FormalPara Key Summary Points

Enhancement of diagnostic capabilities: The integration of artificial intelligence (AI), particularly deep learning (DL) frameworks, into ophthalmology significantly enhances diagnostic accuracy by effectively detecting and quantifying ophthalmic features in imaging data. This advancement offers ophthalmologists a powerful tool for improving patient care and treatment outcomes.

Importance of methodological transparency: Emphasizing the need for high-quality, diverse datasets and transparent reporting of methodologies, the paper highlights these as essential for ensuring the reproducibility and reliability of AI studies in ophthalmology. Transparent practices help mitigate the risk of biases and enhance the credibility of research findings.

Ethical and clinical considerations: The paper addresses critical ethical issues and potential biases in AI models, advocating for continuous monitoring and improvements to AI systems in clinical settings. It also stresses the importance of striking a balance between minimizing false negatives and reducing false positives, which is vital for optimizing patient outcomes without causing unnecessary interventions.

Introduction

The integration of AI into the healthcare industry represents a transformative shift, improving the accuracy of diagnosis, raising the standard of patient care, and streamlining operations [1]. The application of AI in healthcare is diverse, encompassing medical imaging, pathology, large-scale analysis of patient data, and provision of telemedicine services [2].

In medical radiology, for instance, AI algorithms have demonstrated exceptional capabilities in analyzing medical images, locating lesions, and predicting disease progression [3]. Moreover, in the sphere of diagnostic histopathology, AI is refining image analysis, streamlining tissue segmentation, and aiding in predictive analytics. In healthcare education, AI is redefining approaches to training and simulation. By offering personalized and adaptive learning experiences, AI meets the unique needs of each medical professional, promising to boost educational outcomes and equip practitioners with the skills required to adeptly manage the intricacies of modern medicine [3]. The impact of AI on telehealth is particularly significant, facilitating remote patient monitoring, tailored medical treatments, and patient-centric care.

The integration of AI in ophthalmology provides benefits by offering effective solutions for diagnostics, treatment planning, and patient monitoring [4]. AI has the capability to screen for and diagnose a variety of eye conditions such as diabetic retinopathy (DR) [5,6,7], glaucoma [8, 9], age-related macular degeneration (AMD) [10,11,12], retinopathy of prematurity (ROP) [13], cataracts [14], and other anterior segment diseases [15,16,17], and the study showed that the sensitivity was 97.5% for detecting DR [5] and 100% for detecting AMD [18]. AI has significantly automated the screening and prioritization of conditions, including DR, ROP, and glaucoma. This automation has heralded significant improvements in the early detection and intervention, thereby playing a crucial role in preventing vision impairment. The application of AI extends into treatment planning, leveraging imaging technologies such as optical coherence tomography (OCT) and OCT angiography (OCTA) to detect minute structural changes in ocular tissues. AI systems are also used for forecasting surgical outcomes and in monitoring for complications that may arise from systemic medications [19]. The integration of AI with imaging techniques has contributed to patient monitoring in ophthalmology. Deep convolutional neural networks (CNN)-based AI systems are utilized for recognizing and categorizing pathologic myopia from fundus images [20]. In addition, AI plays a significant role in meibography analysis, which is essential for diagnosing and tracking meibomian gland dysfunction [21]. The workflow of AI in ophthalmology is shown in Fig. 1.

Fig. 1
figure 1

Workflow of AI in ophthalmology typically involves data collection and preprocessing, model development, training, validation, testing, and implementation. It is designed to leverage machine learning and deep learning techniques for diagnosing, treating, and managing eye diseases

Full size image

The rapid advancements in AI, particularly in image-based diagnostics such as DR and ROP, underscore the importance of incorporating AI training into ophthalmology education [22]. Studies show that AI systems can match or exceed the diagnostic abilities of experienced clinicians [23]. A solid understanding of AI will not only aid in the accurate interpretation of diagnostic results, but also play a crucial role in their practical application in clinical settings [22]. To effectively integrate AI education into ophthalmology training, a comprehensive AI curriculum is recommended. This curriculum should include an introduction to key mathematical and statistical concepts, foundational principles of AI and machine learning (ML), methods for critically evaluating AI research critically, and an exploration of its clinical applications [24]. The training framework should also incorporate courses in informatics, statistics, and computer science to prepare future ophthalmologists for an AI-enhanced healthcare landscape. Additionally, the curriculum must emphasize the humanistic aspects of medical practice, such as professionalism, communication, empathy, compassion, and respect, to prevent the depersonalization that may result from the use of AI technologies [25].

The integration of AI into telemedicine, notably its cost-effectiveness and utility in simultaneously screening for multiple eye diseases simultaneously in different settings, underscores its potential to improve patient care [26]. Moreover, AI predictive analytics can forecast disease progression, paving the way for more tailored prevention strategies and treatment regimens. Furthermore, AI's contribution to patient triage by evaluating the severity and urgency of conditions ensures that patients receive the appropriate level of care [23]. The explanation of professional terms is presented in Table 1.

Table 1 Explanation of professional terms
Full size table

Ethics Compliance

This article is based on previously conducted studies and does not contain any new studies with human participants or animals performed by any of the authors.

Machine Learning in Ophthalmology

Supervised, Unsupervised, and Reinforcement Learning

The application of ML techniques such as supervised, unsupervised, and reinforcement learning has demonstrated significant potential in advancing eye care and treatment methods. Supervised learning, characterized by the use of labeled datasets to train algorithms, has been used in disease screening and classification in the field [27]. In this method, the algorithm learns through a process of comparison between its generated outputs and the correct outcomes, allowing it to identify and correct errors.

Unsupervised learning, on the other hand, operates on unlabeled data to discover inherent patterns or structures without any prior knowledge of potential outcomes. This technique is particularly valuable for managing voluminous datasets in ophthalmology, such as those derived from corneal topography [28]. Reinforcement learning focuses on training algorithms to make decisions through a trial-and-error approach, rewarding or penalizing actions to foster the adoption of optimal behaviors. Its application in ophthalmology spans several areas, including the management of chronic eye conditions, the refinement of ophthalmic surgical techniques, and the improvement of image segmentation processes [29].

Collectively, these ML methods are spearheading innovations in ophthalmology, offering new avenues to diagnose disease, optimize treatment and improve patient care. Their integration into ophthalmic practices heralds a future where technology and healthcare converge to deliver superior outcomes for patients worldwide.

Training and Validation Datasets

The foundation of advanced ML models significantly relies on the meticulous selection and utilization of training and validation datasets. These applications draw upon a wide array of ophthalmic imaging techniques, including color fundus photography, OCT images, ultra-widefield imaging, and even retinal images captured by smartphones. Central to these endeavors are expansive datasets, such as the Intelligent Research in Sight Registry and the Smart Eye Database, which furnish comprehensive patient data indispensable for the cultivation of sophisticated AI algorithms [30].

The quality and diversity of these datasets are crucial for the development of efficacious ML models. High-quality images are crucial for AI, and initiatives such as Deep Fundus have been instrumental in improving the quality of datasets by automating the assessment of image quality. This not only helps to weed out poor-quality images but also aids technicians in capturing images of superior quality [31]. Diversity in datasets is essential for the algorithms to generalize well to new, unseen data, and to be representative of the population being studied.

Interpretation of AI studies in ophthalmology focusing on performance metrics such as sensitivity, specificity, false positives, and false negatives of developing ML models (Table 2).

Table 2 Definition of metrics parameters
Full size table

The rationale for setting an operating threshold during the training phase must be described, and sensitivity and specificity must be demonstrated on independent datasets using a consistent operating threshold [32]. The development process typically involves the separation of the dataset into a development subset and a test subset [33]. It is imperative that this partitioning is executed in a manner that preserves the overall representativeness of the datasets and adequately addresses any potential class imbalances, leveraged to refine the model’s architecture and optimize its training phase. A comprehensive consideration of these metrics can provide a thorough understanding of the performance of AI system on specific tasks, aid developers in optimizing models, and offer crucial guidance to end-users in selecting and utilizing AI systems. Careful evaluation of these metrics can better ensure the effectiveness of AI systems in practical applications, particularly in the field of ophthalmology, where accuracy is highly valued.

Deep Learning Techniques

Overview of Neural Network Architecture

These pioneering studies have laid the foundation for the development of DL and the field of computer vision. Krizhevsky’s “AlexNet” introduced innovative features, demonstrating the profound capabilities of deep networks in handling computer vision tasks [34]. The study by Simonyan and Zisserman, employing small convolutional filters (especially 3 × 3) and deeper network structures, demonstrated the advantages of emulating the receptive fields of larger filters by multiple layers of small filters while reducing the number of parameters [35]. The residual networks introduced by He et al., solved the vanishing gradient problem through skip connections, making it feasible to train extremely deep networks [36]. Szegedy et al.’s introduced inception modules, embedding networks within networks, and executing multiple convolutions simultaneously with filters of various sizes, as well as reducing dimensions through 1 × 1 convolutions, thereby enhancing computational efficiency [37]. Vaswani et al.’s “Transformer” revolutionized natural language processing methods through the self-attention mechanism, enabling parallel processing of data sequences, surpassing the capabilities of recurrent neural networks in managing long sequences, and leading to the creation of advanced models such as BERT and GPT, redefining benchmarks in natural language processing applications [38].

The collection of these scholarly articles delineates a progressive timeline of innovation in the realm of neural network architecture, with each subsequent study drawing upon the wisdom of its predecessors to extend the frontiers of technological feasibility. Commencing with AlexNet’s groundbreaking introduction of DL for vision tasks and culminating in the Transformer model’s transformative influence on sequence modeling, these architectures have transcended merely achieving unparalleled results.

CNNs in Ophthalmic Imaging

CNNs have been applied to ophthalmic imaging with remarkable success. The development and validation of AI-powered retinal diseases prioritization tools illustrate this trend. These tools are assessed through rigorous metrics including accuracy, sensitivity, specificity, and area under the curve (AUC), underscoring the precision and reliability of AI in ophthalmology [39].

Moreover, the synergy between CNNs and diverse ophthalmic imaging techniques has catalyzed advancements in the field. The integration of CNNs with modalities such as fundus autofluorescence (FAF), OCT, and color fundus photographs has significantly improved the diagnostic accuracy for conditions such as glaucoma, corneal ulcers, and macular edema has been substantially enhanced [40]. In the realm of OCTA, advancements in DL have paved the way for significant improvements in image clarity and quality. A pioneering algorithm, leveraging the U-net architecture for DL, has been introduced to reduce noise in enface OCTA images. This technique called Intelligent Denoise, markedly improves the image quality by diminishing background noise and bolstering the continuity of the vascular structures [41]. In parallel, the utility of CNNs is extending into the domain of retinal diagnostics through smartphone-based imaging systems. These systems, such as iExaminer et al., have been rigorously analyzed for their effectiveness in detecting DR [42]. This assessment highlights the potential of mobile-based platforms to expand access to retinal screening, ushering in a new era of preventive ophthalmology.

Applications in Ophthalmology

The introduction of the Retinal AI Diagnosis System marks a major step forward [43]. This DL algorithm is capable of simultaneously detecting up to ten distinct retinal diseases from fundus photographs, demonstrating the potential of AI to enhance the efficiency and accuracy of retinal disease diagnosis.

Recent advancements in AI and ML have significantly enhanced the management of inherited retinal diseases (IRDs). A diverse array of sophisticated techniques, such as CNNs and others, have been employed to undertake a variety of tasks [44]. These tasks encompass segmentation, prediction, detection, classification, and regression within the realm of retinal image analysis. This multidisciplinary approach has not only improved the precision in diagnosing and understanding of IRDs, but has also paved the way for innovative treatment strategies.

The deployment of AL for the automated disease detection and classification has undergone remarkable progress. The incorporation of DL models, including CNNs and Vision Transformers, has played a crucial role in refining the precision and streamlining the process of diagnosing retinal conditions. A significant breakthrough has been achieved through the creation of an automated mechanism designed to detect and classify 28 different ocular diseases. The strategic use of an ensemble method that fuses diverse CNN architectures. It has achieved a commendable area under the receiver operator characteristic (AUROC) score of 0.96 for disease screening purposes, alongside an average AUROC of 0.93 for accurate disease classification [45]. Furthermore, another pioneering study introduced a DL framework based on the VGG-19 network architecture. By employing transfer learning techniques and incorporating pre-trained weights from the extensive ImageNet database, this model outperformed previous state-of-the-art methods. It boasts an impressive accuracy rate of 99.17% in classifying retinal diseases [46]. Exploring of the diagnostic capabilities of FAF imaging in identifying IRDs has yielded promising results. Researchers have adeptly applied a CNN model to categorize FAF images. This classification covers a range of conditions, including normal retinal health, Stargardt’s disease, Best’s disease, and retinitis pigmentosa. Impressively, the model demonstrated a high level of precision in its classifications, achieving an overall accuracy rate of 95% [47].

Recent advances in the field of ophthalmology have seen a significant shift towards refining DL models for the precise classification of retinal OCT images. Researchers have worked diligently to enhance the capabilities of conventional CNN architectures. By integrating sophisticated methods such as spatially separable convolutions (SSC), these models not only achieve remarkable accuracy, but also significantly reduce computational demands [48]. This innovative approach ensures that DL models can be used efficiently in clinical settings, where computational resources may be limited. The incorporation of SSC and similar techniques marks a pivotal step towards making advanced diagnostic tools more accessible and practical for everyday clinical use.

Examples of Successful DL Applications in Ophthalmology

Burlina et al. developed a DL algorithm for the automated detection of AMD and achieved a 92% accuracy rate in identifying individuals with moderate and advanced stages of AMD [49]. Ting et al. analyzed a dataset of 125,189 fundus photographs. The DL algorithms demonstrated a sensitivity of 96.4% and a specificity of 87.2%, further underlining their diagnostic accuracy [50]. Tong et al. developed a neural network specifically designed to detect retinopathy of prematurity (ROP), utilizing 36,000 fundus images for training. The system achieved a diagnostic accuracy rate of 0.90, demonstrating diagnostic capabilities comparable to or surpassing those of retinal subspecialists [51]. In the field of keratoconus detection from corneal topographies, Lavric et al. developed KeratoDetect, a system that achieved an accuracy of 99.33% [52]. Zhang et al. developed an AI system based on DL and transfer learning for efficient and accurate segmentation of meibomian gland images and assessment of their density, achieving an accuracy of 92% in meibomian gland segmentation, with the sensitivity and specificity of meibomian gland density of 88% and 81%, respectively [21].

Challenges and Considerations in Model Development

Despite the notable achievements in incorporating DL into ophthalmology, several challenges and considerations remain paramount to ensure its successful and sustainable integration. These include the need to increase the diversity and volume of data, standardize reporting protocols, enhancing the interpretability of DL models, ensure seamless clinical implementation, address ethical concerns, and foster collaboration among various stakeholders.

To begin with, the effectiveness of DL algorithms largely depends on the availability of a wide and diverse dataset. It is imperative to compile datasets that cover a wide range of patient demographics, disease stages, and imaging modalities. Such comprehensive datasets are vital for the development of algorithms that are both robust and able to generalize across different clinical scenarios. Moreover, the establishment of standardized reporting protocols and regulatory guidelines is crucial to ensure the safety, efficacy, and ethical utilization of DL applications in clinical settings [39]. This standardization will facilitate the consistent evaluation and comparison of DL tools, ensuring their reliable application in patient care. The integration of DL systems into existing clinical workflows requires careful consideration. This entails ensuring that these systems are user-friendly, compatible with the current healthcare IT infrastructure, and conducive to enhancing, rather than impeding, clinical efficiency [53].

Ethical Considerations

Patient Privacy and Data Security

The key ethical issues related to patient privacy and data protection touched upon by AI, are crucial to maintaining the integrity of the healthcare system and ensuring compliance with regulatory standards (Fig. 2).

Fig. 2
figure 2

The Ethical Consideration figure indicates that AI applications in ophthalmology must safeguard patient privacy and data security while effectively addressing bias and fairness

Full size image

In healthcare, ethical guidelines, exemplified by foundational texts such as the Belmont Report, emphasize the criticality of honoring patient autonomy and ensuring the confidentiality of their information [54]. This ethical mandate is even more important in ophthalmology. The management of sensitive information, encompassing retinal images and comprehensive personal health records, demands rigorous safeguards against unauthorized disclosure and exploitation. Ensuring the privacy and security of patient data is not only ethical, but also fortifies trust in the healthcare ecosystem.

Both practitioners and researchers are obliged to comply with strict data protection laws, including the Health Insurance Portability and Accountability Act (HIPAA) in the United States, the General Data Protection Regulation (GDPR) in Europe [55] and newly ratified Artificial Intelligence Act in the EU. These regulatory frameworks necessitate the secure management of patient data, the fulfillment of consent provisions, and the establishment of safeguards to maintain confidentiality.

The protection of patient data is a cornerstone of ethical practice, especially as the frontiers of AI in healthcare expand. Adhering to data protection laws, surmounting the hurdles associated with data sharing, guaranteeing transparency, securing informed consent, and pioneering novel approaches to privacy protection are integral to maintaining an ethical stance in the use of AI technologies in the medical sector. As this domain evolves, it is crucial to foster ongoing discussions among all stakeholders-healthcare professionals, patients, scientists, and policymakers—is crucial. This collaborative effort is vital to ethically navigate the complexities of AI applications in ophthalmology, and ensure that the advance in technology enhances patient care without compromising patient privacy or the trust placed in the medical community.

Bias and Fairness

Significant progress has been made in the field of AI to amass datasets that mirror the vast diversity of the world’s population. This initiative is pivotal in diminishing bias and enhancing the universal applicability of AI algorithms. Moreover, the field of ophthalmology has seen the implementation of randomized controlled trials (RCTs) utilizing AI. These studies have been meticulously assessed for their compliance with the CONSORT AI guidelines [56]. Such guidelines are instrumental in augmenting the transparency and reproducibility of AI research, ensuring that findings are both reliable and replicable across various contexts.

The process of identifying and reducing bias within AI algorithms encompasses a variety of strategies, broadly categorized as pre-processing, in-processing, and post-processing measures [57]. Preprocessing strategies aim to create equitable datasets to counteract initial biases. In contrast, in-processing techniques seek to refine the learning algorithms themselves to prevent discriminatory bias. Post-processing strategies, on the other hand, are designed to correct biases in the AI system’s output. One noteworthy method has been specifically developed to mitigate bias in surgical AI systems and has shown potential applicability in the field of ophthalmology [58].

The pursuit of equity in AI-enabled healthcare services necessitates a nuanced understanding and extension of the ethical principle of fairness within an AI framework. This concept of fairness extends beyond merely eliminating bias and ensuring non-discrimination; it encompasses both distributive justice and socio-relational justice. Such an approach guarantees that all individuals have equal access to opportunities, and have the right to seek explanations for AI-influenced healthcare decisions. In order to realize equitable healthcare machine learning algorithms, it may be imperative to develop and deploy compensatory mechanisms that specifically support the most vulnerable populations. Moreover, these interventions must be adaptable, taking into account the diverse socio-political and economic landscapes in different regions.

Challenges and Future Directions

The integration of AI in ophthalmology represents a transformative path for the future, but also presents challenges that still need to be addressed (Fig. 3).

Fig. 3
figure 3

Overview of challenges and future trends of AI in ophthalmology

Full size image

A significant challenge to the integration of AI in healthcare is its opaque nature, often referred to as the “black box” phenomenon [59]. This opacity makes it difficult for healthcare professionals and patients alike to grasp the underlying processes that lead to AI-generated conclusions. This issue is not only about transparency, but also about trust and usability in clinical settings.

Furthermore, the emergence of AI in healthcare has led to a complex dilemma known as the “responsibility gap”. This predicament revolves around the ambiguity surrounding accountability when AI systems malfunction or produce errors [54]. The question of whether responsibility lies with the developers, the clinicians who implement these systems, or the manufacturers remains unresolved. This uncertainty creates a legal and ethical dilemma that has yet to be adequately addressed in the healthcare sector.

Moreover, the integration of AI into ophthalmology brings to the fore several key considerations regarding the economic and professional landscape of healthcare. Among these considerations are the economic implications, which encompass cost–benefit analyses and the intricacies of reimbursement policies. These financial aspects are crucial, yet they lack clear guidelines and definitions within the realm of AI applications in ophthalmology. Additionally, this technological shift is prompting a re-evaluation of the role of healthcare providers and necessitates a deeper understanding and acceptance of AI by patients. These changes signify a transformative period in healthcare, heralding both advancements and challenges in the integration of AI into clinical practice.

Additionally, the presence of algorithmic biases and insufficient domain adaptation limits the applicability of AI systems across various demographics and environmental settings. Consequently, these limitations may lead to AI models that exhibit robust performance during their development phase, but falter in providing precise outcomes when applied in heterogeneous clinical contexts. This discrepancy underscores the need for a more adaptable and inclusive approach to the development and evaluation of AI-based healthcare solutions.

To ensure the successful integration of DL systems into ophthalmology, it is imperative to adopt a holistic approach that meticulously balances technical excellence with ethical, legal, and societal considerations. Prior to their full-scale development, it is crucial to address and rectify any issues stemming from faulty instructions, dubious methodologies, and inappropriate AI platforms.

There is also a pressing need to focus on the development of sophisticated multi-class and multi-modal AI networks. These networks are essential for the early detection of diseases, monitoring of disease progression, prognosis, and guiding treatment decisions. Enhancing the sensitivity and specificity of DL in ophthalmology and broader medical diagnostics will undoubtedly improve patient outcomes. This includes tackling complex challenges such as utilizing fundus photographs to assess the optic nerve head, screening for various optic nerve pathologies in both dilated and non-dilated eyes, and forecasting myopia progression in pediatric patients.

Advancements in segmentation networks will refine the localization of anatomical structures and the detection of pathological changes, thereby bolstering the stability and reliability of AI systems in clinical practice. Despite the promising potential of numerous diagnostic AI algorithms proposed in the past, most have been limited by their training on small datasets, and only a handful have been tested in real-world clinical environments. The heterogeneity of patient populations poses a significant challenge, potentially diminishing the accuracy of AI algorithms in practical settings.

Before deploying of AI technology in clinical environments, it is essential to thoroughly evaluate potential hurdles, including the generalizability of models, the interpretability of algorithms, the medical community’s understanding of AI, and the legal and regulatory landscape. Taking these considerations into account will improve the consistency and transparency of reporting, enabling regulators and stakeholders to more effectively gauge the cost-effectiveness of AI-driven interventions.

As AI algorithms demonstrate their efficacy in clinical trials and are subsequently adopted in practice, this technology will be shared with all relevant parties, marking a significant milestone in the field of ophthalmology.

Conclusions

The integration of AI into ophthalmology has transformative potential, offering opportunities to refine patient care and streamline diagnostic methods. By facilitating the early detection of diseases and the development of tailored treatment strategies, AI underscores its pivotal role in advancing ophthalmic healthcare. Nevertheless, the effective deployment of AI technologies in this domain is not without its hurdles. Paramount among these challenges are the imperatives of maintaining the accuracy and reliability of AI systems, safeguarding patient confidentiality, and navigating the complex terrain of ethical considerations. Moreover, it is incumbent upon ophthalmologists to augment their expertise with a comprehensive understanding of AI and its applications to ensure that they can harness these sophisticated tools without compromising on patient welfare and the integrity of treatment outcomes. In essence, while AI is a beacon of innovation in the realm of ophthalmology, its successful adoption relies on a nuanced appreciation of both its vast capabilities and the multifaceted challenges it presents. Only with a deep and thorough understanding of AI can ophthalmologists unlock its full potential to improve standard of patient care and effectively overcome the barriers to its implementation.

To maintain their leadership in the rapidly advancing technological landscape of ophthalmology, it is imperative for ophthalmologists to engage in continuous education focused on AI. Creating an accessible and well-rounded curriculum that addresses the intricacies of AI is crucial. By surmounting the educational hurdles associated with AI and seizing the opportunities it presents, professionals in the field can significantly augment their capacity to deliver unparalleled care to their patients. As AI technology progresses, the educational frameworks for ophthalmologists must also evolve to ensure they have the necessary expertise and competencies to excel in the future landscape of ophthalmic medicine.

The seamless incorporation of AI into ophthalmology, alongside other medical disciplines, hinges on the synergistic relationship between healthcare professionals and AI technologists. This collaboration is essential to developing of AI applications that are not only at the pinnacle of technological innovation, but also have significant clinical value and adhere to ethical standards. By fostering this partnership, the medical community can fully harness the capabilities of AI to improve patient care, increase operational efficiency, and lay the foundation for future healthcare breakthroughs.