Advances and challenges from pathological mechanisms to intelligent quantified diagnosis in diabetic optic neuropathy

Differential diagnosis among DON types

One of the primary challenges lies in the difficulty of diagnosing different type of DON.^57,58 The clinical spectrum of DON ranges from acute ischemic events, such as DP or anterior ischemic optic neuropathy, to chronic, insidious neurodegeneration. Each subtype demands a distinct diagnostic approach, yet they often share overlapping features.^59–61 For instance, OD edema—a hallmark of acute presentations like papillopathy—may be mild and transient, easily overlooked during routine examination.^62,63 Conversely, the more common chronic progressive form presents with subtle, slowly evolving signs like OD pallor or RNFL thinning, which are nonspecific and may be attributed to other optic neuropathies, particularly glaucoma. ²⁶ This overlap complicates differential diagnosis, as DON must be distinguished from compressive, inflammatory, hereditary, and other acquired optic neuropathies, a process that often necessitates additional, costly investigations such as neuroimaging or functional tests.^12,36,64 The absence of universally accepted diagnostic criteria or a standardized classification system further exacerbates this problem, leading to heterogeneity in clinical practice and research.

Compounding this issue is the strong diagnostic lag inherent in DON. The neurodegenerative process begins long before clinical symptoms become apparent or detectable on conventional imaging. By the time patients report visual complaints such as reduced contrast sensitivity or minor field defects, a substantial proportion of retinal ganglion cells may have already been irreversibly lost. This lag is partly due to the brain's compensatory mechanisms, which mask early functional deficits. Consequently, the window for early intervention—potentially the most effective period for neuroprotective strategies—is often missed. This delay transforms management from prevention to damage mitigation, significantly impacting long-term visual outcomes.^13,65–68

The limitations of current examination methods, particularly imaging technologies (Table 3), represent another critical barrier. While tools like OCT have revolutionized objective assessment by quantifying structural loss (e.g. RNFL and ganglion cell complex thickness), they primarily detect atrophy only after significant neural tissue loss has occurred.^69,70 This structural approach fails to capture earlier, potentially reversible phases of the disease characterized by metabolic dysfunction, inflammation, or microvascular compromise. OCT angiography (OCTA), though promising for visualizing capillary dropout in the peripapillary network, lacks standardized quantitative thresholds for DON and is influenced by media opacities and artifacts. ⁷¹ Functional tests, such as visual field perimetry, are subjective, variable, and often insensitive to early neural loss.^72–74 Moreover, techniques like fluorescein angiography are invasive and provide limited specific information about neural tissue health. ⁷⁵ Thus, while each tool offers valuable insights, none alone provides a comprehensive picture of the disease's early evolution or severity. Currently, despite not universally adopted, some clinicians and researchers grade the severity of chronic DON based on structural and functional criteria,^40,76,77 similar to DR.

OPEN IN VIEWER

Table 3.

Key diagnostic tests in diabetic optic neuropathy.

Test	Role in diagnosing	Typical findings
Functional Assessments
Visual acuity test	A standard eye chart test to measure how clearly a patient can see at various distances.	In diabetic optic neuropathy, visual acuity may be blurred, diminished, or change frequently with blood sugar fluctuations.
Visual field perimetry	Maps the functional deficits in the patient's vision corresponding to optic nerve damage.	Altitudinal defect, arcuate scotoma, generalized depression, or central scotoma.
Visual evoked potential	Measures the electrical signals generated by the visual cortex in response to visual stimuli. Assesses functional integrity of the visual pathway.	Prolonged latency (delay) of the P100 wave, indicating slowed nerve conduction.
Color vision testing (dyschromatopsia)	Assess the patient's ability to distinguish between colors.	Faded or “washed-out” colors, particularly reds and greens, can indicate damage to the optic nerve.
Ocular exams and imaging
Comprehensive dilated eye exam (ophthalmoscopy)	Foundation of diagnosis. Assesses optic disc morphology, retina, and blood vessels.	Optic disc pallor, retinal nerve fiber layer thinning, venous sheathing. May have coexisting diabetic retinopathy.
Optical coherence tomography	The most important objective test—A noninvasive imaging test that uses light waves to capture high-resolution, cross-sectional images of the retina and optic nerve. Precisely measures the thickness of the retinal layers.	Thinning of the retinal nerve fiber layer and the ganglion cell layer, even before classic retinopathy appears. This is quantitative evidence of neural loss. Also vital for diagnosing macular edema, which often co-occurs with diabetic optic neuropathy.
Optical coherence tomography angiography	Advanced, noninvasive version of optical coherence tomography that images the retinal and optic nerve head blood flow without requiring an injected dye.	Reduced vessel density and perfusion in the radial peripapillary capillary network around the optic nerve head.
Fluorescein angiography	Primarily used to assess diabetic retinopathy. In diabetic optic neuropathy, it can rule out other causes and may show nonspecific leakage or hypofluorescence.	For diabetic optic neuropathy, wide-field or ultra-wide-field FA is often used to get a comprehensive view of the peripheral retina, where much of the vascular damage occurs. Can help rule out inflammatory or ischemic conditions. Findings are often nonspecific in diabetic optic neuropathy.
Magnetic resonance imaging	Helps exclude conditions such as optic neuritis (which shows optic nerve enhancement) or tumors (which would appear as a compressive mass).	No enhancement is typically seen in diabetic nonarteritic anterior ischemic optic neuropathy.

Finally, the lack of sensitive and specific biomarkers for DON remains a major obstacle. Unlike other diabetic complications, such as nephropathy (e.g. albuminuria) or DR (e.g. microaneurysms), DON lacks a validated biomarker for early detection, progression monitoring, or therapeutic response. Blood-based biomarkers reflecting neural injury (e.g. neurofilament light chain, NGF/BDNF, VEGF) or systemic oxidative stress (e.g. MDA, GSH/SOD, 8-OHdG, NFE2L2, PON-1, GGT) and inflammation (e.g. hs-CRP, tumor necrosis factor-α, IL-6, Homocysteine, AGEs, sRAGE, MCP-1, nuclear factor kappa-B, Adiponectin, GSK-3β, and miRNAs like miR-146a, miR-155 and miR-126) remain exploratory and nonspecific.^78,79 Imaging biomarkers beyond layer thickness, such as textural analysis, functional connectivity, or metabolic profiling, are areas of undergoing research but have not yet entered routine clinical practice.^80–83 Without such tools, clinicians are left relying on late-stage structural changes, perpetuating diagnostic delays and hindering the development of targeted therapies.

Distinguishing DON from other diabetic eye complications

Currently, a central clinical challenge in diagnosing DON lies in its significant phenotypic overlap with other common optic neuropathies, most notably glaucoma, but also optic neuritis and compressive neuropathies.

Glaucoma represents the most critical and common differential diagnosis for chronic progressive DON, as both present as painless, progressive optic neuropathies characterized by OD cupping and RNFL thinning. However, key distinguishing features exist. In glaucoma, neuroretinal rim thinning typically follows a pattern that violates the ISNT rule (where the rim width is normally largest in the Inferior, then Superior, Nasal, and Temporal quadrants), showing preferential loss in the superior and inferior poles, leading to vertical elongation of the cup. In contrast, DON more often presents with diffuse pallor and a more generalized RNFL thinning that tends to respect the ISNT rule. Fundus examination in DON may also reveal venous sheathing or features of concurrent DR. Furthermore, a notable clinical clue is that visual acuity in DON can fluctuate with blood glucose levels, a feature absent in glaucoma.^84,85

Optic neuritis is distinguished by its typical acute or subacute presentation, including vision loss associated with orbital pain exacerbated by eye movements, reduced color vision, and a relative afferent pupillary defect. Magnetic resonance imaging (MRI) is often definitive in the acute phase (within three months), showing optic nerve contrast enhancement and T2 signal hyperintensity—findings absent in DON. While both conditions can lead to RNFL atrophy in the chronic phase, the history is pivotal: optic neuritis follows an acute inflammatory event, whereas DON atrophy develops insidiously. ⁸⁶

Compressive optic neuropathies cause progressive visual loss through direct mechanical compression and are frequently asymmetric or unilateral. They may produce distinct patterns of quadrantic visual field defects corresponding to the site of compression. In contrast, DON typically manifests with altitudinal loss, arcuate scotomas, generalized depression, or central scotomas. Neuroimaging (MRI or computed tomography/CT) is the gold standard for excluding a compressive lesion, as it directly visualizes mass effects on the anterior visual pathway. ⁸⁷

Therefore, the current diagnostic paradigm for DON is hampered by its clinical variability, significant lag between pathological onset and detection, reliance on imperfect tools that prioritize late-stage structural changes, and the absence of validated biomarkers. Overcoming these challenges requires a multifaceted strategy integrating advanced imaging analytics, functional testing refinements, and the discovery of novel molecular or imaging biomarkers to enable earlier, more precise diagnosis and improve patient outcomes. Under such circumstances, AI and computer algorithms emerge as promising tools, offering the potential to translate the complex pathophysiology of DON into actionable intelligent diagnostic strategies.

AI in DR: the foundation

The application of AI in ophthalmology is most advanced and successfully translated into clinical practice in the domain of DR (Supplemental Table S1), where it has received regulatory approval (e.g. the first FDA-approved autonomous system, IDx-DR) for autonomous screening. These systems are predominantly built upon CNNs, a class of deep learning models exceptionally proficient at deconstructing and analyzing pixel-level data within images. The process involves training these algorithms on hundreds of thousands of fundus photographs that have been meticulously graded by panels of human retinal experts. Through this training, the AI learns to identify the presence of DR-referable lesions—such as microaneurysms, dot-blot hemorrhages, and hard exudates—with a high degree of accuracy.^20,88 For instance, the Singapore I Vascular Assessment software can identify microvascular lesions that significantly associated with the progression of DR through quantifying retinal vascular parameters, such as venular fractal dimension and central retinal vein equivalent values. ¹⁸ AEYE Diagnostic Screening is the first FDA-approved autonomous AI tool for DR screening. It is compatible with both portable handheld retinal cameras and conventional tabletop fundus cameras, enabling point-of-care screening to ensure regular monitoring of patients with diabetes. ⁸⁹

Beyond mere detection, these systems can classify disease severity according to standardized scales like the International Clinical DR Severity Scale or the more granular Early Treatment DR Study scale.^16,18 For example, the Multi-scale Spatial-aware Transformer Network, which comprises a spatial-aware module, a multiscale encoder, and a cross-fusion classifier. It employs a multiple instance learning strategy to accurately diagnose and grade DR. ⁹⁰ The EyeArt AI Automated DR Detection System is an FDA-cleared cloud-based retinal diagnostic software device that can accurately detect both more-than-mild DR and vision-threatening DR by analyzing digital color fundus photographs of patient eyes for signs of DR without physician oversight or need for dilation in most individuals. ¹⁶

This automation is transformative for public health, enabling high-volume, cost-effective screening that can be deployed in remote and underserved areas with a shortage of ophthalmologists, AI-powered DR screening stands as a powerful tool for reducing the global burden of preventable blindness.

The translational leap to DON

The remarkable success of AI in detecting and grading DR has established a powerful foundation and provided the technical blueprint for its application to DON. However, this translation represents a significant paradigm shift, as the diagnostic task is fundamentally different. While AI in DR focuses on identifying hemorrhagic, exudative, and proliferative vascular lesions,^91,92 its primary goal in DON is to detect far more subtle signs of neurodegeneration.^29,36,93 Specifically, the structural and functional changes affecting the optic nerve and the inner retinal neural layers. This requires a move from analyzing the retinal vasculature to quantifying the neuronal architecture, and some of the AI systems have the ability to meet certain requirements. For image segmentation, representative architectures including U-shaped Network (U-Net) and fully convolutional networks are deployed to achieve high-precision delineation of critical ocular anatomical structures. These models are distinguished by their end-to-end learning capability and superior segmentation accuracy, yet their clinical translation is impeded by the imperative for large-scale and annotated training datasets. ⁹⁴ For disease classification and recognition, CNNs and support vector machines (SVMs) serve as primary tools for DON severity grading and differential diagnosis. Notably, CNNs enable automated feature learning from raw imaging data but mandate substantial datasets to mitigate overfitting and ensure generalizability. And SVMs show strong performance in situations with small sample sizes but are hindered by the need for time-consuming manual feature engineering.^94,95 For feature extraction and regression, models such as Random Forests and Linear Regression are leveraged to predict continuous clinical variables, with the key advantage of enhanced interpretability. However, their predictive performance depends on the initial feature engineering processes. For multimodal fusion, Multiinput CNNs and Graph Neural Networks integrate the datasets spanning OCT, OCTA, Visual Evoked Potential and functional MRI (fMRI) to facilitate holistic DON assessment. By capitalizing on complementary information, these models exhibit great clinical potential. Nevertheless, their implementation is challenged by inherent model complexity and practical hurdles pertaining to multimodal data acquisition standardization. ⁹⁶ Table 4 summarizes common algorithm models and their potential applications in DON.

OPEN IN VIEWER

Table 4.

Common algorithm models and their potential applications in diabetic optic neuropathy.

Algorithm type	Representative models	Application	Advantages	Limitations
Image segmentation	U-Net, FCN	Precise segmentation of optic disc, optic cup, retinal layers (e.g. RNFL, GCL-IPL)	Capable of end-to-end learning, high precision	Requires large amounts of accurately annotated data for training
Disease classification and recognition	CNN (e.g. VGG, ResNet), SVM	Grading DON, differential diagnosis	CNN can learn features automatically; SVM performs well on small datasets	SVM requires manual feature engineering; CNN requires big data
Feature extraction and regression	Random Forest, linear regression	Predicting continuous variables like RNFL thickness, visual function scores	Relatively strong interpretability	Performance depends on the quality of feature engineering
Multimodal fusion	Multiinput CNN, graph neural networks	Fusing OCT, OCTA, VEP, fMRI data for comprehensive assessment	Leverages multisource information, high potential	Complex models, challenges in data acquisition and registration

U-Net: U-shaped network; FCN: fully convolutional network; CNN: convolutional neural network; VGG: visual geometry group; ResNet: residual network; SVM: support vector machine; DON: diabetic optic neuropathy; RNFL: retinal nerve fiber layer; GCL: ganglion cell layer; IPL: inner plexiform layer; OCT: optical coherence tomography; OCTA: optical coherence tomography angiograph; VEP: visual evoked potential; fMRI: functional magnetic resonance imaging.

The initial and most critical step in this quantitative analysis is the precise delineation of ocular structures. Recent advancements in OD and optic cup (OC) segmentation have leveraged sophisticated adaptations of deep learning architectures. M-Net establishes a robust U-Net-based foundation, employing a one-stage, multi-label system that integrates a multi-scale input layer, side-output layers, a multilabel loss function, and a polar transformation to segment the OD and OC jointly. ⁹⁷ Building upon this foundation, the Coarse-to-Fine Transformer Network represents a significant evolution by hybridizing the U-Net architecture with a Transformer module. This integration, augmented by a Multi-Scale Dense Skip Connection module, substantially boosts segmentation accuracy. ⁹⁸ Further pushing the boundaries of this architectural shift, ONH Detection Transformer (ODFormer) is a dedicated optic nerve head detection network based entirely on the Swin Transformer. It introduces two key innovations: a lightweight bidirectional feature recalibrator and a multiscale context aggregator, which work in concert to refine feature maps and capture global contextual information, thereby achieving state-of-the-art performance. ⁹⁹ Collectively, these models illustrate a clear trajectory from enhanced CNNs to hybrid and finally pure Transformer architectures, all aimed at delivering more precise and reliable OD or optic nerve head measurements. In the context of DON, these models perform several essential functions: they precisely define the boundaries of the OD and OC to calculate the cup-to-disc ratio—a key metric for differentiating DON from glaucoma ¹⁰⁰ —and they automatically segment and measure the thickness of the RNFL and the ganglion cell complex from OCT scans. This provides highly reproducible, objective quantitative data on neural loss, offering a sensitivity far greater than qualitative clinical assessment.

Beyond segmentation, AI is directly applied to disease detection and classification as a diagnostic support tool.^101,102 CNNs—such as Residual Network (ResNet), Visual Geometry Group, and Inception—are trained to analyze both fundus photographs and OCT scans.^103–106 These models learn to identify direct signs of DON from fundus images, including OD pallor, RNFL defect patterns (e.g. localized wedge-shaped defects), and venous sheathing. More powerfully, when applied to entire 3D OCT volumes, CNNs can classify eyes as healthy or affected by DON by identifying complex, often imperceptible patterns of thinning and textural changes across layers. A crucial application is their ability to differentiate DON from other optic neuropathies, particularly glaucoma, by recognizing the distinct patterns and topography of RNFL loss characteristic of each disease.

Perhaps the most promising and advanced application of AI in DON is in the realm of predictive analytics. Using Recurrent Neural Networks or hybrid deep learning models integrated with traditional survival analysis techniques, algorithms can analyze longitudinal data from sequential OCT scans. By modeling the individual rate of RNFL thinning over time, these systems can predict which patients are at the highest risk for rapid disease progression. This enables a shift toward personalized medicine, allowing clinicians to tailor more aggressive monitoring and treatment plans for high-risk individuals.

Finally, AI excels at the discovery of novel imaging biomarkers through a process known as radiomics.^107–109 An OCT-omics prediction model, including 11 key features screened using the Mann–Whitney U test, utilizes logistic regression, SVM, and backpropagation neural network classifiers to assess anti-VEGF treatment response in patients with diabetic macular edema. ¹¹⁰ An unsupervised deep learning architecture for OCT images enables unbiased, data-driven biomarker discovery. It identifies subclinical markers, including known and novel ones, with no prior domain knowledge. The novel one showed positive correlation with visual acuity and were validated via clinical outcomes (visual acuity, lesion activity, retinal morphology) using correlation and machine learning regression. ¹¹¹ By extracting hundreds of quantitative texture features (e.g. entropy, contrast, fractal dimension) from the RNFL and ganglion cell complex on OCT scans, AI can detect subvisual changes that represent the earliest signs of axonal dysfunction. These alterations may be identified while the disease is still in a potentially reversible stage, often long before clinically significant thinning, which is more likely to reflect irreversible structure damage, becomes apparent. The power of AI lies in its ability to identify which specific combination of these radiomic features best correlates with a functional deficit (e.g. on visual field testing) or, most importantly, predicts future atrophy. This paves the way for the validation of novel, highly sensitive imaging biomarkers that could revolutionize the early diagnosis and management of DON.

Current dataset limitations

One primary limitation is data scarcity. Unlike DR, for which large, publicly available, well-annotated datasets exist, curated datasets for DON are severely limited. This scarcity originates from several key obstacles: the low diagnostic rate of DON, heterogeneity in imaging protocols across clinical centers, the high cost and expertise required for multimodal data acquisition and interpretation, and a lack of consensus labels for definitive diagnosis and staging. Consequently, the available data is often insufficient in both volume and quality. This data paucity forms a critical bottleneck, as it directly restricts the training, generalizability and robustness of AI models. Models trained on small, unrepresentative datasets are prone to overfitting, exhibit reduced sensitivity and specificity for early-stage disease, and often perform poorly when applied to diverse populations or imaging equipment not seen during development.

Translational limitations

Another limitation is translational challenge of certain experimental models to patients with long-standing or poorly controlled diabetes. In these clinically relevant populations, the applicability of such models faces substantial challenges. Notably, these models are inherently limited in their ability to recapitulate the full chronicity and cumulative metabolic burden characteristic of human disease. They also lack the granularity to distinguish between varying degrees of disease severity and progression, thereby failing to capture the heterogeneity observed in clinical practice.

Furthermore, while these models may be sensitive to the pathophysiological features typical of early neuropathy, this sensitivity is often diminished in the context of advanced disease. In patients with long-term metabolic dysregulation, the cumulative and severe tissue alterations that define chronic diabetic complications can obscure the more subtle signs of early optic nerve damage. Consequently, experimental findings derived from models of early or uncomplicated disease may not directly translate to the complex, multimorbid reality of patients with poorly controlled or long-standing diabetes, underscoring the need for cautious interpretation when applying such data to these populations.

Potential strategies

To overcome the significant data scarcity and heterogeneity challenges in DON AI research, a focused technological approach is essential. Promising strategies are emerging to build more robust and generalizable models. Foremost among these is federated learning, a decentralized framework that enables collaborative model development across multiple institutions without centralizing sensitive patient data. By training algorithms locally and aggregating only model updates, it effectively pools heterogeneous datasets while rigorously preserving patient privacy and adhering to data governance regulations. This directly addresses the core bottleneck of acquiring large-scale, diverse training data.

Complementing collaborative learning, synthetic data generation offers a powerful method for data augmentation. Advanced generative models, such as generative adversarial networks or diffusion models, can produce high-fidelity, annotated images of the optic nerve that mimic the appearance of true DON pathology. These synthetic datasets can artificially expand limited training cohorts, mitigating overfitting and enhancing a model's ability to generalize to rare or subtle presentations. Furthermore, to ensure models trained in one environment perform reliably in another, domain adaptation and generalization techniques are critical. These algorithms are specifically designed to minimize “domain shift"—the performance drop caused by differences in imaging devices, acquisition protocols, or patient demographics between sites. By aligning feature distributions or learning domain-invariant representations, these methods enhance model robustness and transferability across the varied landscape of clinical practice.

Integration of emerging quantitative biomarkers for AI-based detection

The true power of AI will be unlocked by fusing these diverse data streams. Beyond traditional structural parameters, several emerging imaging biomarkers show promise for early DON detection and can serve as endpoints for AI models, including Structural, Vascular, and Functional–Structural Correlation. ¹¹² Besides OCT, advanced imaging techniques provide a richer dataset for AI algorithms to analyze. OCTA can noninvasively quantify vascular perfusion around the optic nerve.^113,114 Diffusion Tensor Imaging (DTI), an MRI technique, can assess the microstructural integrity and organization of the optic nerve fibers themselves by measuring water diffusion. ¹¹⁵ fMRI can reveal altered functional connectivity within the entire visual pathway and the brain.^82,116 A key future direction is the development of sophisticated AI architectures—such as multiinput CNNs or graph neural networks—capable of integrating these multimodal imaging data (OCT, OCTA, DTI, fMRI) with clinical variables (e.g. HbA1c, diabetes duration) to build a comprehensive, sensitive, and specific assessment system for DON.

AI-driven radiomics and multimodal integration will validate these biomarkers as robust endpoints for AI-based DON detection. By prioritizing these emerging parameters, AI models can shift DON diagnosis from late-stage atrophy detection to early identification of disease onset. Specifically, AI has the potential to integrate and interpret an array of emerging indicators that surpass the sensitivity of conventional clinical metrics. Firstly, unearthing earlier, more subtle signs of DON by analyzing textural and microstructural changes in the RNFL and ganglion cell layer on OCT that precede overt thinning. Secondly, finding patterns in blood flow derived from OCTA, such as reduced peripapillary capillary density. It is more dynamic that may reveal impaired perfusion states earlier than static tests. Thirdly, detecting specific functional changes that correlate with nerve damage, by integrating data from visual field testing and other psychophysical tests. AI excels at modeling the complex relationship between structure and function, which can serve as a specific signature for DON. Finally, the most transformative application of AI will be in longitudinal analysis and predictive modeling. By applying AI to sequential imaging data from multiple patient visits, it becomes possible to dynamically track subtle progression patterns that are invisible to the human eye. This will enable the development of models that can predict an individual's risk of DON onset or their specific rate of progression, shifting the management paradigm from a one-size-fits-all approach to true early intervention and personalized management. This could fundamentally change patient outcomes by enabling preventative therapies to be deployed before significant vision loss occurs. Table 5 summarizes the potential clinical applications and the associated techniques of AI in the management of DON.

OPEN IN VIEWER

Table 5.

Clinical potential applications of AI in diabetic optic neuropathy.

Clinical applications	How AI is used	Key technologies
Early detection and diagnosis	AI analyzes OCT scans to identify patterns of RNFL and GCC thinning that are characteristic of early DON, even before it's clinically apparent.	CNNs can classify eyes as healthy or DON-affected. They often detect texture changes and patterns of thinning that are imperceptible to human observers.
Differential diagnosis	AI helps differentiate DON from other optic neuropathies, like glaucoma.	Algorithms are trained on datasets from both diseases. They learn that glaucoma typically has a specific pattern of RNFL loss (inferior/superior poles), while DON may be more diffuse.
Discovery of novel biomarkers	Using radiomics, AI extracts hundreds of quantitative texture features (e.g. entropy, fractal dimension) from the RNFL and GCC.	These subvisual texture features may change before measurable thinning occurs, potentially serving as the earliest-ever biomarkers for DON.

AI: artificial intelligence; RNFL: retinal nerve fiber layer; OCT: optical coherence tomography; GCC: ganglion cell complex; DON: diabetic optic neuropathy; CNNS: convolutional neural networks.

Proposed diagnostic workflow of DON

The current diagnostic reliance on parameters like RNFL thickness measured by OCT presents a fundamental problem: these metrics often only detect damage after significant and irreversible neural loss has already occurred. This diagnostic lag underscores the critical need for novel, computer algorithm-based quantitative indicators that can identify the earliest stages of disease. Therefore, the proposed diagnostic workflow for DON would integrate AI into a multistep clinical pathway to enable earlier and more precise diagnosis (Figure 2). It begins with the initial assessment of high-risk diabetic patients using standard ophthalmic exams, OCT, and visual fields. For complex or early-stage cases, the workflow incorporates an AI-augmented analysis. Here, a secure platform aggregates de-identified, multimodal data—including advanced imaging like OCTA and clinical history—for processing by specialized AI models. These models perform multimodal integration, extracting and quantifying subtle structural, vascular, and functional–structural correlation biomarkers that are imperceptible to conventional analysis. The AI generates a comprehensive decision-support report, providing a probability score, visualizing abnormalities, and offering differential diagnostic guidance. This assists clinicians in confirming DON, staging the disease, and ruling out mimics like glaucoma. Finally, the system facilitates personalized management and enables AI-powered longitudinal monitoring. By tracking micro-changes over time, it can detect progression earlier than standard methods and holds the potential to predict individual disease trajectories, shifting management toward proactive, personalized intervention before significant vision loss occurs.

OPEN IN VIEWER

Figure 2.

Diagnostic workflow of diabetic optic neuropathy.

Transparency and validation of future AI models

For successful clinical adoption, medical AI must transcend the “black box” problem. ¹¹⁷ Developing explainable AI (XAI) models is paramount.^118,119 Techniques like Gradient-weighted Class Activation Mapping (Grad-CAM) can generate heatmaps that visually indicate which regions of an image (e.g. a specific sector of the neuroretinal rim or a particular texture pattern in the RNFL) most influenced the algorithm's decision. This transparency is dual-purpose: it not only enhances clinician trust and facilitates integration into clinical workflow but also, in reverse, can discover new, biologically significant imaging biomarkers. ¹²⁰ By highlighting previously overlooked features that are highly predictive of disease, XAI can directly contribute to a better understanding of DON pathophysiology.

The clinical translation of AI's transformative potential for DON diagnosis hinges on addressing three pivotal challenges through forward-looking innovations: ensuring explainability, managing cross-domain variability, and establishing rigorous model validation. For XAI, moving beyond current heatmap-based interpretability, future models could leverage causal reasoning frameworks to map AI-derived imaging features to specific pathophysiological processes in DON. This could involve integrating biological prior knowledge into model architecture to provide biologically plausible explanations. Regarding cross-domain variability, the next frontier lies in developing adaptive AI systems that can dynamically adjust to new imaging protocols, patient populations, or clinical settings without retraining. This might involve meta-learning approaches where models are pretrained on diverse multimodal datasets (OCT, OCTA, and fMRI) to acquire generalizable feature representations. For model validation, future frameworks should move beyond traditional evaluation metrics to incorporate clinical utility and longitudinal predictive power. This could include prospective, multicenter trials that assess whether AI-assisted diagnosis improves patient-centered outcomes. Collectively, these advancements could transform AI from an auxiliary diagnostic tool to an integral component of personalized DON management.

Therefore, AI will evolve progressively with the discovery of new biomarkers and the growth of the volume of dataset. The ultimate goal is to develop explainable and sophisticated AI architectures which can integrate multimodal imaging data with clinical variables. Successfully developed AI systems promise to improve diagnostic efficiency, validate novel biomarkers, refine our understanding of disease progression and ultimately achieve a personalized and precise management paradigm for DON.