Introduction

With the incorporation of artificial intelligence (AI) into clinical practice, healthcare professionals get ready to face innovations. The U.S. Food and Drug Administration first approved an AI diagnostic device in 2018 that doesn’t require a specialist doctor to interpret the results. The software program, called IDx-DR, was able to detect a type of eye disease by looking at retinal photos. After this event, the field that AI started to cover in ophthalmology began to expand rapidly [1].

AI has become one of the most influential information technology revolutions of our time. The field of AI is essentially the use of machines to perform tasks that typically require human intelligence. It encompasses machine learning (ML), where machines can learn by experience and acquire skills without human intervention [deep learning (DL)]. AI enables a technical system of human-like behavior that consists of retrieving, interpreting, and learning from data before achieving a specific goal. Among the methods of AI, the intricacies of ML and DL come first [2, 3].

DL is a state-of-the-art ML technique in the last few years that has had a great impact all over the world. DL uses representation-learning methods with multiple levels of abstraction to process input data without the need for manual feature engineering and automatically recognizes complex structures in high-dimensional data by projecting them onto a lower-dimensional manifold [4]. Compared to traditional techniques, DL has been shown to achieve significantly higher accuracies in many areas, including natural language processing, computer vision [3–5], and voice recognition [6].

This study will review the current roles of AI applications in ophthalmology.

Core components in AI system

Most ophthalmologists and clinicians unrelated to computer science can become overwhelmed with AI work because of the complex technical terminology. Not all AI readers need to fully understand all technical architecture or mathematical formulas. However, it is important to evaluate the essential components of AI systems required for clinical translation. An AI system consists of 2 phases: (1) training and validation, and (2) testing. Training and validation in the AI system require 2 main components: (1) a training dataset of images and clinical data, and (2) the selection of a technical network [7, 8]. Without regard to splitting the training and validation datasets into their components, it’s more important to have enough robust and preferably several independent test datasets [9, 10].

ML and DL techniques

In AI studies, there are several technical methods to develop and train an algorithm, it largely depends on the question of the research and training dataset type. In general, the training dataset is based on only clinical data, images, or multimodal models. Many research groups have used conventional methods of ML and statistical methods to analyze clinical data [9]. These conventional methods seem adequate to build robust predictive algorithms to train the model. A newer method, DL, can be adopted for predictive and longitudinal clinical problems with temporal sequence [11]. On the other hand, for image-based training data [optical coherent tomography (OCT) or fundus photographs], the most preferred technique is DL by far because of its ingenerate diagnostic capability. Readers can encounter DL algorithms named convolutional neural networks (CNNs) made up of neuron-like quantitative layers that could have various numbers of layers. In addition, a technique known as transfer learning has been used to increase the performance of diagnosis. In this method, CNNs are usually pre-trained via the ImageNet database (a database containing millions of photos, such as animals, cars, etc.) before applying them to a particular dataset [9].

ML and DL in ophthalmology

The field of ophthalmology is well-placed to tap into the power of AI [12]. A subset of ML, DL consists of a network of interconnected algorithms called neural networks. These networks allow “filtered learning” from large datasets crossing multiple layers of algorithms to automatically extract features. The goal is to closely align the classification or regression outputs with the target output provided by the training set [13]. In ophthalmology, AI algorithms have been developed for the detection of diabetic retinopathy (DR) [14], glaucoma [15], age-related macular degeneration [16], and retinopathy of prematurity [17]. There have been many recent studies on the diagnosis, prognosis, and post-treatment predictions of the mentioned diseases.

Ahn et al. [18] obtained a total of 1,542 photos (786 normal controls, 467 advanced glaucoma, and 289 early glaucoma patients) by fundus photography. Datasets were used to construct simple logistic classification and convolutional neural networks using TensorFlow. The same datasets were used to fine tune pre-trained GoogleNet Inception v3 model. The convolutional neural network achieved accuracy and area under the receiver operating characteristic curve (AUROC) of 92.2% and 0.98 on the training data, 88.6% and 0.95 on the validation data, and 87.9% and 0.94 on the test data. Transfer-learned GoogleNet Inception v3 model achieved accuracy and AUROC of 99.7% and 0.99 on training data, 87.7% and 0.95 on validation data, and 84.5% and 0.93 on test data. The ML algorithm was able to accurately identify both advanced and early glaucoma using fundus photos alone. Compared to previously published models, their new model—which is trained using a convolutional neural network—was more effective at diagnosing early glaucoma.

In the field of ophthalmology, especially in diseases that cause blindness and diseases with a high incidence, AI-assisted medical scanning, and image-based diagnosis emerge in various sub-branches. AI combined with medicine can make the physician’s job more precise and effective.

AI for identifying DR

Ryu et al. [19] developed a fully automated classification algorithm to diagnose DR and determine referral status using OCT angiography (OCTA) images with a CNN model and validated its feasibility by comparing its performance to a traditional ML model. In their study, they presented an end-to-end deep CNN-based method for classifying DR severity automatically from OCTA images. Baseline facts for classifications were based on ultra-wide-field fluorescent angiography to improve the accuracy of data description. The proposed CNN classifier achieved 91–98% accuracy, 86–97% sensitivity, 94–99% specificity, and 0.919–0.976 area under the curve (AUC). In external validation, overall similar performances were also obtained. The results were similar regardless of the size and depth of the OCTA images; this indicates that DR can be satisfactorily classified even with images containing a narrow area of the macular region and a single image plate of the retina. CNN-based classification using OCTA is expected to create a new diagnostic workflow for DR detection and referral. This study demonstrates the idea that the high accuracy, sensitivity, and specificity rates shown in previous studies [20, 21] can be advanced by using new modalities and programs.

Early diagnosis of DR can provide rapid treatment that effectively prevents irreversible vision loss. For this reason, as a precautionary measure, the American Diabetes Association and the American Academy of Ophthalmology recommended that patients with diabetes who do not yet have retinopathy in their eyes should undergo a fundus examination once a year. For patients who have developed retinopathy with intraretinal bleeding or another complication (e.g., diabetic macular edema), follow-up visits are required at more frequent intervals, such as every 1 to 6 months, depending on the severity of the disease [22].

The results of the study by Wongchaisuwat et al. [23] suggest that “Patients will be evaluated early and treated quickly before the disease progresses to irreversible blindness. DL models have confirmed their clinical efficacy in screening for referrable DR and improving visual outcomes.”

Prediction of postoperative visual acuity after vitrectomy using DL-based AI

Macular holes (MHs) are vertical full-thickness defects of the neurosensorial retina in the fovea. Various factors have been shown to affect the visual acuity (VA) after the operation, namely the patient’s age, size of the MH, duration of the symptoms, pre-operative VA, and vascular density of the fovea [24]. In clinical practice, pre-operative VA is an indicative factor for prognosis, often considered the strongest predictor of postoperative VA [25]. However, only the pre-operative VA is not sufficient for the prediction of postoperative VA.

Obata et al. [26] constructed a model for the estimation of postoperative VA from preoperative OCT images using DL-based AI and compared the results with predictions made using a sequential multinomial logistic regression model based on pre-operative VA and pre-operative factors. The methods of DL-based AI indicated a little bit better accuracy but better correlation with actual post-operative VA than the postoperative VA estimating based on pre-operative VA alone. Moreover, DL-based AI methods indicated a little bit better accuracy but better correlation with actual post-operative VA than the estimation of postoperative VA using a multivariate linear regression model or an ordinal multinomial logistic regression.

The results of this study can be considered important because it not only supports the success of AI in the correct identification of diseases, but also inspires the prediction of post-treatment outcomes.

Advances in imaging pathologic myopia using AI

One of the causes of permanent visual impairment is pathological myopia. There has been a great increase in the prevalence of myopia in the world in recent years. This is a sign that the incidence of pathological myopia will increase in the future. Therefore, it has become important to be able to predict the progression of myopia more accurately in order to detect pathological changes early and to provide early intervention when necessary. For these reasons, imaging of myopic eyes becomes a priority.

Significant advances have been made recently in OCT techniques. Thus, it contributed to improving clinical management and phenotyping determinations and renewing the grading system for maculopathy in myopia and traction maculopathy in pathological myopia. The ability to perform wide-field fundus and OCT imaging has made significant progress in detecting posterior staphyloma. Non-invasive OCTA provided a more in-depth visualization of choroidal neovascularization in myopia [27].

Among the globally increasing myopic human population, DL algorithms are effective for the classification and screening of high-risk myopia and for detecting macular degeneration in myopic people. To this end, the blockchain platform can act as a reliable platform for the performance testing of AI models among future medical applications [28].

AI has made very useful contributions to the detection of pathological myopia and the identification of myopia-related complications. Imaging capability has advanced with the help of assistive AI analytics. Thus, it can provide important developments in monitoring the progress of pathological myopia and using it as a guide in treatment. In a review by Li et al. [29], the classification of pathological myopia, the detection and clinical evaluation of myopia-related complications by imaging, and how imaging improves management were tried to be presented in detail.

Thus, a broad update was provided on the development of AI algorithms for pathological myopia detection, classification, and clinical evaluation of myopia-related complications.

Automatic detection of glaucoma via fundus imaging and AI

One of the irreversible visual impairments worldwide is glaucoma. Cases of glaucoma are on the rise worldwide. Visual loss can be prevented if timely treatment and intervention can be made. Therefore, early diagnosis is extremely important. Diagnosis provided through the analysis of medical images using a computer is known as computer-aided diagnosis (CAD). In the past years, various techniques have been proposed to develop an effective CAD system for glaucoma detection. The CAD system can assist the ophthalmologist in batch screening of glaucoma. They outlined the lessons that can be drawn from existing approaches using segmentation, non-segmentation, and classification methods through studies conducted at many research centers around the world and the limitations that need to be addressed to develop a robust CAD for glaucoma detection [30].

In order to detect glaucoma promptly, it is important to evaluate the optic nerve head (ONH) and optic disc margins. Fundus oculi imaging is a non-invasive, low-cost but very important examination. Image review is a subjective assessment most of the time. It is therefore based on time-consuming and costly expert evaluations [31].

This is exactly why it is an important question whether AI can imitate glaucoma assessments made by experts by AI. That is, can AI automatically find the boundaries of the ONH and disc like an expert? Can it then use the segmented image to identify glaucoma with high accuracy?

AI helps anterior segment diseases of the eye

AI in medicine has the potential to help with the early detection, quantification, and monitoring of diseases affecting the anterior segment of the eye, including microbial keratitis, keratoconus (KCN), dry eye syndrome (DES), and Fuchs endothelial dystrophy (FED). To identify and categorize different forms of microbial keratitis as well as to quantify characteristics related to microbial keratitis, algorithms have been developed.

Numerous advancements have been made in the identification, division, and quantification of traits associated with FED and DES. Studies examining the application of AI in the screening, diagnosis, staging, or treatment of corneal conditions, with a focus on conditions like microbial keratitis, KCN, DES, and issues like meibomian gland dysfunction (MGD) and FED [35].

AI helps neuro-ophthalmology practice

The visual system extends from the eyes to the most posterior parts of the brain (occipital cortex). As a result, intracranial pathologies often lead to visual disturbances, prompting patients to consult an ophthalmologist. Neuro-ophthalmology is an integrative medical discipline that includes the study of pathologies along the visual pathway. The most common neuro-ophthalmic conditions are:

• (1)

  Those affecting the afferent system: conditions affecting the afferent visual system lead to various visual dysfunctions.

• (2)

  Affecting the efferent system: These are conditions that cause central ocular-motor disorders, ocular-motor cranial neuropathies, gaze instability, and pupillary disorders in addition to systemic dysfunctions affecting the neuromuscular junction or extraocular muscles [1].

Detection of eye movement disorders

The ocular deviation in infantile and acquired strabismus observed in children and adults may be associated with muscle restriction, lack of convergence or divergence, or refractive errors. Ocular misalignment can be detected clinically with the Hirschberg test and Krimsky test, among other methods, and the gold standard is the prism covering test (PCT). Eye movements are affected by cortical control, subcortical centers, premotor coordination of conjugated eye movements, ocular motor cranial nerves (especially cranial nerves III, IV, and VI), and extraocular muscles. This comprehensive system aims to create stable binocular single vision. Any damage to the ocular motor pathways can lead to ocular misalignment, concomitant gaze abnormalities, or abnormal involuntary oscillating movements of the eyes called nystagmus [64].

Additional applications for AI in conditions with neuro-ophthalmology

Although not directly seen in a neuro-ophthalmology clinic, various diseases such as neurodegenerative, neurodevelopmental, trauma, etc. can show ocular symptoms. In such cases, ML techniques have been considered in the parameters of view for neurodegenerative [78] [e.g., Parkinson’s disease (PD), and Alzheimer’s disease (AD)] and neuro-psychiatric diseases [79]. DL techniques related to pupillometry have been used in the investigation of psychiatric and neurodevelopmental disorders. Moreover, visually evoked potential (VEP) responses taken during intracranial surgery, particularly in the sellar region, can be automatically evaluated by neural network algorithms, directing real-time monitoring during surgical removal [80].

Detecting systemic diseases using AI from ocular images

Since the 20th century, research has used color fundus photography (CFP) as a standard screening technique for retinal diseases. AI-based on ophthalmological images has the potential to diagnose some systemic diseases. In this way, it has been shown that chronic kidney diseases (CKDs) can be detected in the early stages [81]. Large amounts of fundus photos give a possibility to assess predicting models, and the recent development of smartphone-based and handheld CFP cameras enhanced the utility of CFP [82]. However, the traditional use of information from CFP images through subjective judgments or simple statistics made by the human eye is limited compared to the use of AI. AI offers opportunities to supply key features from retinal images with ML and DL algorithms. This worthy information can be used as a chooser via heatmaps according to retinal specialists’ interests [83].

However, more research is needed to validate these models for real-world application. With the increasing number of relationships found between the eye and other body systems, it would be useful to explore the possibility of a universal scanning tool using ocular images.

Future directions

By integrating prior clinical knowledge with ocular image data in the ML models, it is possible to enhance the transparency and interpretability of the AI models. This approach can provide insights into how the models arrive at their predictions. It also helps establish a clearer understanding of the correlations between ocular characteristics and systemic diseases. Enhancing AI model transparency fosters trust and reliability in clinical settings, reducing reporting bias and correlation errors.

A major concern that should be raised is the lack of explainability in AI models, especially when they are applied to systemic diseases for which correlation with retinal features may not have been established through clinical trials. To address this issue, a combination of previous clinical knowledge and ocular image data can be leveraged using ML methodologies such as the deep regressor model [105].

Neuro-ophthalmologic studies benefit significantly from AI systems. These systems are valuable for screening and characterizing ONH structure and function and, to a lesser extent, certain eye movement disorders. These systems automate complex diagnostic procedures, ensuring timely and accurate results. They are particularly valuable in neuro-ophthalmological conditions, where ocular dysfunctions may signal underlying life-threatening or systemic diseases. Today, there’s a growing need for prospective, multicenter studies to assess the practical value of AI systems in both neuro-ophthalmic and non-neuro-ophthalmic settings, where experts may face readiness challenges. Furthermore, there’s a growing emphasis on DL-guided quality assessment of retinal images to minimize non-diagnostic datasets. This is an even more important consideration in neuro-ophthalmology, where data are scarce [106]. In the long term, ophthalmologists may leverage personalized treatments through “AI-assisted prognosis and disease monitoring”. This could enable more effective patient care. Simultaneously, non-ophthalmologists may benefit from AI-powered ocular exams, facilitating the automatic detection of neurological and systemic conditions.

It is imperative to continue developing methodologies that prioritize explainability in AI models. This ensures that insights derived from ocular imaging are interpretable and can be effectively integrated into clinical decision-making. Such transparency is crucial for the broader acceptance and adoption of AI-based ocular imaging in healthcare settings. A critical point to emphasize is the need for generalizability in predictive models for systemic diseases. While internal validation methods are valuable for proof-of-concept and initial AI implementation studies, they may not suffice for real-world deployment. To establish the robustness and viability of these models, external validation datasets comprising diverse and representative real-world data are essential. External validation datasets offer a more accurate representation of societal demographics, accounting for variations in race and ethnicity. These factors can significantly impact the performance of prediction models. For instance, retinal pigmentation varies among different populations, such as Asians and Europeans, with observable differences in the loss of retinal pigment cells and rod photoreceptors as individuals age [107]. This implies that age prediction models trained on a population with less retinal pigmentation may have lower performance when applied to a population with more retinal pigmentation.

To enhance the generalizability of AI models for systemic diseases, future studies should prioritize training datasets that exhibit greater biodiversity and closer resemblance to the intended validation datasets. This approach helps ensure that the models can perform effectively across diverse populations and minimizes biases associated with specific demographics. By incorporating a wider range of demographic characteristics and variations into the training data, the models are more likely to achieve superior generalizability. As a result, they can reliably produce predictions when applied to real-world scenarios.

While the potential of AI to enhance medical practice is promising, there is limited research on its influence on the relationship between healthcare providers and patients. From a psychosocial perspective, emerging technologies have the potential to reshape the dynamics between clinicians and patients in several ways. The integration of AI into healthcare settings is revolutionizing the delivery of care to patients. AI-generated information regarding diagnosis, treatment, and medications is increasingly guiding decisions across the entire healthcare spectrum. This includes choices related to treatment options, lifestyle changes, disclosure of health information to family members, and even the communication of unfortunate news.

As AI is poised to play a significant role in medical practice in the foreseeable future, it becomes crucial to comprehend its influence on shared decision-making and the overall patient-doctor relationship. In many healthcare systems worldwide, patients have the right to be informed about the tools, resources, and approaches employed to address their medical conditions. Patients will increasingly need to be aware that their diagnosis or medical prescriptions may originate from a machine, complementing the expertise and efforts of human doctors [108]. While the field of medicine inherently embraces innovation and technology, it’s noteworthy that certain healthcare professionals hold reservations about technology in patient care. Their concerns primarily stem from the perceived risk of dehumanizing patients and the fear that unfamiliar tools may be used against them in medical controversies [109].

The data clearly demonstrate that the process of technological innovation in the healthcare field is rarely seamless, even when dealing with technologies different from AI. While the benefits of AI and ML devices in improving the technical aspects of medicine, such as diagnostic accuracy, are evident, their influence on the patient-doctor relationship remains largely uncharted. It is presumed that the mere awareness of the presence of a “machine” in the healthcare process could influence the attitudes of both doctors and patients. From a psychological perspective, attitudes towards something can develop even before direct experience with it.

Conclusions

When considering the integration of AI into the diagnosis and treatment identification process, it becomes conceivable that the concept of “shared decision-making” should evolve to incorporate the contributions of artificial entities. AI will no longer be viewed as just another “app” on doctors’ devices; it will be seen as an active interlocutor capable of providing diagnoses, prognoses, and intervention recommendations to both doctors and patients. The implementation of AI in healthcare holds significant potential for positive effects, contingent upon the attitudes of doctors. For instance, if AI were to handle administrative and technical tasks in medicine, doctors would gain the opportunity to reclaim valuable time for consultation and empathetic listening to their patients, thus enhancing shared decision-making [108]. In this context, AI would serve as an active intermediary between healthcare providers and patients. Furthermore, patients must receive appropriate information regarding these important concepts. Many patients may not be familiar with the concept of over-diagnosis, which can make it challenging for them to assess the relative risks of unnecessary diagnosis and treatment compared to the risk of missing a cancer diagnosis. Additionally, patients may not always have clear preferences regarding these outcomes.

As AI advances through ML and DL, addressing factors like ethnicity, race, and gender remains a challenge which signifies that AI’s journey is far from complete [110]. As a result, it is critical to develop algorithms that prioritize patient well-being and dignity. Such a strategy is essential for improving AI’s implementation and integration in the healthcare industry. Hence, it becomes essential to define a reasonable default decision threshold for situations in which patients either choose not to express their preferences or do not have specific preferences [111].

At present, the scientific literature lacks sufficient research data to comprehensively address these questions. However, by drawing upon studies examining healthcare providers’ responses to technological innovation and insights from the field of medical psychology, it can be anticipated certain social-psychological phenomena that may emerge in future healthcare scenarios. Understanding these phenomena is crucial for preparing and managing any potential undesirable side effects. These developments will not only trigger social and psychological events but also inevitable ethical and legal consequences. As such, they demand evaluation from a comprehensive perspective that encompasses these aspects. By using these assessments, it is possible to guarantee that the required preparations are completed on time [112].