Abstract
The rapid advancements in deep learning algorithms and the availability of large, open-access databases of fundus and OCT (optical coherence tomography) images have contributed greatly to advancements in computer-assisted diagnostics and the localization of various disorders affecting the retina. This study offers a comprehensive examination of retinal diseases and various recent applications of deep learning strategies for categorising key retinal conditions, such as diabetic retinopathy, glaucoma, age-related macular degeneration, choroidal neovascularization, retinal detachment, media haze, myopia, and dry eyes. Open-access datasets continue to play a critical role in the advancement of digital health research and innovation within the field of ophthalmology. Thirty open-access databases containing fundus and OCT (optical coherence tomography) pictures, which are often utilised by researchers, were carefully examined in this work. A summary of these datasets was created, which includes the number of images, dataset size, and supplementary items in the dataset, as well as information on eye disease and country of origin. We also discussed challenges and limitations of novel deep learning models. Finally, in conclusion, we discussed some important insights and provided directions for future research opportunities.
Keywords
Introduction
Background
Eye is an intricate and delicate organ that has a vital impact on our everyday existence. Eye disorders such as glaucoma, diabetic retinopathy, age-related macular degeneration, choroidal neovascularization, retinal detachment, retinal vein Occulusion and retinoblastoma can have serious consequences. These include visual impairment and potential loss of sight, if not detected early.
The significant impact on public health is clear, vision impairment can result in a decrease in quality of life, a loss of autonomy, and an elevated potential for accidents. Furthermore, there are socio-economic repercussions like increased healthcare costs, decreased productivity, and the personal and societal burdens of providing support for those who become visually impaired [1]. Early identification and precise diagnosis of retinal diseases are critical for effective treatment and management, potentially preventing or delaying vision loss. As such, advancements in detection and classification techniques, including computer-aided detection systems and deep learning, are essential to tackle these public health challenges [2].
Over the last few decades, there have been significant advancements in the detection and classification of retinal diseases. This progress has shifted from traditional manual examination by ophthalmologists to the creation of automated computer-aided systems. The primary factors fueling this change comprise the rising frequency of retinal conditions, the demand for improved and precise screening methods, and the swift progress in digital imaging and deep learning technologies. Ophthalmologists have traditionally relied on manual retinal examinations as the established method, but this process was laborious, subjective, and susceptible to differences between observers.
The development of digital fundus imaging, including colour fundus photography and OCT, has transformed the detection and classification of retinal diseases using deep learning. Both human experts and automated computer algorithms can analyse the high-resolution digital images produced by these imaging techniques. The integration of machine learning and, more recently, deep learning techniques has transformed the detection and tracking of retinal diseases.
Deep learning models have shown tremendous potential in computer-aided diagnosis and localization of various retinal diseases. Convolutional neural networks (CNNs) in particular have demonstrated state-of-the-art performance in classifying retinal images into different disease categories. This is in part due to the availability of large, open-access databases of retinal images, such as OCTA, OCTA500, EyePacs, Messidor, DR1, DR2 and DRIVE, which have enabled the training and evaluation of robust deep learning frameworks.
Key contributions
In this review, we examined the most recent developments in deep learning-based techniques for the detection and classification of retinal diseases.
Specifically, we:
Provided a comprehensive review of the major retinal diseases and their clinical characteristics. Discussed the Challenges faced by ophthalmologist in diagnosing retinal diseases. Discussed the role of digital imaging techniques, like fundus photography and OCT, in diagnosing retinal disorders. Examined the application of deep learning algorithms, including convolutional neural networks and transfer learning, in automating the identification and classification of retinal disorders. Highlighted the development of thirty open-access databases of retinal images and their significance in advancing research and clinical implementation. Discussed the limitations of Deep learning models and Future prospects in the area of diagnosing retinal diseases.
This extensive review seeks to act as a guide for researchers, clinicians, and healthcare professionals interested in leveraging the power of deep learning and open-access retinal image databases to improve the detection and management of retinal disease
The retina serves as the principal mechanism for receiving, arranging, and transmitting visual stimuli to the brain via the optic nerve. This entire process facilitates the ability to see. Retinal disorders are a major contributor to visual impairment and blindness globally, impacting millions of individuals. These encompass a range of ailments, such as age-related macular degeneratio, diabetic retinopathy, retinal
Summary of various retinal diseases diagnosed by different Machine Learning and Deep Learning algorithms
Summary of various retinal diseases diagnosed by different Machine Learning and Deep Learning algorithms
vein occlusions, retinitis pigmentosa, glaucoma, choroidal neovascularization (CNV), Myopia, epiretinal membrane, retinoblastoma, and retinal detachment. Ageing populations, an increase in chronic conditions like diabetes, and lifestyle choices all have an impact on the prevalence of these diseases.
Early identification of these retinal conditions is essential for prompt and efficient treatment, as they have the tendency to result in visual impairment and blindness if not addressed timely. Table 1 summarises various retinal diseases diagnosed by different Machine Learning and Deep Learning algorithms.
Accurately diagnosing and classifying retinal diseases can be challenging due to several factors:
Many diseases of the retina share common symptoms, such as blurry vision and floaters, which poses a challenge in differentiating between them without thorough imaging and examination The progression of retinal diseases may exhibit subtle and gradual changes. Detecting early-stage conditions or monitoring advancement over time could necessitate the use of highly sensitive and specific diagnostic tools. The quality of images of the retina is vital, as it can impact the precision of diagnosis and categorization when affected by issues such as cataracts, small pupils, or movement in the eye. The presentation of retinal diseases can vary greatly among individual patients, and can be influenced by factors like age, ethnic background, and coexisting health issues. Skilled ophthalmologists may be in short supply, particularly in rural or underdeveloped areas. This has led to a dependency on automated systems that still need to undergo validation. Some diagnostic tools may not be widely accessible in all clinical environments, and advanced equipment such as Optical Coherence Tomography, while capable of producing detailed images, is expensive and not universally available. The creation of efficient computer-aided diagnosis systems demands extensive, varied, and accurately annotated datasets, which can be difficult. Imaging techniques like OCT provide detailed views of retinal structure, but they require specialized training and expertise for interpretation.
Imaging modalities are of utmost importance in the field of ophthalmology for the purpose of diagnosis and treating eye diseases. Each modality offers unique insights on the anatomy and function of the retina as well as related medical conditions. Some of the imaging techniques are listed below:
Fundus imaging
This method takes color images of the retina and is essential for recording the condition and progression of the retina. It is especially valuable for detecting bleeding, small swelling in blood vessels, discharge, and other abnormalities in the fundus that indicate diseases like diabetic retinopathy [6].
Optical coherence tomography (OCT)
It helps ophthalmologists diagnose and track conditions like macular holes, epiretinal membranes, retinal detachment, and age-related macular degeneration by providing comprehensive, cross-sectional scans of the retina [14].
OCT angiography
OCTA is a retinal imaging technique that offers detailed, three-dimensional visualisation of the retinal vascular structure with high resolution at the micron level. Unlike optical coherence tomography, it does not measure blood flow directly but provides functional details about retinal blood vessels and microvascular systems as an expansion of the OCT platform [32].
Fluorescein angiography
Fluorescein angiography includes injecting a fluorescent dye into the blood vessels, then taking photos as the dye travels through the retinal blood vessels. This procedure assists in detecting conditions such as retinal vein occlusions, diabetic retinopathy, and macular edema by revealing areas of non-perfusion, leakage, or new vessel growth [2].
Multimodal imaging
Multimodal imaging, involving the use of different imaging techniques in a single session, has the capacity to comprehensively assess the retina by offering both structural and functional data. Utilizing multiple imaging modes can aid in precisely identifying retinal conditions and customizing an effective treatment strategy for managing the disease [2, 13].
Open access retinal image databases
Publicly available retinal image databases have become invaluable resources for researchers and clinicians in the field of ophthalmology. These databases provide a vast collection of annotated retinal images that can be used for algorithm development, model training, and performance evaluation. Throughout the development process of a deep learning application, data plays an important role. In the modern era of digital technology, with widespread use of electronic health records, centralized storage of medical images, and prevalent digital solutions, there is a demand for repositories containing colour fundus and OCT images. To discover these repositories, various terms related to “fundus”, “OCT” “retina,” and “retinal images” in combination with terms like “dataset,” “database,” and “repositories” were entered into search engines such as Google Dataset Search, Kaggle, data.mendeley, Ieeedataport and Retinal image bank.
The Retina Image Bank was established in August 2012 and contains a substantial open-access assortment of over 25,000 distinct and downloadable retina images. Kaggle functions as a platform for both data science and AI, allowing users to provide their data for analysis by the community. The reliability of a dataset can often be assessed by examining its up votes or reviewing accompanying notebooks that have been shared alongside the dataset. Table 2 summarizes thirty open access databases of Fundus, OCT and OCTA images.
Application of Deep Learning for retinal disease diagnosis and classification
Literature review
About 15 million Indians are sight-impaired. A shocking 75% of these instances can be healed. India has a doctor-patient ratio of 1:10,000. Research shows that Glaucoma and Diabetic Retinopathy cause most blindness in India. Long-term diabetes leads to diabetic retinopathy, which is the main reason of
Summary of open access fundus, OCT and OCTA image datasets
Summary of open access fundus, OCT and OCTA image datasets
blindness in young individuals in both developed and developing nations. Glaucoma damages the visual nerve, causing blindness. Both illnesses have no early signs, making them hard to diagnose. If not treated quickly, these disorders might cause permanent vision loss [3].
Bali and Mansotra, 2023 [1] reviewed various studies that used deep learning models to detect conditions like diabetic retinopathy, age-related macular degeneration, cataracts, and glaucoma. They highlight the effectiveness of these models, emphasising the significance of large datasets for training purposes.
Sengar et al., 2023 [10] developed the EyeDeep-Net, a multi-layer neural network, to train and assess images for detecting different retinal diseases, including AMD, cataract, diabetic retinopathy, Media haze, and amblyopia. The CNN was utilised to identify important characteristics from the input colour fundus image dataset. Multi-class images of fundus were obtained from RFMiD dataset and subsequently used to inform predictive diagnostic judgements based on processed features.
Patil et al., 2023 [5] provided a system that uses deep learning techniques to automatically diagnose five categories of visual disorders using two types of stack ensemble models: ensembles of ensembles and faster ensembles. These categories include cataract, glaucoma, pathological myopia, DR, and eyes with no disease. To increase the robustness and generalization of the models, multiple ocular image datasets were pooled and assessed using metrics such as precision and recall. The ensemble of ensembles framework achieved an accuracy of 88% while the faster ensemble reached 87% accuracy.
Vyas et al., 2022 [9] provided a computer-aided diagnosis system that classifies TBUT frames and assesses the severity and existence of dry eye disease using a modified GoogleNet CNN architecture. Based on the TBUT videos, they classified DED as normal, moderate, or severe, and they have shown a high classification performance with an accuracy rate of 83%. Additionally, a robust 90% correlation was observed with the assessments of ophthalmologists.
Nagamani and Rayachoti, 2024 [7] developed a novel model based on deep learning techniques. The DL-Net aims to enhance the classification and segmentation of retinal diseases by efficiently recording inter-scale characteristics and integrating them using convolutional blocks. The DL-Net model is capable of automatically identifying normal, diabetic macular edoema (DME), choroidal neovascularization (CNV), age-related macular degeneration (AMD), and drusen pictures. The author presented the Modified ResNet-50 methodology and image processing techniques for the classification of multiple OCT images. The authors also proposed an effective detection technique for image segmentation utilising a Bi-LSTM-based deep recurrent convolutional neural network (DRCNN). The revised SqueezeNet model examines the volumetric segmentation of OCT pictures. The suggested framework achieved an accuracy of 99.76%.
Sagvekar et al., 2024 [4] proposed a method for HPLBO “Hybrid hunter-prey ladybug beetle optimization enabled deep learning for DR classification,” The input image was preprocessed using Kalman filter and ROI extraction to reduce noise. The Meyer wavelet, which is a type of wavelet transform, decomposes the image into subbands in order to facilitate subsequent processing. Lesions were segmented through K-Net trained with the Hunter-Prey Ladybug Beetle Optimizer. Veins and arteries were classified from the wavelet-transformed image with an automatic vein and artery classification network enhancement feature extraction phase. Diabetic retinopathy classification is performed using a Deep Maxout network trained with the proposed HPLBO method, achieving 93.6% accuracy.
Thanki, 2023 [17] proposed a smart computer-aided diagnosis system that utilizes machine learning and deep neural networks to analyze colour retinal fundus images in order to classify glaucomatous retinal images. The aim was to enhance the early identification and treatment of glaucoma, with the potential to reduce the risk of blindness. This proposal outline the objective of surpassing the current maximum accuracy of 97% achieved by the existing model for classifying retinal images.
Keenan et al., 2022 [82] Developed a deep learning model called DeepLensNet to automatically diagnose and classify age-related cataracts using anterior segment photographs. After being trained, the model can tell the difference between nuclear sclerosis and other eye conditions by looking at slit-lamp photos and retroillumination photos. The Age-Related Eye Disease Study dataset, which has 18,999 anterior segment images from 1137 eyes, was used in this study to train and test DeepLensNet’s ability to diagnose and classify the severity of cataracts.
Peng et al., 2019 [8] proposed a DeepSeeNet deep learning model to automatically categorize images according to the Simplified Severity Scale of the Age-related Eye Disease Study (AREDS) using bilateral colour fundus pictures. DeepSeeNet utilises a deep CNN with an Inception-v3 architecture for image classification. With 317 layers and over 21 million weights, the model underwent training using the Adam optimizer on powerful CPU and GPU hardware. It was consisted of three sub-networks: Drusen-Net for detecting drusen, Pigment-Net for identifying pigmentary abnormalities, and Late AMD-Net for determining late AMD presence or absence. Training of model involved 58,402 images, and testing included 900 images from follow-up data of 4,549 AREDS participants.
Al-Fahdawi et al., 2024 [83] developed a Fundus-DeepNet model to identify various eye diseases by combining feature representations from set of fundus images. Eight ocular diseases were classified using Discriminative Restricted Boltzmann Machine (DRBM) A series of comprehensive experiments were carried out on the Ophthalmic Image Analysis-Ocular Disease Intelligent Recognition dataset. The experiment results demonstrated F1-scores 88.56%, Kappa scores 88.92%, AUC 99.76%,and final scores of 92.41% in off-site test set,and Fi-scores 89.13%, Kappa scores 88.98%, Auc 99.86%, and Final scores of 92.66% in on -site test set.
Latif et al., 2022 [84] proposed a 2-phase network for localizing the optic disk and diagnosing glaucoma, called the ODGNet Model. The initial phase uses a visual saliency map combined with a shallow CNN to locate the optic disk in fundus images effectively. Pre-trained models based on transfer learning, like VGGNet, ResNet, and AlexNet integrated with saliency maps were employed in second phase for glaucoma diagnosis. These models were assessed using five open access datasets DRIONS-DB, HRF, DR-HAGIS, ORIGA and RIM-ONE. The experimental findings indicate that the ODGNet model, proposed for diagnosis of glaucoma on ORIGA Dataset, is highly predictive and achieves accuracy 95.75%, specificity 94.90%, Sensitivity 94.75%, and AUC 97.85%.
Mistry et al., 2023 [85] described the creation of a computer vision system utilising machine learning and image analysis to identify retinoblastoma, an eye cancer, by automatically analysing fundus photographs. The system integrated innovative approaches in deep learning and segmentation to achieve highly accurate and fast detection of retinoblastoma, enabling early diagnosis compared to conventional methods.
(Uppamma et al., 2023) [27] examined different methodologies employed in the identification and categorization of diabetic retinopathy through the analysis of fundus images. A hybrid machine learning method that combines various learning strategies and feature extraction techniques was proposed. Additionally, a Deep Transfer Learning method was employed, which utilises Inception-v3 to detect different layers of Convolutional Neural Networks. Furthermore, a Toboggan segmentation and multiagent procedure was utilised for damaged retinal vascular segmentation. Lastly, image quality was enhanced using Gaussian and modified Kirsch filters.
Shaikh et al., 2022 [86] presented a machine learning model developed to aid in the quicker and more efficient detection of eye diseases. The researchers trained a CNN model on a dataset of 372 images, categorized into normal eyes, glaucoma, and retinopathy. Images were preprocessed and converted to grayscale for enhanced feature extraction. The model achieved high accuracy rates: 99.01% for normal images, 99.99% for glaucoma, and 98.99% for retinopathy.
Malik et al., 2019 [87] proposed a framework that focuses on gathering structured diagnostic data in a globally recognised format to predict various eye diseases listed in ICD-10. It took into account all potential symptoms of both the front and back portion of eye. Medical experts entered a total of fifty-two diagnoses in the software. A two-stage diagnostic method was created to exclude patient information concerning refractive error prior to analysis by machine-learning algorithms. Four different classification techniques were evaluated, specifically the decision tree, random forest, naïve Bayes, and ANN. The random forest algorithm outperformed the artificial neural network, although it did have a slightly longer execution time compared to the decision tree algorithm.
Sait, 2023 [11] developed a DL-based EDC (Eye Disease classification) model to identify eye diseases like DR, glaucoma and cataracts early from complicated retinal pictures. In their approach, firstly, Image preprocessing was done using denoising autoencoders, then features were extracted using SSD approach. The selection of critical features was conducted using the Whale Optimisation Algorithm (WOA) with the Levy Flight and Wavelet mutation strategies. The ShuffleNet V2 model’s ideal parameters for categorising eye diseases were found using the Adam Optimizer. The suggested model exhibited substantial improvement, achieving an impressive accuracy rate of 99.1% and a sensitivity rate of 98.9% when evaluated on the ODIR dataset. Similarly, when tested on the EDC dataset, the model achieved an accuracy rate of 99.4% and a sensitivity rate of 98.7%.
Marouf et al., 2022 [12] Introduced an effective methodology for accurate prediction of five distinct ocular conditions: cataracts, primary congenital glaucoma, acute angle-closure glaucoma, exophthalmos, and ocular hypertension. Moreover, offered a comprehensive evaluation of nine machine learning techniques including Decision Tree, Random Forest, Naive Bayes, AdaBoost,Logistic Regression, kNN, Bagging, Boosting and SVM. In terms of performance, SVM emerged as the top performer with an impressive accuracy rate of 99.11% in a 10-fold cross-validation process. Logistic Regression also achieved noteworthy accuracy at 98.58% using an 80:20 split ratio.
Lee et al., 2020 [88] introduced a new technique for generating retinal images that can facilitate accurate diagnoses after retinal exams. This proposed approach ensures the capability to distinguish vessel areas in images, which are critical for examinations, even when working with low-resolution retinal images. This was accomplished by selecting the identical high-resolution images from a given dataset and isolating their high-resolution vessel segments. These segments were then combined with the low-resolution ocular image to create a distinct vessel image, aiding in making accurate diagnoses.
Winston et al., 2019 [89] proposed two modified SOM self organising map networks that were trained with the extracted feature vectors. The training process of the conventional SOM was enhanced by using sarm optimization technique.Their main objective was to speed up the learning time and improving the performance of the iris classification system.
Prasad et al., 2019 [3] proposed a deep neural network framework that utilizes Convolutional Neural Networks to automatically diagnose the occurrence of DR and glaucoma at their initial phases from fundus images sourced from Kaggle and Medimrg. The developed model can identify the diseases with 80% accuracy.
Zhu et al., 2024 [6] examined various deep learning architectures like CNN models, transfer learning strategies, and the utilization of datasets like APTOS 2019 and EyePACS for diabetic retinopathy classification. According to their findings the CNN models, particularly those based on the ResNet and VGGNet frameworks, are the most popular for Diabetic Retinopathy categorisation due to their depth, which can range from tens to hundreds of layers, resulting in excellent classification outcomes. Table 3 Shows the Summary of Deep Learning and Machine Learning Models for classification of multiple ocular diseases.
While the reviewed literature showcases the potential of AI-powered eye disease detection, several challenges and limitations remain:
Image Quality: Since the quality of images can differ due to various imaging conditions in the real world, there is a challenge in ensuring the model performs consistently across images of varying quality. Summary of DL and ML models for eye disease classification
The limitations of data availability and quality present challenges in developing accurate and reliable model
Less illuminated video for diagnosis, frequent blinking interfering with analysis, and the occurrence of blurred video affect data quality and diagnosis accuracy. Lack of public datasets makes model training and validation difficult.
Data Imbalance: Real-world medical datasets often have class imbalances, which can affect the model’s ability to detect less common conditions.
The presence of highly imbalanced datasets posed significant challenges during the pre-processing of the fundus images. There exists a need for effective image pre-processing techniques in order to address the issues of image noise and artefacts.
Processing data using deep learning methods can be quite challenging and time-consuming due to the extensive quantity of data that needs to be handled.
Using a single view image as input to the model for retinal illness identification can limit the model’s performance. Multi-view techniques have the potential to significantly improve model performance.
Data distribution: The effectiveness of the system can be impacted by the unequal distribution of data among different eye disease classification. Insufficient data on specific diseases may result in biased predictions, as the system may encounter difficulties in learning the characteristics of less prevalent diseases.
The limited number of publicly available datasets restricts the development and testing of robust models.
Difficulties in identifying the damaged lesions within the images, which affects the accuracy of the classification results.
Variation in how medical data is recorded and described can lead to inconsistencies that pose a challenge to general automated solutions and the generalizability of the model.
In this study, we explored various retinal disease then delved into various imaging modalities, including fundus photography, OCT, and fluorescein angiography, used to capture detailed information about the eye’s internal structures. We identified open-access datasets for each modality, facilitating further research in this domain. Next, we presented a comprehensive literature review on deep learning and machine learning models employed for eye disease classification. The reviewed literature emphasises the significant potential of deep learning techniques and machine learning techniques for detecting and classifying eye disorders. The results demonstrate that ShuffleNet V2 is highly effective in classifying DR, Glucoma, AMD with an accuracy of 99.1% using the ODIR dataset and 99.4% using the EDC dataset. Meanwhile, the DL-Net classification model achieves a highest accuracy of 99.76% in distinguishing DME, CNV, drusen, and AMD. Among machine learning models, SVM outperformed for classification of cataract, glaucoma and hypertension with an accuracy of 99.11%, using real-world patient data from 563 cases. However, we also acknowledged the challenges and limitations associated with deep learning models including handling high-quality diverse datasets effectively addressing imbalances in data performance across different imaging conditions as well as ensuring consistent proficiency throughout evolution should also be considered crucial to developing robust reliable computer-aided diagnostic systems truly complementing healthcare professionals.
Future work
Novel Deep learning algorithms can simultaneously detect several retinal disorders. This can assist clinicians identify patients with several retinal diseases. Despite the existence of previous research in this field, such as the simultaneous diagnosis of cataract, diabetic macular edema (DME), and diabetic retinopathy (DR), as well as the simultaneous diagnosis of age-related macular degeneration (AMD), diabetic retinopathy (DR), and glaucoma, the exploration of retinal diseases such as choroidal neovascularization (CNV), retinoblastoma, retinal detachment, and retinal vein occlusion remains relatively limited. In the future, it is anticipated that novel deep learning models will be employed for the diagnosis of such retinal disorders.