Artificial Intelligence in Clinical Diagnosis and Treatment of Dry Eye: Workflows, Effectiveness, and Evaluation

Abstract

Purpose:

To introduce the applications of artificial intelligence (AI) in the clinical diagnosis and treatment of dry eye (DE) and to explore its common workflows, effectiveness, challenges, and future development directions.

Methods:

This article conducts a literature review, focusing on the applications of AI in the diagnosis and treatment of DE. The primary search terms include “artificial intelligence”, “machine learning”, “deep learning”, “computer-aided”, and “Dry Eye”.

Results:

A total of 48 relevant original studies were identified, and their algorithms, sample characteristics, and data types were summarized. Through data analysis and image recognition, AI assists in DE examinations, identifies risk factors, aids diagnosis, and manages and monitors treatment. AI excels in enhancing diagnostic efficiency, accuracy, and objectivity, and shows promise in cloud-based treatment management. However, the applications of AI in DE also face certain challenges that need to be addressed.

Conclusions:

AI has the potential to revolutionize the diagnosis of DE and recommend personalized treatment strategies. This review summarizes existing challenges and offers clinicians and researchers a comprehensive, objective overview of AI applications and workflows in DE.

INTRODUCTION

Dry eye (DE) is a common ocular surface disease that affects the quality of life and vision of millions of people worldwide.1 The definition of DE is “a multifactorial disease of the tears and ocular surface that results in symptoms of discomfort, visual disturbance, and tear film instability with potential damage to the ocular surface. It is accompanied by increased osmolarity of the tear film and inflammation of the ocular surface”.2 The pathogenesis of DE involves multiple aspects including tear secretion, tear film evaporation, ocular surface sensation, neuroendocrine, immune inflammation, making the classification and diagnosis of DE very complex and challenging.3,4

Currently, diagnosing DE involves various subjective and objective methods.5 These examination methods face several issues: Inconsistent and inaccurate diagnoses due to a lack of unified standards,5,6 influence from subjective factors of patients and doctors,7,8 and insufficient sensitivity and specificity,9,10 leading to unreliable results. Therefore, there is an urgent need to develop new technologies and methods to improve the efficiency and accuracy of DE clinical diagnosis.

Recently, the use of artificial intelligence (AI) in medicine has led to increased research on its potential and effectiveness in aiding DE clinical diagnosis.11 AI is defined as “the science and engineering of making intelligent machines”,12,13 where intelligence is “the ability to achieve goals in a variety of environments”. AI can be divided into three levels: Weak AI, strong AI, and super AI.14 Weak AI refers to systems or models that can only exhibit intelligent behavior in specific domains or tasks. Strong AI refers to systems or models that can exhibit intelligence behavior equivalent to or surpassing humans in any domain or task.15 Super AI refers to systems or models that can exhibit intelligence behavior far beyond humans in all domains or tasks.16 Currently, The medical field primarily employs weak AI technologies like machine learning (ML), deep learning (DL), computer vision, and image processing to aid doctors in data analysis, image recognition, and diagnostic prediction.

There are currently some reviews that discuss the applications of AI in DE.11,17,18,19,20,21 Twelve reviews (from PubMed, accessed on June 2024) involving the applications of AI in DE were found during our edit of this review. However, most reviews discuss the applications of AI in anterior segment diseases or ocular surface diseases, with a secondary portion addressing the applications of AI in DE, rather than being specifically designed to focus on DE [Table 1]. There are only a few reviews that specifically and relatively comprehensively discuss the applications of AI in DE.11,17,18

Table 1.

Summary and comparison of existing reviews involving the applications of artificial intelligence in dry eye (from PubMed, accessed on June, 2024)

PMID	First author	Publication year	Purpose/specifically designed to focus on DE
34843999	Storås AM	2022	Briefly introduce AI, discuss its current use in DE research, and explore its potential for clinical applications/Yes
35724917	Swiderska K	2022	Summarize recent studies on meibomian gland imaging and share the latest technologies and approaches with the community/No
35656580	Brahim I	2022	Review traditional methods and automated diagnostic methods for four types of tests for DE/Yes
35819899	Kang L	2022	Review the most current clinical AI applications in anterior segment diseases/No
36605721	Ji Y	2022	Summarize the research progress of AI in the diagnosis of OSDs/No
36553174	Yang HK	2022	Provide an overview of studies regarding the application of AI in DE disease and discuss the recent advances in the integration of AI into the clinical approach for DE disease/Yes
36875760	Zhang Z	2023	Summarize AI research and technologies for diagnosing OSDs/No
36884203	Xu Z	2023	Provide an overview of AI applications and potential future applications in anterior segment diseases/No
36380089	Pur DR	2023	Review the literature on the application of AI and bioinformatics for analysis of biofluid biomarkers in corneal and OSDs/No
37421481	Pucchio A	2024	Review summarizes the applications of AI and bioinformatics methodologies in analysis of ocular biofluid markers/No
38448961	Tey KY	2024	Provide an update regarding recent advances in AI technologies pertaining to corneal diseases/No
38611606	Kryszan K	2024	Review focuses on the application of AI in analyzing in vivo confocal microscopy images for corneal diseases/No
Our review			Summarize the application scenarios and objectives, application methods, and evaluation of AI in the diagnosis and treatment of DE, and analyzes the challenges in current research and future improvement directions/Yes

AI: Artificial intelligence, DE: Dry eye, OSDs: Ocular surface diseases

Considering previous reviews, our narrative review aims to provide nontechnical readers, such as doctors and clinical researchers, with a layman’s explanation of the AI methods being used in DE today. This review includes three main aspects: (1) exploring the various applications and goals of AI in DE; (2) discussing AI’s application workflows, primarily introducing the general process of supervised learning to help nontechnical readers understand this process; and (3) evaluating the applications of AI, including its effectiveness in the diagnosis and treatment of DE, the challenges and limitations faced by AI systems or models in clinical applications, and future development directions and suggestions for improvement. Table 1 summarizes the first author, publication year, and purpose of the existing reviews involving the applications of AI in DE and compares them with our review.

Table 2.

Search strategies

Databases	Keywords
PubMed
#1	“Artificial Intelligence” [MH]
#2	(((((“Artificial Intelligence” [TW]) OR (“Machine Intelligence” [TW])) OR (“Deep learning” [TW])) OR (“Computer Aided” [TW])) OR (“AI”[TW])) OR (“Computer Vision System”[TW])
#3	#1 OR #2
#4	“Dry Eye Syndromes”[MH]
#5	((((((((((“Dry Eye Syndrome”[TW]) OR (“Dry Eye Disease”[TW])) OR (“Dry Eye Diseases”[TW])) OR (“Dry Eye”[TW])) OR (“Dry Eyes”[TW])) OR (“Evaporative Dry Eye Disease”[TW])) OR (“Evaporative Dry Eye Syndrome”[TW])) OR (“Evaporative Dry Eye”[TW])) OR (“Dry Eye, Evaporative”[TW])) OR (“Evaporative Dry Eyes”[TW]))
#6	#4 OR #5
#7	#3 AND #4
Web of science
#1	TS=(Artificial Intelligence OR Machine Intelligence OR Deep learning OR Computer Aided OR AI OR Computer Vision System)
#2	TS=(Dry Eye Syndromes OR Dry Eye Syndrome OR Dry Eye Disease OR Dry Eye Diseases OR Dry Eye OR Dry Eyes OR Evaporative Dry Eye Disease OR Evaporative Dry Eye Syndrome OR Evaporative Dry Eye OR Dry Eye, Evaporative OR Evaporative Dry Eyes)
#3	#1 AND #2

The publication dates range from January 2000 to December 2023

METHODS

The PubMed and Web of Science databases were searched (from January 2000 to December 2023) using the keywords “Artificial Intelligence”, “Machine Learning”, “Deep Learning”, “Computer-Aided”, as well as “Dry Eye”, “Dry Eye Syndrome”, and “Dry Eye Disease”. The screening was conducted independently by two researchers (XD, KY), following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. Irrelevant and duplicate studies were removed after reviewing titles and abstracts, and eligible studies were confirmed through full-text review. Any disagreements during the screening process were resolved through discussion among multiple researchers. Figure 1 shows the detailed flowchart of the study inclusion and exclusion process. The inclusion criteria include: (1) original studies involving AI algorithms; (2) studies focused on DE; (3) literature published between January 2000 and December 2023; and (4) studies published in peer-reviewed journals. The exclusion criteria include: (1) conference abstracts, editorials, and other forms of nonoriginal research; (2) articles unrelated to DE; (3) review articles; and (4) studies that do not involve AI algorithms.

Figure 1.

Flowchart of the study selection process

RESULTS

In the preliminary search, a total of 272 articles were identified from PubMed (n = 108) and Web of Science (n = 164). After removing duplicate records and irrelevant studies based on a review of the titles and abstracts, 59 articles were selected for full-text review. Finally, we identified a total of 48 relevant original studies. The main types of models or algorithms used in these studies, as well as descriptions of the sample characteristics, data types, potential application scenarios, and evaluation metrics, are summarized in Table 3.

Table 3.

Summary of data description, main methods, applications, and evaluation in relevant dry eye studies

Study	Data description	Main methods	Applications	Evaluation
Grus and Augustin22	Numerical data: Tear protein	ANN	AE; diagnosis	ACC: 89%
Grus et al.23	Numerical data: Tear proteins	ANN	AE; diagnosis	AUC: 93%
Yedidya et al.24	Images and videos: Tear film	IPT	AE; diagnosis	ACC: 91%
Koh et al.25	Images: Meibomian gland imaging images	SVM	AE; diagnosis	SPE: 96%, SEN: 98%
Rodriguez et al.26	Images: Conjunctiva	IPT	AE; diagnosis	ACC >64%
Ramos et al.27	Images and videos: Tear film videos	IPT	AE; diagnosis	SPE: 80%, SEN: 82%
Rodriguez et al.28	Images: Fluorescently stained corneal images	IPT	AE; diagnosis	CCC: 93%
Remeseiro et al.29	Images: Tear film interferometry images	IPT	AE; diagnosis	ACC: 96%
Remeseiro et al.30	Images: Tear film map	SVM	AE; diagnosis	ACC: 91%
Peteiro-Barral et al.31	Images: Tear film lipid layer images	Multiple criteria decision-making	Diagnosis	NR
Hwang et al.32	Images: Tear film lipid layer	IPT	Diagnosis	NR
Baǧbaba et al.33	Images: Fluorescein-stained cornea images	IPT	Diagnosis	CCC: 98%
Fernández et al.34	Numerical and categorical data: Tear sample	LR	IRF	AUC: 74%
Cartes et al.35	Numerical data: Tear film osmolarity	Logistic classifier	AE	ACC: 85%
Yang et al.36	Numerical and categorical data: Questionnaire	ANN	IRF	AUC: 79%
Nam et al.37	Numerical and categorical data	Decision trees	IRF	AUC: 70%
da Cruz et al.38	Images: Tear film lipid layer	RF	Diagnosis	ACC: 97%
da Cruz et al.39	Images: Tear film	RF	AE	ACC: 99%
Ćirković40	Numerical and categorical data: Virtual patient	NR	Diagnosis; treatment	NR
Zhou et al.41	Images: Meibomian gland images	CNN	AE; diagnosis	ACC: 92%
Stegmann et al.42	Images: Tear meniscus	CNN	AE	SPE: 98% SEN: 96%
Wang et al.43	Images: OCT images	CNN	AE	DSC: 97%
Setu et al.44	Images: Meibography images	CNN	AE	AUC: 96%
Elsawy et al.45	Images: OCT images	CNN	Diagnosis	AUC: 99%
Chase et al.46	Images: OCT images	CNN	Diagnosis	ACC: 85%
Xu et al.47	Images: IVCM images	CNN	AE	AUC: 96%
Wang et al.48	Images: Anterior segment images	CNN	AE	AUC >90%
Zheng et al.49	Images: Blink videos	CNN	AE; diagnosis	DSC >90%
Yu et al.50	Images: Meibography images	CNN	AE; diagnosis	ACC: 92%
Saha et al.51	Images: Meibography images	GAN	AE; diagnosis	ACC: 73%
Fineide et al.52	Numerical and categorical data: DE patients	RF, multilayer perceptron	AE; IRF	ACC >97%
Deng et al.53	Images: Meibography images	CNN	AE; diagnosis	DSC >69%
Jing et al.54	Images: IVCM images	CNN	AE	NR
Zhang et al.55	Images: Meibography images	CNN	AE; diagnosis	DSC: 93%
Jing et al.56	Images, numerical and categorical data: IVCM images	CNN	AE	NR
Zhang et al.57	Images and numerical data: Blink videos	CNN	AE; diagnosis	ACC >96%
Jing et al.58	Images and numerical data: Questionnaire, IVCM images	CNN	AE	NR
Edorh et al.59	Numerical and categorical data: Corneal epithelial thickness	RF; LR	AE	NR
Shimizu et al.60	Images and videos: DE videos	CNN	AE; diagnosis	AUC: 88%
Levine et al.61	Images: IVCM images of the cornea	CNN	AE	NR
Kikukawa et al.62	Images: Tear film	CNN	AE; diagnosis	AUC: 90%
Abdelmotaal et al.63	Images: Ocular surface videos	CNN	AE; diagnosis	AUC: 98%
Yokoi et al.64	Videos: Keratography data	CNN	AE; diagnosis	ACC >72%
Storås et al.65	Numerical and categorical data: Tear proteins	LGBM classifier	AE; diagnosis	ACC >72%
Li et al.66	Images, numerical and categorical data: Meibography images	CNN	AE; diagnosis	NR
Wang et al.67	Images: Tear meniscus height	CNN	AE; diagnosis	DSC: 99%
Wan et al.68	Images: Ocular surface images	CNN	AE; diagnosis	DSC: 88%
Wang et al.69	Images: Meibography images	CNN	Diagnosis	AUC >93%

DE: Dry eye, NR: Not reported, AE: Ancillary examination, IRF: Identification of risk factors, ANN: Artificial neural network, SVM: Support vector machines, IPT: Image processing technique, LR: Logistic regression, RF: Random forest, CNN: Convolutional neural networks, GAN: Generative adversarial network, ACC: Accuracy, AUC: Area under curve, SPE: Specificity, SEN: Sensitivity, CCC: Concordance correlation coefficient, DSC: Dice similarity coefficient, OCT: Optical coherence tomography, IVCM: In vivo confocal microscop

Through the summary of these studies, it was found that the application scenarios and objectives of AI in the clinical diagnosis and treatment of DE include AI assisting in DE examinations, identifying risk factors, aiding diagnosis, and managing and monitoring treatment. In addition, AI excels in enhancing diagnostic efficiency, accuracy, and objectivity, and shows promise in cloud-based treatment management. However, the applications of AI in DE also face certain challenges that need to be addressed. We will introduce and discuss these in the discussion section, which will be divided into the following three parts: (1) application scenarios and objectives of AI in clinical diagnosis and treatment of DE; (2) AI workflows for clinical diagnosis and treatment of DE; and (3) effectiveness and evaluation of AI in clinical diagnosis and treatment of DE.

DISCUSSION

Application scenarios and objectives of artificial intelligence in clinical diagnosis and treatment of dry eye

Diagnosing DE is a complex process that includes patient screening, symptom assessment, choosing examination methods, analyzing results, and making a diagnosis.63,70 Each of these steps requires physicians to make treatment plans based on the specific circumstances of the patient and their own experience, which can be influenced by various factors such as level of knowledge and experience, bias, and time pressure.71 Therefore, clinical diagnosis and treatment DE of faces the following issues: (1) lack of objectivity,72 meaning different physicians may give different diagnosis results for the same patient; (2) lack of comprehensiveness, meaning physicians may overlook or disregard certain important examination methods or indicators; and (3) lack of personalization, meaning physicians may not be able to provide the most suitable diagnosis results and treatment plans based on individual differences and needs of the patient.73

The findings of the search suggest that AI technology may be able to address the issues associated with clinical diagnosis of and treatment DE in the following ways: (1) assistance in DE examinations: AI assisted in DE examinations through data analysis and image recognition.24,25,26,27,28,29,32,35,39,41,42,43,44,45,47,48,51,54,55,56,57,58,59,60,61,62,64,65,66,67,68,69 AI could analyze eye surface images to detect DE signs and use meibography to assess gland dysfunction.44,48 In addition, AI algorithms accurately evaluate tear film stability.27,32,74 (2) Identification of risk factors of DE: AI algorithms aid in the identification of risk factors associated with DE.34,36,37,52 For example, Nam et al. utilized decision trees and Lasso regression to identify 13 significant factors associated with DED out of 78 potential variables.37 (3) Assisting in the diagnosis of DE: AI has shown high accuracy in diagnosing DE by objectively analyzing patient-reported symptoms and eye-related images,22,23,24,25,26,27,28,29,30,31,33,38,46,49,53,67,68,69 while also supporting the differentiation between DE and other ocular diseases.45 (4) Management and monitoring of DE treatment: Currently, there is optimism about using AI to manage and monitor DE.11,29,40,75 Some studies have attempted to integrate AI models into portable devices or cloud platforms for the detection and treatment recommendations of DE.29,60,75 These application scenarios of AI technology in clinical DE have been summarized in Figure 2.

Figure 2.

Current application scenarios of artificial intelligence technology in clinical diagnosis and treatment of dry eye

Currently, AI research for DE treatment is nascent, with ophthalmologists still leading treatment decisions.76,77 Clinicians tailor treatments based on symptoms, medical history, and exams, offering a flexibility AI has not yet matched. Meanwhile, the primary focus of AI’s applications in DE has been on disease diagnosis. There was a lack of emphasis on research related to treatments based on the physiological and pathological mechanisms of DE. There may be more possibilities to explore in this area.

Artificial intelligence workflows for clinical diagnosis and treatment of dry eye

The applications of AI in clinical diagnosis and treatment of DE mainly include two aspects: Algorithm models based on ML and DL, and technical means based on computer vision and image processing.78 DL is a subset of ML,66 the two differ significantly in their applications to the clinical diagnosis and treatment of DE, but they can also be combined to jointly diagnose DE.79 ML is favored for structured data and interpretability, while DL is better for unstructured data like images, focusing on automatic feature extraction.80

ML methods can be broadly categorized into two primary categories based on their methodologies: Supervised learning and unsupervised learning.81 Supervised learning predominantly encompasses tasks such as classification and regression, while unsupervised learning is primarily focused on tasks like clustering and association analysis, among others.82 Supervised learning, which depends on labeled data where the training set includes input features and corresponding output labels,83 is commonly used in the applications of AI in the clinical diagnosis and treatment of DE. In contrast, unsupervised learning, which works with unlabeled data to analyze and categorize it, is less commonly used. Figure 3 illustrates workflows to supervised learning. One notable feature of DL is the omission of feature engineering, resulting in substantial time savings.84

Figure 3.

General workflows of supervised learning in dry eye

The workflows of supervised learning, from preprocessing to diagnosis and treatment, involve key stages such as data collection and preprocessing, feature extraction, model training, and validation.85 Each stage is essential for maintaining the accuracy and effectiveness of AI models. The first step in the AI workflow is data collection, which involves gathering ocular images and videos, as well as numerical or categorical data for relevant clinical measurements and assessments. Both DL and ML models can utilize these data types.86 Preprocessing ocular images and videos usually involves key frame extraction, image enhancement (like contrast adjustment, noise reduction, and normalization), and image segmentation to highlight areas of interest such as the cornea, tear film, and meibomian glands.87 For numerical and categorical data, preprocessing typically includes removing missing values and standardizing the data for consistency and comparability.88

Feature extraction in AI technology for the clinical diagnosis of DE is widely applied and complex. Feature extraction in traditional image processing techniques primarily relies on mathematical algorithms.89 These techniques can be used to represent images as feature vectors for tasks such as image classification and object detection. Rodriguez et al. used a Blob detection algorithm to detect and count point-like areas (Blobs) in images for identifying point-like defects, which can help in detecting corneal staining scores related to DE.28 DL can automatically extract features from raw image, but they may lack interpretability.90 Some studies in DE diagnosis models try to improve model interpretability. Abdelmotaal et al. used class activation maps to show that the inferocentral paracentral cornea is crucial for CNN models in detecting DE from ocular surface images.63 Yokoi et al. used Grad-CAM++ heatmaps to demonstrate that severe aqueous-deficient DE showed extensive red areas with high blur, whereas increased evaporation DE had fewer red areas and less blur.64

After extracting features, AI models are trained to identify patterns associated with DE.91 To guarantee accuracy and reliability, models are validated with separate datasets to ensure they generalize to new data.92 Key performance metrics, including accuracy, sensitivity, specificity, and the area under the curve (AUC), are used to evaluate the performance of model. Some studies compare the results of model with those of ophthalmology experts to demonstrate that the AI’s decisions are more objective and consistent.50,63

AI models can be preliminarily used for clinical screening and diagnosis of DE after validation. Besides diagnosis, AI systems can also differentiate between different DE subtypes to propose personalized treatment strategies.64 Some devices also show promise for patient monitoring and management.11 Ćirković et al. and Remeseiro et al. investigated the incorporation of diverse AI models into assistive software and web-based systems,29,40 which not only offer predictive capabilities for DE but also provide treatment recommendations for DE. Similarly, Shimizu and colleagues integrated the Swin Transformer model into a portable smartphone attachment to collect ocular surface videos,60 potentially aiding in out-of-clinic DE detection.

Generally speaking, the workflows of AI in the clinical diagnosis and treatment of DE encompass stages such as data collection, preprocessing, feature extraction, model training, and validation. By collecting more DE-related images and data, and utilizing manual or semi-supervised labeling, the quality of the data and the balance of the samples can be ensured.93,94 In the future, further research on model interpretability is needed to enhance the reliability of AI models.

Effectiveness and evaluation of artificial intelligence in clinical diagnosis and treatment of dry eye

Studies have demonstrated that AI-driven diagnostic tools can match the accuracy of human ophthalmologists and are faster than ophthalmologists. In 2021, Elsawy et al. reported on a multi-disease DL neural network for the diagnosis of common corneal diseases, including DE, using anterior segment optical coherence tomography (AS-OCT) images.45 Their model achieved an AUC >0.99 and an F1 score above 0.90 for DE. In addition, Yu et al. developed a DL model based on CNN to automatically detect meibography images. The model achieved high accuracy levels, with a mean average precision exceeding 0.976 for conjunctiva and 0.92 for meibomian glands. According to the recent meta-analysis, the pooled estimate of accuracy for AI models in diagnosing DE is as high as 91.91%.95 Furthermore, the model processed each image in just 480 msonds, nearly 21 times faster than human specialists.50 In these studies, annotations by ophthalmologists are typically regarded as the gold standard. AI not only achieves a similar level of accuracy to that of ophthalmologists, but it also usually receives higher consistency assessment ratings than ophthalmologists.47,68 This suggests that AI models demonstrate superior reliability and reproducibility, making them particularly valuable in clinical settings where consistent and accurate diagnoses are essential.

Compared to traditional methods, AI has several advantages in clinical diagnosis of DE:11,96 (1) improving diagnosis efficiency, i.e., reducing the time and cost required for diagnosis; (2) increasing diagnosis accuracy, i.e., improving the accuracy, sensitivity, specificity, and other indicators of diagnosis results; (3) enhancing diagnosis objectivity, i.e., reducing the impact of subjective factors on diagnosis results; (4) increasing diagnosis comprehensiveness, i.e., increasing the amount and dimension of diagnostic information.

However, there are also several challenges and limitations of using AI in clinical diagnosis of DE, including difficulties in data acquisition, inconsistent data quality, lack of standardized diagnostic criteria, actual application risks, and algorithm monitoring issues.28,32,90 In addition, the integration of AI in DE raises various ethical, economic, and potential bias concerns.97 Ethically, the lack of transparency in DL algorithms can create opacity in decision-making for both medical professionals and patients, and there are also issues related to data privacy.98,99 Economically, the implementation of AI necessitates costly equipment, software, and ongoing support, potentially straining healthcare budgets.100,101 Data bias is also a significant issue,102 as most research data currently comes from single sources, which may lack diversity, affecting the generalization ability of model and leading to biases when applied to different regions and populations. Furthermore, AI systems are currently unable to provide humanistic care, yet psychological factors play an important role in the diagnosis and treatment of DE, particularly in aspects of patient communication and psychological support, areas in which AI systems are still unable to replace clinicians.103

To address these problems, we propose the following feasible suggestions:104,105,106 (1) data acquisition: Increasing the channels of data acquisition for DE through collaboration with partners, building data sharing platforms, and ensuring data quality; (2) data quality: Establishing a data collection system to ensure the accuracy of data sources; in addition, strengthening data cleaning efforts, enhancing noise and abnormal data processing, and improving data quality; (3) standardized diagnostic criteria: Holding regular work meetings, inviting industry experts to discuss and establish and improve diagnostic and treatment criteria for DE; (4) treatment risks: Establishing a risk control system for DE treatment to ensure the accuracy and safety of AI algorithms in treatment; simultaneously, establishing full-process records and tracking of the treatment process to promptly identify and address any problems.(5) data combinations: Exploring different types of data (such as biomarkers, clinical tests, demographic features, and various imaging data like AS-OCT) combinations can improve the accuracy of models, especially in the context of multimodal data fusion. The integration of multiple data sources provides more dimensional information, further enhancing the performance of the models.

This review may have the following limitations: it only searched the PubMed and Web of Science databases, which may induce bias. The workflows of AI were illustrated using only the recently more common supervised learning techniques. However, compared to previous work, the potential advantages of our review include a more detailed description of application scenarios and goals, making it more suitable for nontechnical readers to understand. In addition, it not only focuses on the applications of AI in the diagnosis of DE but also covers its applications and evaluation in treatment management.

In conclusion, AI has the potential to revolutionize DE diagnosis and recommend personalized treatment strategies. We reviewed the applications and workflows of AI in the diagnosis and treatment of DE, summarized existing issues, and proposed solutions for their optimization. We hope that this review can provide clinicians and researchers with a detailed and objective introduction while inspiring further related research.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

This work was supported by the Youth Project of the Medical Scientific Research Foundation of Wuhan Municipal Health Commission (No. WX21Q33), and the Scientific Research Fund Project of Aier Eye Hospital Group (No. AR2210D2). The funders had no role in the design, data acquisition or manuscript preparation of the present study.