diagn. <1%

47

Publication

David B. Olawade, Kusal Weerasinghe, Mathugamage Don Dasun Eranga Mathug. <1%

48

Publication

Francesco Pichi, Alessandro Invernizzi, William R. Tucker, Marion R. Munk. "Optic. <1%

49

Publication

Sukhpreet Kaur, Amanpreet Kaur, Manish Kumar. "Recent Advances in Computati. <1%

50

Publication

Thangaprakash Sengodan, Sanjay Misra, M Murugappan. "Advances in Electrical . <1%

Tuan D. Pham, Simon Holmes, Domniki Chatzopoulou, Paul Coulthard. "Artificial I. <1%