Arora et al. employed convolutional layering and compound scaling strategies in AI models to develop the EfficientNet framework (15). Specifically, the EfficientNetB0 model was optimized in terms of depth, width, and resolution to classify DR severity (16). Recent studies have utilized Generative Adversarial Networks (GANs) to generate synthetic fundus images, thereby enriching the diversity of training datasets. These synthetic images were incorporated into DiaGAN-CNN, a transfer learning-based deep learning model, to improve the accuracy of image-based DR classification (16). The final model achieved an accuracy of 0.84255, a quadratic weighted Kappa of 0.79565, and an AUC exceeding 0.90 (17).

Chao Chen et al. adopted Inception-V3 as the base architecture, reduced network complexity using Global Average Pooling (GAP), and integrated a fully connected layer with ReLU activation and a Softmax output layer. The model was trained on the ImageNet and Kaggle DR datasets, and the system was optimized through asynchronous calling mechanisms, multithreading, and microservice architecture (18). Using the Baidu EasyDL platform, Cao et al. applied a transfer learning approach, retraining the upper-layer network parameters to construct a new classifier based on a pre-trained model and a publicly available diabetic fundus disease dataset from Kaggle. Their model achieved a Kappa coefficient of 1.00 (19).

Collectively, these technical advancements have substantially improved the accuracy, consistency, and clinical applicability of AI-based diagnostic tools for DR and PDR detection.

Researchers have noted that limited demographic diversity in training datasets may compromise the generalizability of AI models, resulting in reduced accuracy when applied to broader or underrepresented populations. To mitigate this issue, large-scale public datasets containing DR images of varying severity levels are commonly used to ensure adequate variability (15). In addition, a two-stage training strategy was employed in conjunction with the NASNet-Large pre-trained model to address the risk of overfitting due to excessive oversampling. This approach incorporated decision tree algorithms and the Synthetic Minority Over-sampling Technique (SMOTE) to manage data imbalance and improve model generalization (20).

A dataset of retinal scans from Brazilian patients was used to develop and validate a deep learning algorithm capable of diagnosing various stages of DR, including PDR. The algorithm achieved a sensitivity of 97.8%, a specificity of 61.4%, and an area under the ROC curve of 0.89, highlighting its high sensitivity and overall diagnostic performance (21).

Beyond imaging, Fatma et al. proposed a hybrid model that integrates metabolomics data to enhance both the interpretability and predictive power of DR diagnosis. Model performance was validated using multiple metrics, including 10-fold cross-validation and SHAP (SHapley Additive exPlanations), demonstrating improved predictive accuracy and greater transparency in clinical interpretation (22).

To further improve generalizability across imaging devices, Zhang et al. introduced a novel preprocessing method, Single-Channel Standard Deviation Normalization (SCSDN). SCSDN maintained consistent performance across images acquired from various fundus camera models, minimized algorithm dependency on specific hardware, reduced clinical deployment costs, and significantly improved diagnostic accuracy in real-world DR validation scenarios (23).

Collectively, these clinical validations underscore the robustness, adaptability, and increasing reliability of AI-based diagnostic systems in detecting and stratifying PDR across diverse clinical settings (Table 2).

Accurate preoperative planning enabled by AI can be seamlessly integrated with real-time intraoperative guidance systems to enhance surgical outcomes. Building on diagnostic advancements, AI plays a pivotal role during surgery by leveraging detailed structural and flow information derived from non-invasive imaging modalities such as OCT and OCT angiography (32). The incorporation of AI into surgical workflows has markedly improved the precision of PDR treatment.

For example, U-Net-based ensemble models have been employed for the non-invasive segmentation of microaneurysms (33). Other deep learning algorithms applied to OCTA not only achieve high diagnostic accuracy for DR and referable status but also generate class activation maps (CAMs), which visualize specific pathological regions—such as foveal avascular zone (FAZ) alterations and vessel density changes—thus potentially guiding intraoperative decision-making directly on OCTA images (34). Additionally, AI-driven systems can distinguish pathological neovascularization (NV) from compensatory angiogenesis, enabling targeted interventions such as suppression of NV using high-intensity, low-intensity pulsed ultrasound (LIPUS; 0.5 MHz, 210 mW/cm (2)), inducing endothelial apoptosis via the p38 MAPK/ER stress signaling pathway (35). Accurate AI-aided segmentation of these features is essential for precise localization and surgical planning (32). As a result, real-time intraoperative imaging integration becomes feasible, significantly improving surgical precision.

PDR pathogenesis involves retinal microvascular occlusion, which promotes NV formation. These newly formed, fragile vessels are prone to rupture, frequently leading to vitreous hemorrhage and pathological foci (27). To address this, AI models are instrumental not only in anatomical localization but also in assessing neovascular activity. For instance, deep learning methods applied to ultra-widefield fluorescein angiography (UWF-FA) can detect neovascular leakage with high accuracy (AUC = 0.96), effectively distinguishing active lesions requiring treatment from confounding retinal features, thereby guiding laser or surgical planning (36). In ischemic regions, AI-guided application of low-intensity LIPUS (1.5 MHz, 30 mW/cm²) can promote AKT-mediated angiogenesis, enhance local vascular density, and support metabolic recovery (37). Michael et al. developed a deep learning algorithm capable of identifying NV in fundus images and providing real-time intraoperative feedback, thereby supporting targeted and adaptive interventions during surgery (27). Collectively, these AI-assisted technologies substantially enhance the safety and precision of PDR surgery.

AI also plays a pivotal role in postoperative management by enabling predictive analytics and personalized medicine strategies for patients with DR. For instance, few-shot learning (FSL) combined with explainable AI (XAI) has been used to quantify macular features from OCTA images, enabling accurate assessment of recovery trajectories and facilitating individualized treatment planning (38).

In another study, Siri et al. developed an XGBoost-based predictive model using electronic health record (EHR) data to estimate 7-day and 30-day postoperative infection risks, thereby supporting systematic monitoring and early clinical intervention (39). These AI-powered tools contribute to optimizing recovery protocols and reducing complication rates, ultimately enhancing long-term clinical outcomes in DR management.

Current AI studies specifically targeting anti-VEGF therapy in PDR remain scarce, necessitating extrapolation from DME evidence. Research on AI applications targeting anti-VEGF therapy for PDR remains limited—a gap that underscores the need for a translational research paradigm informed by findings from DME. This approach is justified by three core pathological commonalities:

First, both DME and PDR share a VEGF-driven pathogenic mechanism. The pathogenesis of these conditions is closely associated with elevated VEGF levels (5, 40). Notably, oxidative stress under hypoxic conditions is significantly and positively correlated with VEGF concentrations in the vitreous fluid of patients with DME and PDR (41, 42). VEGF activates several downstream signaling pathways, including PLCγ–PKC–MAPK, PI3K–AKT, and RAC, which collectively regulate angiogenesis (43). While anti-VEGF therapy is the first-line treatment for DME (10), it also serves as an alternative or adjunctive option in PDR, with demonstrated therapeutic efficacy (44). Furthermore, VEGF forms complexes with proteins such as copper transport protein 1 (CTR1) via disulfide bonds (45), and assembles ternary structures involving endoglin (ENG), neuropilin 1 (NRP1), and VEGFR2 to enhance pro-angiogenic signaling (46). These pathogenic complexes can now be structurally analyzed using AI-based tools such as AlphaFold 3 (AF3), which predicts interactions among proteins, nucleic acids, small molecules, ions, and modified residues, providing novel insights into anti-VEGF therapeutic mechanisms (47).

Second, both diseases exhibit common imaging biomarkers of vascular mobility. DME is primarily induced by ischemia, which increases retinal capillary permeability and promotes microaneurysm formation (48). The resistance index (RI = [PSV – EDV] / PSV), an indicator of distal microvascular resistance, has been significantly associated with progression toward retinal non-perfusion. The vasculopathic processes seen in DME are expected to induce RI-like changes in vascular parameters in PDR as well (49). Wang et al. reported that diabetic patients with proliferative retinopathy exhibited the highest mean RI value (0.83), compared to significantly lower values in healthy controls (0.54, p ≤ 0.001) (50).

Third, there is a generalizability of treatment response prediction. Gross et al. demonstrated that anti-VEGF agents such as ranibizumab are effective in both PDR and DME, reducing the risk of visual field loss associated with panretinal photocoagulation (PRP) (51). Li et al. proposed a multimodal fusion architecture that separately extracts and hierarchically integrates structural and blood flow information, enhances feature representation via multiscale feature interaction, and applies a weighted average (Avg) of OCTA grading outcomes to avoid alignment-dependent feature fusion. The model uses ResNet50 as the backbone, optimized with the Adam optimizer and supported by data augmentation techniques. This multimodal fusion architecture demonstrates applicability to both PDR and DME (52).

Understanding these molecular interactions enables AI to predict therapeutic responses, as demonstrated in the following predictive models. Medical AI, particularly in predicting patient responses to treatment, has demonstrated considerable promise—especially in forecasting outcomes of anti-VEGF therapy for DME. Deep learning models—particularly those based on OCT image analysis—have been extensively employed to evaluate initial patient responses to therapy.

In one study, Rasti et al. employed a deep learning algorithm to analyze OCT images obtained before and after treatment to assess the therapeutic response in DME patients undergoing anti-VEGF therapy. Performance metrics such as AUC, sensitivity, precision, and specificity were used to evaluate the predictive model (53).

To accurately predict best-corrected visual acuity (BCVA), central subfield thickness (CST), cube volume (CV), and cube average thickness (CAT) from multimodal data in DME patients receiving anti-VEGF therapy, Leng et al. developed a deep learning model integrating a convolutional neural network (CNN) and a multilayer perceptron (MLP) (54).

Alternatively, generative adversarial networks (GANs) have been employed to predict OCT image outcomes. These models effectively identified key biomarkers—including intraretinal fluid (IRF), subretinal fluid (SRF), and hard exudates (HE)—enabling more detailed forecasts of treatment response in DME patients. Such approaches further assist clinicians in predicting both short- and long-term therapeutic outcomes (55).

Xin et al. proposed a separate prediction model, demonstrating that cube-measured foveal volume (CMFV) provided more accurate estimates of initial anti-VEGF treatment efficacy than CST. Their deep learning model estimated CMFV from OCT images, incorporating additional differentiation techniques to enhance predictive performance (56).

In another study, Sastry et al. developed the Notal OCT Analyzer (NOA), a machine learning system for quantifying retinal fluid volumes—including SRF, IRF, and total retinal fluid (TRF)—which are critical for evaluating patient responses to anti-VEGF therapy. Their results showed that fluctuations in these volumes were associated with treatment efficacy, providing clinicians with improved insight into therapeutic outcomes and prognostic potential (57).

Building upon these foundational studies, our analysis proposes a unified framework to optimize the role of AI in personalized anti-VEGF therapy.

China bears a substantial burden of diabetic retinopathy (DR), and AI-assisted solutions can be particularly valuable in economically underdeveloped regions. These approaches have the potential to significantly reduce the workload associated with image grading and lower capital expenditures (58).

To enhance treatment planning and minimize the unnecessary use of high-cost drugs, Anwesa et al. developed a hybrid deep learning model to predict responses to anti-VEGF therapy in patients with DME (59). Ruijie et al. applied a machine learning regression model trained on real-world data to predict the short-term efficacy of anti-VEGF therapy in DME patients. Model performance was assessed using mean absolute error (MAE), mean squared error (MSE), and the coefficient of determination (R²). This predictive capability is essential for both clinical and economic decision-making related to the short-term outcomes of anti-VEGF therapy (60).

AI-based prediction of anti-VEGF treatment response in DR and DME supports the development of personalized and effective therapeutic strategies (61).

In clinical decision-making, AI provides predictive insights to assess responses to anti-VEGF therapy, mitigate associated risks, and optimize therapeutic outcomes. Ying et al. developed a machine learning model to predict changes in best-corrected visual acuity (VA) in patients with DME 1 month after anti-VEGF treatment. This model supports clinicians in making informed treatment decisions, tailoring individualized therapeutic strategies, and managing patient expectations regarding treatment outcomes (62).

To obtain a more objective evaluation of visual function, researchers have used OCT imaging to infer visual acuity in DME patients (63). The resulting predictive models achieved R² values of 99.9% for DR, 97.7% for early DR, 93.9% for DME, and 98.4% for strong responders in the training set, and 96.3, 96.8, 79.9, and 96.3%, respectively, in the validation set.

Yuhui et al. combined multi-omics analysis with machine learning to enable early diagnosis of DR and DME and to predict responses to anti-VEGF therapy. Their model accurately forecasted early DR progression and treatment response in DME patients, offering a novel tool for clinical diagnosis and therapeutic planning (64).

Soumya et al. integrated AI algorithms to automate OCT data segmentation and analysis, resulting in reduced processing time and cost, faster diagnostic reporting, and improved workflow efficiency (65) (Figure 1).

Diagrams illustrating the associations investigated in this study are provided below: (A) Fundus photographs and OCT images are input and preprocessed using GAN and SCSDN to enhance image quality and improve model accuracy. (B) The core algorithm leverages CNNs, hybrid deep learning models, and ViT networks, incorporating optimization steps to enhance model performance, classification accuracy, and generalization. (C) The AI model detects lesion features in fundus images and classifies PDR, providing clinicians with a reliable diagnostic basis while enhancing treatment efficiency and patient quality of life. (D) Data Validation: The model’s performance was evaluated by calculating the area under the ROC curve, sensitivity, specificity, and other metrics. Its accuracy and feasibility were verified by comparing it with traditional physicians and handheld fundus cameras. (E) OCT, Optical Coherence Tomography; GAN, Generative Adversarial Network; SCSDN, Single Channel Standard Deviation Normalization; CNN, Convolutional Neural Network; ViT networks, Vision Transformer networks; ROC Curve, Receiver Operating Characteristic Curve.

One major limitation of current AI models is their heavy reliance on training data from limited geographic regions, resulting in notable generalization issues. For instance, Ting et al. developed a deep learning system (DLS) for diabetic retinopathy screening using multi-country datasets, with the AUC for referable diabetic retinopathy ranging from 0.889 to 0.983 across 10 externally validated datasets (66). This variability stems from the inclusion of datasets from diverse countries, leading to fluctuations in model performance and introducing uncertainty in clinical applications. Additionally, differences in the quality of fundus color photographs, the algorithms used, and the performance of imaging equipment across studies contribute to significant inconsistencies in AI diagnostic outcomes (67). Notably, these models have been found to be more sensitive to changes in camera equipment than human physicians.

Another critical challenge is the so-called algorithmic “black box” and the resulting clinical trust issues. In AI, “black box models” refer to algorithmic systems—particularly deep learning models—whose internal decision-making processes are difficult to interpret. Although such models often achieve high predictive accuracy, their opacity hampers trust, reproducibility, and clinical adoption. While models like ExplAIn have demonstrated a balance between performance and interpretability for DR diagnosis—through an end-to-end weakly supervised segmentation architecture with generalized occlusion regularization (68)—the issue of transparency remains unresolved. In most cases, interpretability is limited to identifying correlations. For example, Herrero-Tudela et al. employed SHAP to quantify and visualize feature contributions, marking a step forward in model interpretability (69). However, SHAP has inherent limitations: it can indicate the relative importance of metabolites but fails to uncover the causal pathways underlying key pathological mechanisms. This “trustworthy yet uninterpretable” nature continues to pose a major barrier to clinical adoption.

The real-world implementation of AI models continues to face substantial challenges. A primary issue is the regulatory divide and the absence of standardized validation protocols. Ong et al. conducted a global analysis of AI as Medical Devices (AIaMDs), identifying 36 devices from 28 manufacturers—97% (35/36) approved in the EU, 22% (8/36) in Australia, and only 8% (3/36) in the United States (70). These findings highlight significant disparities in regulatory approval processes across countries.

Moreover, the false negative and false positive rates of AI models remain critical concerns. A meta-analysis by Wang et al. reported that, despite improvements in AI-based diagnosis of ocular diseases, the false negative rate (FNR) of 12% and false positive rate (FPR) of 8.8% remain non-negligible (71). Even more concerning is the lack of globally harmonized standards for validating AI’s ability to assess dynamic disease progression.

AI deployment is further constrained by resource allocation challenges. Increasing sensitivity may enhance the identification of high-risk patients and improve clinical outcomes, but it also elevates healthcare costs. Conversely, increasing specificity may reduce unnecessary testing but risks missed diagnoses (72). This trade-off places a disproportionate burden on low-income regions. Additionally, AI systems trained on homogeneous datasets may underperform in diverse populations, leading to racial and ethnic disparities in detection rates (73).

Although Vision Transformers (ViTs) have shown promise in detecting diabetic retinopathy in clinical settings, their application is limited by high computational demands. Training ViTs requires high-performance GPUs, with memory usage exceeding 20 GB and power consumption approaching 400 watts per GPU. These resource-intensive requirements pose significant barriers to clinical deployment, particularly in settings with limited infrastructure and funding (74). Collectively, these multifaceted barriers underscore the long and complex path toward real-world adoption of AI in healthcare.

Currently, three critical gaps characterize AI research in this field: the absence of cross-model evaluation standards, which leads to fragmentation among diagnostic and therapeutic models—resulting in treatment delays and resource inefficiency; the lack of model generalizability, which impedes large-scale application; and the deficiency of multi-center validation, which causes the performance of otherwise high-precision, single-center models to deteriorate in real-world settings.

To address these challenges, future advancements should focus on three key areas: