Diabetic retinopathy (DR) is one of the most common microvascular complications of diabetes and a leading cause of blindness in adults worldwide (1). As the prevalence of diabetes continues to rise globally, the incidence of DR is also increasing significantly. Retinal vascular abnormalities, which are hallmarks of DR, gradually lead to a decline in patients’ vision and, in severe cases, can cause blindness (1, 2). Given the current medical capabilities, the disease cannot be completely cured; treatment focuses on maintaining the patient’s existing level of vision. If DR can be diagnosed and treated early, in most cases, patients’ vision can be preserved. Early detection and timely treatment of DR are crucial for preventing vision loss (3).

Traditionally, the detection of DR relies on ophthalmologists’ manual evaluation of retinal images (4). However, this method is time-consuming, labor-intensive, and subject to variability due to the experience and subjective judgment of the evaluators, leading to inconsistent detection outcomes (5). Currently, most ophthalmologists still use traditional methods to diagnose diabetic retinopathy (DR) by analyzing the presence and types of abnormalities in retinal images. Microaneurysms (MIA), hemorrhages (HEM), soft exudates (SOX), and hard exudates (HEX) are the four most common types of lesions (3–5).

Manual detection of diabetic retinopathy (DR) images presents several issues. First, interpreting DR images requires trained ophthalmologists, but in underdeveloped countries, there is a severe shortage of ophthalmologists, leading to many patients being unable to receive timely screening and treatment (6). Additionally, the cost of DR examinations is high, making it unaffordable for many patients and causing them to miss the opportunity for early intervention. These issues contribute to the high prevalence and risk of blindness associated with DR (7). Since timely detection is crucial in preventing vision loss, scientists and engineers have been working to design automated methods to achieve accurate and rapid diagnosis and treatment. Automated methods not only address the shortage of human resources but also significantly reduce the cost of screening, benefiting more patients (8). In recent years, with the rapid development of machine learning (ML) and artificial intelligence (AI) technologies, ML models trained on a large number of fundus images have achieved high accuracy in automated DR classification (9). These models can quickly and efficiently analyze large volumes of images, allowing for a substantial number of screenings to be completed in a short time. To further improve detection performance, substantial effort has been invested in developing automated methods that are both efficient and cost-effective. These methods not only consider the accuracy of detection but also emphasize ease of use and cost control, making them more suitable for implementation in resource-limited settings (10).

Recently, advancements in optical coherence tomography (OCT) and retinal imaging technology have provided high-resolution image data for the early detection of DR (9, 10). OCT technology can generate detailed three-dimensional images of the retina, revealing subtle lesion features, allowing for detection of abnormalities at an early stage of the disease. These high-resolution image data greatly enhance the performance of automated detection systems, enabling more accurate identification and classification of DR lesions, thereby providing timely and effective treatment recommendations for patients (11). These images can capture minute changes in the retina, enabling more accurate detection of DR. With the rapid development of deep learning technology, significant breakthroughs have been achieved in the field of computer vision. Deep learning algorithms, particularly convolutional neural networks (CNNs), have demonstrated exceptional performance in image recognition and classification tasks and have been widely applied in medical image analysis. Kazakh-British et al. (11)conducted experimental research using relevant processing pipelines to extract arteries from fundus images and then trained CNN models to identify lesions (9). Alexandr et al. (12) compared two widely used classical designs (DenseNet and ResNet) with a new enhanced structure (EfficientNet) in their other work. Previous studies have shown that deep learning-based models can automatically analyze OCT and retinal images, accurately identifying and classifying different stages of DR (13).

Despite numerous studies exploring the application of deep learning in DR detection, their results and conclusions often vary, and a unified perspective has yet to emerge. Therefore, there is a need for a systematic review and meta-analysis to comprehensively evaluate the effectiveness of deep learning-based OCT and retinal image analysis for DR detection, clarifying its clinical value and future directions.

This study aims to systematically review and meta-analyze existing research to assess the accuracy and reliability of deep learning models in detecting DR from OCT and retinal images and to identify key factors influencing detection performance. Through this research, we hope to provide scientific evidence for clinical practice and promote the application and popularization of deep learning technology in ophthalmic diagnostics.

Reporting of this review and meta-analysis followed the PRISMA checklist. The study protocol was registered after the initial screening stage. The design of the inclusion and exclusion criteria of this study was based on the five main principles of the Participant-Intervention-Comparator-Outcomes-Study (PICOS) design search principle (14). Our PICO question was as follows: In deep learning applications developed based on retinal images for early screening of diabetic retinopathy (Participants), how does DL (Intervention) compare with traditional landmarks by a single expert or with scripted eye care provider referral and education (Control) in terms of accuracy (Outcome). The systematic evaluation program is registered on the International Prospective Systems Evaluation website (PROSPERO-CRD42024575847).

Currently, the assessment of the severity of diabetic retinopathy in patients heavily relies on manual interpretation of retinal fundus images, which poses significant challenges (31). Therefore, automated image grading systems play a crucial role in the early diagnosis and evaluation of these vision-threatening diseases. For example, deep learning algorithms and image processing techniques can analyze large volumes of fundus images, providing consistent and highly accurate diagnostic results, reducing human error, and improving diagnostic accuracy (32–34). By regularly collecting and analyzing patients’ fundus images, automated image grading systems can continuously monitor the progression of diabetic retinopathy, assisting doctors in timely adjusting treatment plans to achieve the best therapeutic outcomes (34). Multiple studies (35–37) have shown that deep learning algorithms can be used to generate expert-level grading diagnoses for retinal fundus images. However, these methods often achieve good performance at the expense of increased time complexity. Due to the same input image size in these independent models, the robustness of their classification is relatively poor. Therefore, this study employs a systematic review and meta-analysis to analyze the role of deep learning-based optical coherence tomography and retinal images in the detection of diabetic retinopathy.

The results of this meta-analysis confirmed that, compared to standard care, autonomous artificial intelligence improved the completion rate of diabetic eye exams in adolescents with diabetes [OR=1.88, 95% CI=(1.45, 2.44), p=0.031]. Risa M et al. (21) were the first to assess the role of artificial intelligence in narrowing the care gap among racially and ethnically diverse adolescent diabetic patients. The study indicated that closing the care gap for diabetic eye exams, as measured by MIPS and HEDIS quality indicators, was a crucial component of value-based care. The results suggested that autonomous artificial intelligence could help meet these historically challenging benchmarks, particularly among racially/ethnically diverse and resource-limited youth. Li et al. (28) confirmed that, although deep learning (DL) detection often showed larger deviations at points such as the porion, subspinale, gonion, articulare, and anterior nasal spine, DL might not exceed expert detection accuracy but could clearly assist both regular and experienced examiners in landmark detection. Training models on larger datasets might have eventually helped achieve or surpass expert accuracy. The results indicated that DL models included in the studies achieved an accuracy above 83% for identifying diabetic retinopathy. A total of 71% of established DL research models had detection accuracies exceeding 90%. Mehboob A et al. (27) proposed a DL architecture consisting of three phases: image pre-processing, feature extraction, and classification. Deep convolutional networks (CNNs) were trained to extract deep features. Heat maps extracted from the proposed framework highlighted the presence of any exudates, microaneurysms, hemorrhages, cotton wool spots, or new vessels, indicating feature extraction from the affected region and achieving high accuracy. Deep CNNs could take unknown images as input and extract problem-specific features, thereby generating an appropriate response. The results showed that the proposed technique outperformed existing ones in terms of sensitivity. Even with a lighter CNN architecture, it demonstrated competitive accuracy. Moreover, among ensemble-based architectures, the proposed framework achieved the highest accuracy using average pooling when trained on an augmented dataset. F. Mansour Romany (30)used deep convolutional networks to classify data into normal and diseased categories with an accuracy of 97.93%.

Common deep learning ensemble algorithm classifiers include Random Forest, Support Vector Machines (SVM), Neural Networks, K-Nearest Neighbors (KNN), Multilayer Perceptrons, Naive Bayes, Decision Trees, and Logistic Regression. In 2021, an ensemble-based machine learning algorithm was proposed (38), which combined three different classifiers: Random Forest, Support Vector Machines (SVM), and Neural Networks, with a meta-classifier for decision-making. This ensemble-based approach enhanced the robustness and performance of the algorithm. The algorithm was tested on the Messidor dataset and achieved an accuracy of 0.75. Another ensemble-based algorithm for diabetic retinopathy screening was proposed by Nagi, A in 2021 (39). This algorithm employed a two-stage classifier, where the first stage consisted of outputs from six classifiers: SVM, KNN, Multilayer Perceptron, Naive Bayes, Decision Trees, and Logistic Regression, followed by a second stage using a Neural Network to make the final decision based on the classifier outputs. The algorithm achieved a test accuracy of 76.40% on the Messidor dataset. In 2020, an ensemble-based deep neural network architecture was established. This model used ResNet (40) and leveraged four ResNets to perform binary classification among five categories of diabetic retinopathy: normal vs. mild DR, normal vs. moderate DR, normal vs. severe DR, and normal vs. proliferative DR. The results from each classifier in Stage 1 were then processed by an AdaBoost classifier in Stage 2 to obtain the final classification results. The algorithm was evaluated on the Kaggle dataset APTOS 3662 retinal images, resulting in an accuracy of 61.9%.

The study results indicated that to enhance the accuracy of artificial intelligence in diabetic retinopathy detection and assessment, an automated algorithm should have followed a two-step strategy (41). The first method involved automatically defining the acceptability of retinal images to determine if they qualified for automatic grading, and then only applying the automated algorithm if the retinal images passed the acceptability test. The second method suggested that to ensure global applicability of automatic grading, the development of automated algorithms should have used images that reflected the specific acquisition conditions in real-world programs, allowing the model to understand and leverage these unique characteristics. Among the most commonly used DL models in research, CNN included two different convolutional neural network (CNN) architectures: Inception-v3 and Inception-v4 (42). These architectures had significant differences in design and performance. Inception-v3 enhanced feature extraction capabilities mainly through improved Inception modules, which included multiple parallel convolutional and pooling layers, as well as 1x1 convolutions to reduce computational complexity. It also introduced batch normalization and separable convolutions to accelerate training and improve efficiency. In contrast, Inception-v4 built upon Inception-v3 by integrating residual networks (Residual Networks), introducing Inception-ResNet and Reduction-ResNet modules (43–45). These improvements gave Inception-v4 deeper network layers and better feature extraction capabilities, while residual connections addressed gradient vanishing issues in deep networks, enhancing training stability. Although Inception-v3 performed excellently in various computer vision tasks, Inception-v4 generally offered higher accuracy and faster training speed. Li et al. (28) developed an improved Inception-v4 network based on stem, inception, and reduction modules, and created an ensemble of five classification model instances based on this Inception-v4 network. Its performance level was comparable to or exceeded that of ophthalmologists, achieving excellent results on the primary dataset used. The detection accuracy was comparable to Inception-v3, but its responsiveness was notably higher than other Inception-v3-based DL models (25–27).

However, as all studies tested in this same dataset (and most also trained on this dataset), we likely have high comparability but limited generalizability. Future studies should aim to test DL models on broad data, demonstrating robustness and generalizability. This review and the included studies have a number of limitations. First, the precision and recallresults for some types of lesions in our study, which we mentioned above, were limited. More training data for these lesions should becollected to improve the performance of our model. Second, The established dataset was not necessarily a good representation of data from screening programs in clinical practice. Thereby, the built dataset was not sufficient to reflect the algorithm’s performance in broader clinical use. Future studies should consider including a wider outcome set and aim to test DL applications comprehensively in other study designs and settings (e.g., observational studies in clinical care, randomized controlled trials).

DL shows relatively high accuracy for detection of diabetic retinopathy, whether using a self-trained DL model or choosing an established AI model. The majority of studies focused on CNN(Inception-V3)to develop DL models. The results showed that the accuracy of DL models in evaluating diabetic retinopathy was highly consistent across different studies and superior to the control group, with no heterogeneity observed. Further validation with larger datasets is needed, and it is hoped that more randomized controlled trials will be used for model validation, and the true value of using DL in clinical care needs to be demonstrated. Future research should focus on standardizing datasets, improving model interpretability, and validating performance across diverse populations.