Artificial intelligence-based apps for screening and diagnosing diabetic retinopathy and common ocular disorders

Abstract

Artificial intelligence (AI), encompassing machine learning and deep learning, is being extensively used in medical sciences. It is slated to positively impact the diagnosis and prognostication of various diseases. Deep learning, a subset of AI, has been instrumental in diagnosing diabetic retinopathy (DR), diabetic macular edema, glaucoma, age-related macular degeneration, and numerous other ocular diseases. AI performs equally well in the early prediction of glaucoma and age-related macular degeneration. Integrating AI with telemedicine promises to improve healthcare delivery, although challenges persist in implementing AI algorithms, especially in developing countries. This review provides a comprehensive summary of AI, its applications in ophthalmology, particularly DR, the diverse algorithms utilized for different ocular conditions, and prospects for the future integration of AI in eye care.

Key Words: Age-related macular degeneration; Alzheimer's disease; Artificial intelligence; Automatic retinal image analysis; Chronic kidney disease; Convolutional neural networks; Diabetic retinopathy; Diabetic macular edema; International council of ophthalmology; Machine learning; Massive training artificial neural networks; Natural language processing; OCT angiography; Optical coherence tomography; Vision transformers

Core Tip: In this modern world, artificial intelligence-based algorithms have a definitive role in the early screening of different ocular diseases. Early screening helps in rapid referral and appropriate management to prevent sight-threatening complications. However, all said and done, it needs to be applied wisely to preserve the importance of human touch and empathy in clinical practice.

INTRODUCTION

The estimated number of patients suffering from diabetes mellitus (DM) is expected to reach 642 million by 2040[1,2]. The ophthalmic complications of DM, namely diabetic retinopathy (DR) and diabetic macular edema (DME), are leading causes of blindness and visual disability. The increasing need for DR screening and overcoming issues associated with current screening methodologies have led to integrating artificial intelligence (AI) with screening protocols. The concept of AI was first discussed at a conference at Dartmouth College in 1956. However, many healthcare workers may feel overwhelmed as the technology is complex. Therefore, it is imperative to have a basic knowledge of the technical architecture used in AI. AI involves two phases: Training and validation, as well as testing. The training dataset typically includes clinical images and clinical data[3]. Thus, AI is a process in which a machine aids in recognizing specific patterns. It encompasses developing algorithms to represent the population for performance analysis. The research question and type of training dataset are crucial in algorithm development. Recently, AI has been projected to play a critical role in diagnosing and planning the management of patients with DR and other ocular diseases (Video).

METHODOLOGY

For this review, the most cited articles on AI in ophthalmology were searched in standard databases like Cochrane, EMBASE, Web of Science, Google Scholar and PubMed. The Reference Citation Analysis Tool was also applied to search for the keywords, and high-impact factor journal publications written in English were selected.

BASIC CONCEPTS IN UNDERSTANDING OF ARTIFICIAL INTELLIGENCE

Types of AI can be categorized as follows:

General Artificial Intelligence: Capable of performing tasks similar to human intelligence.

Narrow Artificial Intelligence: Specialized in specific tasks or domains.

Artificial Superintelligence: Superior to human intelligence in all aspects[4]. AI is continuously evolving and utilizes various techniques and algorithms to diagnose diseases. Commonly used techniques include natural language processing (NLP) methods and machine learning (ML) techniques. ML is divided into supervised learning (where training data is provided to the computer) and unsupervised learning (where the device creates its own algorithm without explicit training data).

Convolutional Neural Network (CNN) algorithms include LeNet, AlexNet, GoogLeNet, and ResNet. NLP involves processing unstructured data such as medical notes and journals and extracting information from them. ML is a technique for analyzing structured data (e.g., genetic data and images), allowing computers to make diagnostic predictions through repeated learning from existing materials[5,6]. In ophthalmology, machine learning techniques are commonly utilized. These techniques involve computers repeatedly making predictions based on existing materials, particularly for automated analysis and classification of retinal images captured by optical coherence tomography (OCT) or fundus photographs. Abràmoff et al[7] highlighted the role of an AI-based diagnostic algorithm in detecting DR in 900 diabetic patients with no history of DR. Ryu et al[8] developed an AI algorithm using OCT angiography images. The model achieved a sensitivity of 86%-97% and a specificity of 94%-99% for DR classification. The classification is slated for effective diagnostic workflow and detection of DR.

Deep learning (DL) models employ two algorithms: CNN, which categorizes retina images based on quantitative layers, and Massive Training Artificial Neural Networks[4]. In 2008, the United States Food and Drug Administration (FDA) approved the use of AI to detect brain tumours using MRI imaging. As far as the use of AI in retinal pathologies is concerned, various studies have been done. In a study by Ting et al[9], an AI algorithm detected DR using multiple retinal images. AI has a broad range of applications; in addition to diagnosis, AI can provide organizational assistance to surgical procedures to reduce manual workload[10].

CURRENT STATUS OF ARTIFICIAL INTELLIGENCE IN OPHTHALMOLOGY

A summary of various trials establishing the role of AI in ophthalmology is presented in Table 1. Recent advancements in ophthalmology have seen the adoption of innovative techniques to improve privacy and robustness, such as federated learning (FL), and address data scarcity through self-supervised learning[11,12]. New AI models, including DeepSeeNet and PredictNet, have emerged as promising tools for enhancing the detection and management of conditions such as age-related macular degeneration (AMD) and glaucoma[13,14].

Table 1 Various trials on artificial intelligence in ophthalmology.

Ref.	Data included	Results
Kanagasingam et al[43], 2018	A total of 216 patients, out of which 193 agreed to undergo eye screening. 386 images were evaluated by an AI-based system and by an ophthalmologist	Specificity: 92%; Positive predictive value 12%; Detection of false positive cases attributed to poor image quality
Arenas-Cavalli et al[44], 2022	Examination of 1123 diabetic eyes, utilizing a well-designed protocol endorsed by the Chilean Ministry of Health Personnel and Retina Specialists	Sensitivity: 94.6%; Specificity: 74.3%
Peeters et al[10], 2023	Analysis of a dataset comprising 16,772 cases of DR and 16833 cases of DME unique patient visits	Specificity for DR: 94.24%; Sensitivity for DME: 90.91%; Sensitivity in patients aged over 65 years: 82.51%
Brown et al[45], 2018	Utilization of 100 test images in Retinopathy of Prematurity using Inception–V1 and U-Net CNN	Predicted sensitivity: 100%; Predicted specificity: 94%
Ting et al[9], 2019	Incorporation of ten external datasets from various countries (Japan, United States, Hong Kong, Mexico, and Australia) employing the Deep Learning Algorithm VGG-19	Sensitivity in referable DR: 90.5%; Specificity in referable DR: 91.6%
Morya et al[52], 2021; Morya et al[53], 2021	World's first Smartphone based AI Annotation tool for grading multiple retinal images in a shortest span – quantitative and qualitative analysis	DR; AMD; Glaucoma; Retinitis Pigmentosa; CSR etc.

AI: Artificial intelligence; AMD: Age-related macular degeneration; CSR: Central serous retinopathy; DR: Diabetic retinopathy; DME: Diabetic macular edema.

The application of AI in ophthalmology has witnessed significant expansion, with tools providing real-time feedback for screening DR and refined models like DeepSeeNet demonstrating efficacy in predicting AMD progression[15,16]. Moreover, smartphone-based AI applications and AI-driven robotic surgery systems have emerged as notable developments, improving accessibility and precision in eye care[17,18]. Recent studies have introduced more advanced CNNs like EfficientNet, DenseNet, and vision transformers that have shown superior performance in image classification and disease detection in ophthalmology[19].

FL allows the training of AI models without sharing patient data and has been increasingly used to enhance privacy and improve the robustness of AI models in ophthalmology, for example in retinal imaging in relation to myopic maculopathy[20]. New self-supervised learning algorithms have been developed to improve AI performance in scenarios with limited labelled data, which is crucial for rare ophthalmic conditions such as ocular surface squamous neoplasia[21].

AI in DR has been enhanced with real-time diagnostic systems that provide immediate feedback, improving screening efficiency in clinical settings[22]. AI models like DeepSeeNet have been further refined to better predict the progression of AMD and the need for interventions like anti-VEGF therapy[13]. AI algorithms such as U-Net and PredictNet have improved accuracy in detecting glaucomatous optic neuropathy and predicting disease progression[14]. Novel AI approaches are now being used to diagnose and treat various disorders of the anterior segment[23]. Similar improvements are being noted in the prevention and treatment of common eye diseases as myopia, glaucoma, diabetic retinopathy, and cataracts. Medical resource allocation, risk prediction and early warning systems, illness screening and monitoring, ophthalmic data management, health education, and patient management are all significantly enhanced by AI[24].

Combining AI with telemedicine has become more prevalent, particularly in remote areas. This integration has helped reduce patient wait times and facilitate access to specialized care[25]. Significant progress has been made in developing AI-driven smartphone applications for eye screening, particularly useful in resource-limited settings[17]. AI-driven surgical supervision systems have fostered improvements in precision and control, enhancing outcomes in ophthalmic surgeries[26].

VARIOUS AI APPS IN DR

AI algorithms have a significant role in screening DR using fundus photographs. Recently, AI algorithms have been demonstrated to have a higher sensitivity and specificity in detecting referable DR than human graders. Recent advances in AI technology and the data-rich nature of the ophthalmology speciality have played a vital role in using AI in DR screening. Numerous deep learning models have outperformed feature-based machine learning methods (Table 2). The algorithms cleared by the FDA for clinical use in DR screening include IDx-DR, EyeArt, and AEYE Diagnostic Screening (AEYE-DS).

Table 2 Various algorithms used in diabetic retinopathy as follows.

Study conducted	Algorithms used	Identified and diagnosed
A retrospective study by Liu et al[46], 2022, using traditional fundus images	EfficientNet-B5	DME
A retrospective study by Dai et al[47], 2021	ResNet and Mask R-CNN	DR grading
A retrospective study by Lee et al[48], 2021	OpthAI, AirDoc, Eyenuk, Retina AI Health, Retmarker	Referable DR detection
A prospective study by Heydon et al[49], 2020	EyeArt v2.1	Referable DR detection
A prospective study by Gulshan et al[50], 2019	Inception-v3	Referable DR detection
Akram et al[51] proposed an automated module	MESSIDOR database used	Proposed an automated module for the grading of diabetic maculopathy

DME: Diabetic macular edema; DR: Diabetic retinopathy.

FDA-approved algorithms

The first FDA-approved AI algorithm, IDx-DR, utilized AI with a retinal camera and employed deep learning methods. It is used to determine the severity of DR and automatically refers cases with more than mild DR (mtmDR) to ophthalmologists while rescreening negative cases annually, boasting high sensitivity and specificity. IDx-DR's decision-making capability without an ophthalmologist's intervention makes it accessible to any healthcare professional[10].

The second AI system the FDA approved was EyeArt by Eyenuk Inc. (Los Angeles, CA, United States). EyeArt has a sensitivity of 96% and a specificity of 88% for diagnosing moderate DR and 92% and 94% for more severe cases of VTDR (vision-threatening diabetic retinopathy). In addition to its accuracy, this algorithm is incredibly efficient[27].

In November 2022, the FDA granted clearance to another algorithm developed by AEYE Health to screen for mtmDR, the AEYE-DS. It is the first fully autonomous AI diagnostic screening system, now integrated into commercially available portable and table-top devices further enhancing the availability and accessibility of AI technology where it is required[28].

Research-based algorithms

SELENA+, a deep learning system, was developed in 2017 at the Singapore Eye Research Institute, Singapore. It was approved by the Singapore Health Service Authority to help in the national screening programme for Diabetic Retinopathy. It was used to detect vtDR (vision-threatening Diabetic Retinopathy) with considerable success[29].

MONA DR DME is a class 1 medical device recognized by the European Union Medical Device Directive that accurately predicts referable diabetic retinopathy, diabetic macular oedema and a combination of both. This software, requiring one fundus image per eye, efficiently reported three diabetic eye screening results with the image centred between the optic disc and macula. The program demonstrated high sensitivity and specificity in predicting both conditions[10] (Figure 1).

Figure 1 Artificial intelligence algorithm for diabetic retinopathy prediction. CNN: Convoluted neural network; DME: Diabetic macular edema; DR: Diabetic retinopathy; L/R: Left/Right; VGG: Visual geometry group.

EyeWisdom V1, a diagnostic software employing DL algorithms, is designed to aid in diagnosing DR. A study was conducted to evaluate the practicality and safety of EyeWisdom V1 in diagnosing DR compared to manual grading. The study encompassed 1089 fundus images. For all cases of diabetic retinopathy, the sensitivity and specificity of EyeWisdom V1 were 98.23% (95%CI: 96.93%–99.08%) and 74.45% (95%CI: 69.95%–78.60%), respectively[30].

A novel retinal fundus image reading system was developed to provide large, high-quality data to create new screening algorithms. This 5-step system rates image quality, presence of abnormality, findings with location information, diagnoses, and clinical significance[31].

VeriSee DR is a deep-learning image assessment tool for diabetic retinopathy screening. VeriSee DR (Acer Inc., Taiwan). The software has been used in Taiwan and Thailand. It utilizes CNN for training algorithms for DR screening. It was used to diagnose referable DR according to the International Clinical Diabetic Retinopathy Disease Severity Scale. Every image was gradable with a maximum sensitivity of 95.7%, and maximum specificity of 94.2% using various camera models[32].

AI algorithms have utilized the Kaggle public dataset developed to detect severe DR using colour fundus photography. A study wherein > 4000 photographs were selected for lesion annotation and the Inception V3 structure employed as the categorization algorithm, the findings demonstrated enhanced performance with 896-pixel input images, including increased sensitivity, specificity, and area under the receiver operating characteristic curve. The study concluded that this artificial intelligence-driven system can enhance the accessibility and efficacy of severe DR screening[33].

An innovative algorithm was developed for diabetic retinopathy grading training of junior ophthalmology residents and medical students. A total of 520 fundus photographs were included in this study. Photographs were divided into eight groups, with 65 images for each group[34]. Thus, this AI not only grades DR accurately but is also being used to train humans in doing so.

In 2016, another deep-learning system was introduced by Gulshan et al[35] It was developed at Google Inc., United States. It is a CE-marked medical device used to detect referable/non-referable DR.

Retmarker DR software, a deep learning system, was developed at the University of Coimbra, Portugal. It is a CE-marked medical device used for disease/no disease grading and detection of microaneurysm turnover. It also has the capability to compare current and previously captured images, enabling the assessment of disease progression[4].

Other AI techniques, such as morphological component analysis combine k-nearest neighbour algorithms and AI-recognised cues to achieve fast and accurate optic disc detection[36]. Morphological Component Analysis, has shown very high specificity, ranging from 97.45% to 97.53%, as reported in a study by Imani et al[37].

Considering more than minimal DR, which includes microaneurysms, hard exudates, retinal haemorrhages, cotton wool spots, and DME, AI implemented on portable devices and smartphone attachments in primary care settings demonstrate sufficient image quality for algorithmic assessment[38].

Recent advancements include a novel AI network approach combining segmentation and classification tasks. Morphological features are utilized for segmentation, identifying various OCT pathologies like DR, choroidal neovascularization, vitreomacular traction, drusen, macular hole, and serous retinopathy to enable early detection of subtle features and appropriate referral[38].

WHY IS THERE A NEED FOR AI IN DR?

Al-based diagnostic systems offer numerous advantages, including rapid processing speeds, reliable outcomes, and the capability to reduce interobserver variability. These technologies facilitate the screening of diabetes patients on a larger scale and in resource-limited environments by enabling the analysis of substantial volumes of images in a significantly shorter time frame compared to human specialists. The grading of retinal images in diabetic patients through fundus examination or colour fundus photography typically requires specialised retinal specialists, whose numbers are limited, especially in developing countries. Consequently, introducing AI systems has alleviated some of the challenges associated with manual grading, which is both expensive and labour-intensive (Figure 1).

CHALLENGES IN IMPLEMENTING AI

The scientific community reports that most algorithms have not been developed into software for primary or secondary care levels. Another factor is that while most algorithms focus on reporting DR, consideration of both DR and DME is essential and relevant. The real-world use of AI in DR can be challenging due to the technicalities of the technology, especially in the developing world[39]. Also, AI grading of DR in patients with significant cataracts and poor media clarity may be complex. The medicolegal issues associated with the inadvertent wrong diagnosis using AI can be a problematic situation for authorities and the medical fraternity. The application of FL can be a possible solution to address such problems.

There is an urgent need for widespread and accessible DR screening tools for individuals with diabetes mellitus. The same has been emphasized by the International Diabetes Federation, the International Council of Ophthalmology, the World Council of Optometry, and the International Agency for the Prevention of Blindness. The use of AI can play a significant role in making the screening and diagnosis of DR accessible[10].

Telemedicine alongside AI

Telemedicine, coupled with AI, can improve the health status of patients by exchanging medical information electronically between sites. In areas lacking health workers, telemedicine can facilitate the distribution of healthcare to remote regions, reducing waiting times for both patients and clinicians. Additionally, DL algorithms have been applied as visual impairment screening tools, aiding in referring patients to tertiary eye centres based on visual impairments[30]. Community-based teleophthalmology DR screening has been a useful tool in screening of referable DR in India. The MII RetCam (MII Ret Cam Inc, Coimbatore, India) and the Remidio fundus-on-phone (Remidio Innovative Solutions Private Limited, Bengaluru, India) are economical and portable, and several studies have shown their advantage in limited resource settings[39-41].

AI-driven robots

AI-driven robots are valuable and time-saving tools to enhance and improve patient health while making healthcare cost-effective. They are currently utilized in rehabilitation and physical therapy to enhance patients' motor and functional capabilities. Additionally, they find applications in telesurgery and geriatric care, such as in countries like Japan, where Eldocare robots are utilized. Furthermore, they can be employed in new drug development processes to detect adverse effects of medications[18]. However, they currently have limited applicability in the field of ophthalmology. More research and development is needed in the field of robotics driven by AI to devise novel technologies that deliver vision therapy, monitor medication compliance, and provide assistive support to persons having visual impairment, including amblyopia and visual field defects.

Healthcare AI companies

Several healthcare AI companies have developed various software for disease detection. For instance, Viz.ai introduced an Aneurysm detection software in 2018, Beth Israel Medical Centre developed an AI-powered microscope for detecting blood infections in 2017, and VirtuSense launched Movement Anomaly Detection software in 2012. Despite embodying significant advancements, these companies primarily work on a for-profit basis thereby raising concerns about data security and privacy[24]. Pooled multiethnic datasets being used to train AI are commercially marketed and thereby subject to the same security concerns[42].

Main challenges and concerns in implementing AI

The lack of regulations and guidelines for the application and creation of AI poses significant challenges. Further research is needed to develop AI ethically and ensure its appropriate use. Additionally, AI should be tailored and adapted based on different regions' demographics and healthcare systems[38].

Several AI algorithms used in various studies are based on traditional fundus imaging (TFI) or smartphone-based fundus images, especially in DR and other retinal disorders[43-51]. Morya et al[52,53] have analysed various studies in DR and AMD and commented that bias related to sampling, prejudice for example in the context of suspicious optic discs, measurement bias related to instrumentation, a wrong choice of software algorithm and concerns related to quality control of images utilized, all constitute the challenges being faced in the AI world today[52,53].

OTHER USES OF AI IN OPHTHALMOLOGY

In healthcare, AI has the potential to improve the outcomes of diseases. Data from electronic health records has shown that it reduces readmissions, particularly in patients with diabetes, heart disease, and cancer.

In chronic and non-reversible macular disease, AMD, AI can detect early stages automatically by detecting drusen or fluids. Some algorithms can predict the early onset of AMD and the need for anti-VEGF treatment[54]. Morya et al[52,53] developed the World's first smartphone-based AI annotation tool for grading multiple retinal images of 15 common disorders using a qualitative and quantitative deep–learning module. This tool had flexibility, portability, and high accuracy in image annotation. Thus, AI can help in identification and documentation of retinal lesions to guide treatment decisions.

DeepSeeNet, developed by Peng and his collaborators, is an algorithm that assesses the severity of AMD with good precision and accuracy[16]. Recently used algorithms include temporal deep learning models using OCT images, self-supervised non-parametric instance discrimination AI using TFI, and SLIVER-net models using 3d-OCT images for the detection of AMD, and choroidal neovascularization in AMD[55].

A CNN combined with a patch and image-based method has reported high accuracy in diagnosing Retinal Vein Occlusion. Similarly, significant advances in OCT imaging are notable in pathologic myopia, with a renewed grading system aiding in the detection of myopic maculopathy.

Many studies based on two-level sorting have proven effective in diagnosing Retinopathy of Prematurity (ROP). I-ROP DL has the potential to assess disease progression and regression of ROP. Additionally, the patient's response to treatment can also be analyzed[4].

Automatic artificial intelligence segmentation using OCT measures glaucoma patients' retinal nerve fibre and ganglion cell layer. Akter and colleagues developed the first algorithm in 2022[56]. Algorithms used include FusionNet in detecting glaucomatous optic neuropathy, U-Net and PredictNet in predicting glaucoma incidence and progression, ResNet 50 in detecting primary open angle glaucoma among patients with ocular hypertension, and 3d-ResNet-34 and 3d-ResNet-50 in suspected primary angle closure glaucoma with narrow iridocorneal angles and anterior synechiae[20].

In diagnosing dry eye, a DL approach is used to segment Meibomian glands digitally, achieving 95.6% meiboscore accuracy. The accuracy for segmentation of the eyelid and atrophy has been reported as 97.6% and 95.4%, respectively[55].

The Congenital Cataract Guardian System has been incorporated in congenital cataracts, exhibiting high sensitivity and specificity. This system, a smartphone app, comprises three function modules: Prediction, dispatching, and telehealth. Algorithms such as DeepLensNet, ResNet50, XGBoost Classifier, Global-local attention network, Random forest, and adaptive screening methods are utilized for diagnosing and quantitatively classifying age-related cataracts. AI is also employed in Phacoemulsification cataract surgery, utilizing Faster Region-based CNN[20].

Keratoconus, which can result in irreversible vision loss, can be detected using AI to identify subclinical keratoconus, which is crucial for improving visual prognosis. The algorithm for diagnosing subclinical keratoconus, BESTi, has demonstrated high sensitivity and specificity[55].

MobileNet is employed to differentiate between bacterial keratitis (BK) and fungal keratitis (FK)[57,58]. Random Forest is another algorithm used for differentiating BK, FK, and viral keratitis[59] (Figure 2).

Figure 2 Artificial intelligence algorithm for corneal disorders. DL: Deep learning.

Specular microscopy images have been instrumental in developing artificial intelligence algorithms to differentiate between normal and pathological corneas for corneal grafting[60,61].

U-Net and DenseNet networks are used. Optic atrophy and optic disc drusen were observed, with 96.4% sensitivity and 84% specificity in detecting papilloedema and normal ONH.

AI LIMITATIONS

One primary consideration is the lower quality of ocular fundus photographs in older populations, particularly in teleophthalmology, where image quality is inversely related to age[62]. Implementation of AI carries a risk of bias, emphasizing the need for continuous monitoring and regulatory guidelines[63]. AI algorithms have been associated with high false-positive rates, with data availability lacking for rare diseases such as eye tumours and common diseases like cataracts not routinely imaged[64]. While AI algorithms are prevalent in developed countries, there's a pressing need for their adoption in developing countries to address the scarcity of clinicians and high costs. More extensive training on image data is necessary to improve the sensitivity and specificity of AI[65]. Data sharing poses ethical and legal challenges. Federated Learning, where data resides in various institutions without centralization, can address these issues, ensuring patient privacy while enhancing AI robustness[66]. The current cost of developing AI algorithms is very high, and there are challenges like data breaches, data theft, and lack of transparency regarding how autonomous machine learning works, leading to reservations about this technology[67]. As ophthalmic images especially pertaining to the retina can be linked to several medical diseases (Table 3), AI acquisition of these images is a black box which can have unprecedented ramifications[68-71].

Table 3 Artificial intelligence use in non-ophthalmologic disease.

Diseases	Artificial intelligence algorithm
Autism spectrum disorder[68]	DL model based on OCT images and automatic retinal image analysis
Chronic kidney disease[69]	Elevated urine albumin/creatinine ratio associated with reduction in retinal and choroid vasculature density in OCT or OCT angiography studies
Iron deficiency anemia[70]	Lower retinal vessel density and reduced vessel light reflectance observed in OCT images
Intracranial hypertension[71]	Brain and Optic nerve study (BONSAI) AI, U-Net and DenseNet networks used and Papilledema, optic atrophy and optic disc drusen observed, with 96.4% sensitivity and 84% specificity in detecting papilloedema and normal ONH
Alzheimer’s disease[62]	DL model using retinal images showed 83.6% accuracy, 93.2% sensitivity, 82.0% specificity, and an AUROC of 0.93 for detecting Alzheimer's disease-dementia

AI: Artificial intelligence; AUROC: Area under receiver operating characteristic curve; DL: Deep learning; OCT: Optical coherence tomography; ONH: Optic nerve head.

CONCLUSION

AI has undergone an industrial revolution recently, particularly in processing input data through deep learning without human intervention. It has demonstrated high accuracy in various fields, including medical image analysis. AI has been employed in analyzing ocular images captured by OCT, OCT Angiography, and fundus photographs. AI shows promise for long-term integration with telemedicine. The use of AI is imperative to improve the screening efficiency of DR in addition to other diseases like AMD, ROP, and Glaucoma. Additionally, AI algorithms aid in understanding the relationships between systemic diseases and ocular conditions. To achieve a broader acceptance, methodologies must be developed to integrate AI into decision-making processes effectively. Incorporating AI into healthcare is imperative for improving patient outcomes through early diagnosis and intervention, leading to cost-effective, streamlined processes. However, implementing AI software can present medical and ethical challenges. The use of AI may be associated with data privacy issues and have to be addressed promptly. AI should be viewed as a tool to assist practitioners.