Abstract

Introduction:

Recent advancements in neuro-ophthalmology necessitate integrative imaging methodologies to address the structural and functional complexities of the visual pathway. Conventional diagnostic tools, including magnetic resonance imaging (MRI) and optical coherence tomography (OCT), are constrained by limitations in spatial resolution, cross-modality integration, and interpretability, often resulting in diagnostic uncertainty in cases involving compressive neuropathies, demyelinating diseases, or unexplained visual field deficits. Deep learning approaches, despite their computational power, lack anatomical specificity and fail to incorporate domain knowledge critical for clinical interpretability.

Methods:

To address these challenges, we propose a multimodal framework that integrates OCT angiography with AI-enhanced MRI analysis through a symbolic-neural architecture. This framework employs the NeuroGraphPath model, which represents the visual pathway as a directed graph with anatomically defined nodes and parameterized transformations between regions, including the retina, optic chiasm, lateral geniculate nucleus (LGN), and visual cortex. The model incorporates spatial embeddings, learned decussation mechanisms, and anomaly detection modules to ensure biologically grounded and interpretable diagnostics. Additionally, the Chiasmatic Flow Inversion strategy facilitates bidirectional reasoning, enabling the tracing of activations to probable lesion sites with quantified uncertainty.

Results and Discussion:

Empirical evaluations demonstrate superior performance in lesion localization, uncertainty-aware reasoning, and interpretability compared to baseline AI models, particularly in complex visual field presentations. This integrated approach advances neuro-ophthalmic diagnostics by bridging imaging modalities and embedding anatomical reasoning, addressing the growing demand for precision and explainability in medical imaging research.

1 Introduction

Advancing the precision and effectiveness of neuro-ophthalmic diagnostics is of paramount importance, particularly in accurately assessing disorders involving the visual pathways. The integration of advanced imaging modalities not only enhances our anatomical and functional understanding but also enables early diagnosis and monitoring of disease progression. Optical Coherence Tomography Angiography (OCTA) provides high-resolution visualization of the retinal microvasculature, serving as a window into neurovascular health (1). Meanwhile, Magnetic Resonance Imaging (MRI), especially when augmented by artificial intelligence (AI), allows for detailed analysis of the brain and optic nerve structures (2). These tools, when used separately, have demonstrated substantial clinical value; however, their integration remains underutilized (3). Not only can a multimodal approach provide complementary information across structural and functional domains, but it also has the potential to uncover subtle pathologies missed by single-modality assessments (4). Furthermore, AI-enhanced analysis introduces capabilities such as automated segmentation, feature extraction, and pattern recognition, which significantly improve diagnostic reliability and reduce observer variability. Consequently, this integrated strategy offers a comprehensive and scalable solution for analyzing the visual pathway, particularly in complex or ambiguous cases.

Initial efforts to enhance neuro-ophthalmic diagnostics relied on manually designed frameworks that encoded clinical expertise into structured systems. These approaches utilized predefined rules and logical mappings to associate visual symptoms with potential pathologies (5). While these systems offered interpretability and consistency, they were often rigid and struggled to adapt to atypical cases or integrate diverse data sources (6). For example, early diagnostic tools were limited in their ability to process multimodal inputs such as imaging data, laboratory results, and patient history in a unified manner. Attempts to extend these systems with more flexible reasoning mechanisms provided incremental improvements but failed to address the growing complexity of neuro-ophthalmic data (7). As a result, these early methods laid the groundwork for computational diagnostics but were insufficient for handling the high-dimensional and heterogeneous data characteristic of modern clinical practice.

To overcome the limitations of these early systems, researchers began employing statistical models capable of learning patterns directly from data. These methods utilized algorithms such as support vector machines, decision trees, and ensemble techniques to analyze imaging datasets and identify disease-specific features (8). In neuro-ophthalmology, these models were applied to tasks like retinal layer segmentation, optic nerve analysis, and disease classification based on multimodal inputs (9). While these approaches improved diagnostic accuracy and adaptability, they often required extensive manual feature engineering, which limited their scalability and generalizability (10). Additionally, their performance frequently declined when applied to external datasets or rare conditions, highlighting challenges related to overfitting and data diversity (11). Despite these limitations, these methods represented a significant step forward by leveraging computational techniques to enhance diagnostic workflows.

The advent of deep learning marked a transformative shift in neuro-ophthalmic diagnostics by enabling end-to-end learning from raw imaging data. Convolutional Neural Networks (CNNs) and other advanced architectures eliminated the need for manual feature extraction, allowing models to automatically learn complex patterns from data (12). These techniques have been successfully applied to tasks such as retinal vessel segmentation in OCTA, optic neuropathy detection in MRI, and multimodal data fusion for comprehensive disease modeling (13). Pretrained models, fine-tuned on large-scale datasets, further improved performance and generalizability, even in data-limited scenarios (14). Moreover, innovations like attention mechanisms enhanced interpretability, enabling clinicians to identify decision-driving regions within the data (15). However, challenges such as the need for large training datasets and concerns about model transparency persist. Nonetheless, deep learning has revolutionized the field by providing scalable, high-fidelity tools for analyzing complex visual data, bridging structural and functional insights in unprecedented ways.

Based on the above limitations—including the rigidity of symbolic AI, the feature-dependence of machine learning, and the interpretability challenges of deep learning—we propose a multimodal imaging framework that integrates OCT Angiography with AI-enhanced MRI to facilitate comprehensive visual pathway analysis. This approach aims to leverage the high-resolution microvascular insights of OCTA alongside the macrostructural and functional detail provided by MRI. By incorporating AI techniques such as deep fusion networks and cross-modal transformers, the proposed framework enables joint analysis of disparate data types, capturing both localized retinal abnormalities and upstream neuroanatomical disruptions. Furthermore, the model is designed to optimize interpretability by embedding visual attention layers and uncertainty quantification, thereby enhancing clinician trust. This unified pipeline not only streamlines diagnostics but also improves early detection and longitudinal monitoring of neuro-ophthalmic conditions. In doing so, it addresses the longstanding fragmentation in visual diagnostics and offers a holistic, data-driven solution for advancing clinical outcomes.

Visual pathway analysis constitutes a foundational aspect of neuro-ophthalmology, offering a systematic approach to understanding the functional integrity and pathological mechanisms of the visual system. The visual pathway, spanning from the retina to the primary visual cortex, comprises anatomically and functionally distinct regions that serve as potential loci for pathological disruptions leading to visual impairments. The analysis of this pathway necessitates the integration of anatomical modeling, computational inference, and clinical signal interpretation to address the multifaceted challenges posed by neuro-ophthalmological diagnostics. To address these limitations, our proposed framework incorporates a deep fusion architecture that jointly processes OCTA and MRI data, thereby enhancing cross-modal feature learning and diagnostic accuracy. It demonstrates robust adaptability across diverse clinical scenarios, ensuring consistent performance among different patient populations and disease subtypes. Experimental evaluations further reveal that the framework surpasses traditional single-modality models in sensitivity and specificity, especially in identifying subtle anomalies along the visual pathway.

2 Related work2.1 OCT angiography vascular pattern analysis

Optical coherence tomography angiography (OCT-A) has significantly advanced the field of retinal vascular imaging by providing high-resolution, noninvasive visualization of microvascular networks. This imaging modality eliminates the need for dye injection, thereby enhancing safety and repeatability (9). Quantitative metrics derived from OCT-A, such as vessel density, perfusion area, and flow index, have been extensively studied as biomarkers for detecting neuro-ophthalmic disorders (11). Alterations in these vascular parameters have been linked to conditions like optic neuritis and ischemic optic neuropathy (12). The evolution of segmentation algorithms, from manual approaches to fully automated deep learning models, has improved the reproducibility and efficiency of OCT-A analyses (13). These algorithms now incorporate advanced features such as projection artifact removal and layer-specific segmentation, which are critical for accurate capillary plexus differentiation (14). However, segmentation errors remain a challenge, particularly in cases involving distorted retinal layers, prompting ongoing refinements in algorithm design (15). OCT-A has also been instrumental in detecting hemodynamic changes that precede structural atrophy, enabling earlier intervention in progressive optic neuropathies (16). Comparative analyses with fluorescein angiography underscore OCT-A's advantages in specificity and depth resolution (17). Normative databases stratified by demographic factors have been developed to establish baseline references, aiding in the differentiation of pathological deviations from normal variations (18). Efforts to integrate OCT-A with structural OCT data are advancing, with deep learning models trained on combined inputs to classify diseases such as glaucoma and papilledema (19). These models not only enhance diagnostic accuracy but also provide interpretable insights into disease mechanisms (20). Longitudinal studies utilizing OCT-A are shedding light on vascular remodeling in response to treatment, offering new avenues for monitoring therapeutic efficacy (21). Challenges such as motion artifacts, limited field of view, and device standardization continue to drive innovation in this domain (22).

2.2 AI-enhanced MRI visual pathway mapping

Magnetic resonance imaging (MRI) has become a cornerstone for investigating the central visual pathways, including the optic tracts, lateral geniculate nuclei, and visual cortex (23). Structural MRI sequences like T1-weighted and T2-weighted imaging are widely used to identify anatomical abnormalities such as demyelination and edema (24). Diffusion MRI, particularly diffusion tensor imaging, provides insights into white matter integrity and optic radiation tractography (25). Functional MRI (fMRI) has been employed to map cortical activation and retinotopic organization, revealing functional connectivity alterations in neuro-ophthalmic disorders (26). The integration of artificial intelligence has revolutionized MRI analysis, with deep learning models achieving expert-level accuracy in segmenting visual pathway structures (27). Machine learning algorithms utilizing diffusion metrics have been developed to differentiate between various etiologies of optic pathway injury, including inflammatory and vascular causes (28). Generative models capable of synthesizing missing imaging modalities have facilitated analysis in incomplete datasets (29). Functional connectivity networks derived from resting-state fMRI have been analyzed using graph-theoretical approaches to detect subtle changes preceding clinical deficits (30). Multimodal fusion frameworks that combine structural, diffusion, and functional MRI data have yielded composite biomarkers for assessing pathway integrity (9). Attention-guided models applied to fMRI data have localized regions of altered activation, correlating these findings with clinical measures of visual function (11). Validation studies have linked AI-based MRI predictions to visual field tests and electrophysiological outcomes, enhancing their clinical relevance (12). Prospective trials are evaluating the utility of AI-enhanced MRI in detecting optic radiation compression in intracranial tumors (13). Standardization efforts aim to harmonize imaging protocols across centers, enabling broader algorithm deployment (14). Challenges such as motion artifacts, limited resolution, and the need for large annotated datasets are being addressed through advanced training techniques and domain adaptation methods (15).

2.3 Multimodal fusion diagnostics development

The development of multimodal fusion diagnostics represents a transformative approach in neuro-ophthalmology, combining complementary imaging modalities to enhance disease characterization and localization (16). Integrative models that fuse OCT-A vascular metrics with MRI-derived markers of tract integrity and cortical function are being actively explored (17). These models utilize co-registered structural landmarks to align retinal and brain imaging data, enabling joint analyses (18). Early implementations have demonstrated improved sensitivity in detecting optic nerve compression by leveraging hybrid feature spaces (19). Dynamic fusion architectures adaptively weight contributions from each modality based on data quality, enhancing robustness (20). Generative adversarial networks have been employed to translate imaging information between modalities, addressing challenges related to missing data (21). Spatial registration techniques align retinal and brain coordinate systems, facilitating cross-scale mapping of pathological changes (22). Longitudinal studies integrating OCT-A and MRI have provided insights into the progression of vascular and neural degeneration, correlating these findings with clinical outcomes (23). Statistical models of cross-modal trajectories are being developed to predict functional prognosis and therapeutic response (24). Usability studies are evaluating the integration of multimodal outputs into clinical workflows, emphasizing the need for intuitive visualization tools (25). Regulatory frameworks are being established to validate these diagnostic systems, ensuring compliance with clinical standards (26). Pilot studies involving conditions such as multiple sclerosis and ischemic optic neuropathy are assessing the clinical utility of multimodal diagnostics (27). Challenges such as data synchronization, alignment precision, and interpretability are being addressed through modular pipelines and interactive interfaces (28). Shared databases with paired OCT-A and MRI datasets are facilitating the development of robust multimodal algorithms (29). Future directions include the incorporation of electrophysiological and visual field data, aiming to create holistic diagnostic frameworks that integrate structural, vascular, and functional metrics (30).

3 Method3.1 Overview

This work introduces a methodological framework designed to enhance the analysis of the visual pathway, addressing critical challenges such as data heterogeneity, signal ambiguity, and spatial-temporal constraints inherent in clinical imaging modalities, including MRI, OCT, and functional visual field assessments. The proposed framework conceptualizes the visual pathway as a multi-component system, wherein each stage of neural transmission is represented through parametric mappings and constraint-satisfying operators. This abstraction enables reasoning across varying anatomical scales while maintaining computational tractability in scenarios involving incomplete or noisy data. The framework is structured into three core components, each corresponding to a distinct layer of abstraction. Section 3.2 formalizes the visual pathway as a sequence of transformations over high-dimensional feature spaces, encoding structural and functional representations of visual signals. Mathematical symbols and operators are introduced to characterize inter-regional interactions, pathway symmetries, and diagnostic uncertainties. Particular emphasis is placed on modeling the optic chiasm, lateral geniculate nucleus (LGN), and the retinotopic organization of the visual cortex, as these regions are frequently implicated in clinically observable alterations. Section 3.3 presents NeuroSymNet, a symbolic-neural architecture designed for hierarchical encoding of visual pathway features. This architecture integrates anatomical priors with parametric symbolic structures, accommodating modality-specific embeddings and propagating symbolic constraints throughout the computational graph. The hybrid design supports interpretable predictions and enables backpropagation through anatomical topologies, facilitating diagnostic reasoning in neuro-ophthalmology. Section 3.4 details the computational strategy termed Chiasmatic Flow Inversion, which extracts diagnostic trajectories by leveraging the symmetrical and partially decussated nature of the visual pathway at the optic chiasm. This strategy employs invertible transformations grounded in functional symmetry operations, constrained inverse mappings, and co-variant vector transport along neural tracts, supporting uncertainty-aware inference in cases involving visual field defects, compressive neuropathies, or demyelinating lesions. By organizing the methodological exposition across these components, the framework provides a biologically informed and computationally robust approach to visual pathway analysis. Each section builds upon the symbolic foundation established in the preliminaries, contributing innovations that address the specific requirements of clinical reasoning in neuro-ophthalmology. The framework enhances interpretability in diagnostic modeling and introduces capabilities for structured generalization across anatomical and pathological variations.

3.2 Preliminaries

To establish a rigorous framework for analyzing the visual pathway in neuro-ophthalmology, the neural transmission of visual signals from the retina to the visual cortex is formalized as a sequence of structured transformations. Let represent the ordered set of anatomical regions in the visual pathway, where R denotes the retina, ON the optic nerve, CH the optic chiasm, LGN the lateral geniculate nucleus, TR the optic radiations, and VC the primary visual cortex.

For each region , a high-dimensional feature space Fi is defined, where any signal xi∈Fi encodes the spatial, temporal, and modality-specific representation of neural activity at that stage. The transition of visual information between two adjacent regions vi→vi+1 is modeled as:

where is a region-specific transformation parameterized by θi, which incorporates anatomical properties such as decussation ratios, projection angles, and conduction delays.

The optic chiasm introduces lateralization and partial crossover of visual signals. This is captured by a decussation operator , which splits the signal into crossed and uncrossed components:

where α∈[0, 1] is the chiasmatic decussation coefficient, and ΠC and ΠU are projection operators for crossed and uncrossed fibers, respectively.

The complete forward transmission of visual information from the retina to the cortex is represented by the composite operator:

resulting in the transformation:

Each transformation is constrained by the geometric structure and functional capacity of the corresponding region. This is formalized by requiring:

where is a family of Lipschitz-continuous mappings with constant Li, ensuring bounded signal distortion within healthy neural tissue.

To account for inter-subject anatomical variability, a deformation field is introduced, where is a template manifold for the visual pathway, and describes diffeomorphic warping from the template to subject-specific anatomy. The spatial perturbation at region vi is denoted as , and the spatially conditioned transformation is defined as:

The bilateral nature of the visual pathway necessitates the definition of paired signal paths ΦL and ΦR for the left and right hemifields. Input visual stimuli are encoded into lateralized signals:

where are lateral encoding functions corresponding to hemiretinal projections.

The cortical representation of the visual field is topographically mapped. A retinotopic projection function is defined as:

where ξ represents the cortical coordinates corresponding to a given visual stimulus. The inverse mapping from cortical location to potential lesion origin is expressed as:

providing a mechanism for tracing cortical disruptions back to their peripheral origins.

Uncertainty in signal transmission, arising from noise, measurement artifacts, or pathological disruptions, is modeled probabilistically. The stochastic transition kernel P(xi+1∣xi) at each stage is used to compute the total uncertainty propagation:

with marginal and conditional distributions approximated symbolically.

To support diagnostic modeling, a structural anomaly function Γ:Fi → is introduced:

This abstraction forms the basis for embedding clinical decision rules and lesion detection criteria.

To further clarify the diagnostic utility of the Chiasmatic Flow Inversion (CFI) strategy, we emphasize its ability to trace cortical activations back to their most likely upstream disruptions. The CFI mechanism leverages the known anatomical crossover at the optic chiasm to generate decussation-aware inverse mappings from the visual cortex to retinal origins. This process involves estimating latent variables per anatomical region, calculating symbolic Jacobians for backpropagation, and using dual-channel reconstruction to resolve the inherent ambiguity between crossed and uncrossed fibers. Furthermore, an auxiliary symbolic anomaly detection module evaluates the posterior probability of disruptions across the pathway. This enables clinicians to not only visualize which anatomical region is likely affected, but also assess the confidence of that inference, thus improving both localization precision and interpretability in complex visual deficit cases.

3.3 NeuroGraphPath: a symbolic-neural model for visual transmission

As illustrated in Figure 1, We propose NeuroGraphPath, a symbolic-neural hybrid architecture for modeling the visual pathway with high anatomical fidelity and computational interpretability. The model is designed to encode neural signal transformations across the primary anatomical regions of the visual pathway, as defined in Section 3.2, while respecting topological constraints, lateral symmetries, and pathological variability.

Schematic representation of the visual pathway analysis framework, spanning from the retina through the optic nerve, chiasm, lateral geniculate nucleus (LGN), optic radiations (TR), and culminating in the primary visual cortex (VC). This complete anatomical path enables a biologically grounded computational model for interpretable diagnostics.

Each synthetic modality is treated as a semantically meaningful input, passed through a modality-specific encoder. These representations are fused via cross-modal attention into the symbolic-neural graph. Anatomical priors constrain how functional proxies like perfusion loss or lesion saliency interact with the topological flow of visual signals.

3.3.1 Multimodal encoder architecture

As illustrated in Figure 2, the architecture of NeuroGraphPath is structured as a directed acyclic graph , where each node represents a specific anatomical region (e.g., retina, optic nerve, LGN), and each directed edge denotes a parametrized transformation capturing neural signal propagation from vi to vj.

Diagram illustrating the sequential neural transmission of visual signals through the visual pathway. Each anatomical region, from the retina (R) to the primary visual cortex (VC), is associated with a high-dimensional feature space (F1–F5) that encodes the neural activity at each stage. The transitions between regions are modeled as structured transformations, capturing the complex processing of visual information.

Each node is associated with a feature vector xi∈Fi, and the update rule for any region is given by:

where σ is a non-linear activation function (e.g., ELU), and is a symbolic-linear operator augmented with learned parameters θij:

where Wij, Bij are learned weights and biases, and Ωij encodes anatomical constraints such as fiber density, angle of projection, and decussation probability. To integrate biological symmetry, we construct parallel subnetworks and for the left and right hemispheric processing, coupled via a chiasmatic fusion layer that implements partial decussation via soft attention gating. Let denote pre-chiasmatic activations originating from the left and right eyes (retinas), respectively. We define the chiasmatic transformation as:

where WC and WU are learned cross (decussated) and uncross (ipsilateral) projection weights, respectively, and α∈[0, 1] is a learned decussation ratio shared across instances. Anatomically, the left LGN receives a crossed contribution from the right eye () and an uncrossed contribution from the left eye (), whereas the right LGN receives a crossed contribution from the left eye () and an uncrossed contribution from the right eye (). This formulation explicitly enforces the ipsilateral origin of uncrossed fibers and the contralateral origin of crossed fibers at the optic chiasm, consistent with known visual pathway anatomy. Each node further incorporates spatial metadata via positional encodings derived from anatomical templates. Let denote the anatomical position of node vi, and define the spatial embedding:

which is concatenated with each feature vector:

3.3.2 Graphical propagation layer

The graph propagation is executed over T time steps (depth levels), with each layer refining signal estimates and capturing hierarchical interactions:

where denotes the set of predecessors of node i. To support clinical reasoning, we integrate symbolic anomaly detection into each node via an auxiliary output head:

where ai∈ℝ is the anomaly activation score for region vi, and γ, Wa, ba are learnable parameters. A region is flagged as anomalous if:

with threshold τ determined via unsupervised calibration on normative data. To model uncertainty and enable bidirectional reasoning, we define a variational latent variable at each node. The mean and covariance are functions of the node's embedding:

with ψi learned per region. These latents are propagated forward through the graph via reparameterized transformations, enabling probabilistic diagnostics and posterior inference.

3.3.3 Cortical projection and symbolic inversion

The final cortical projection layer Ψ maps the propagated features at VC into visual field coordinates. For any input stimulus s, the corresponding cortical activation is:

which can be inverted to locate probable upstream disruption:

via gradient-based symbolic inversion within the NeuroGraphPath graph. NeuroGraphPath represents a biologically faithful, topologically constrained, and symbolically enriched neural model for end-to-end analysis of the visual pathway. Its structure supports both feedforward prediction and inverse inference, and its components are modularly extensible to multi-modal data integration and longitudinal modeling.

3.4 Chiasmatic Flow Inversion: a bidirectional reasoning strategy

As illustrated in Figure 3, to exploit the full expressive power of NeuroGraphPath, we introduce Chiasmatic Flow Inversion (CFI), a symbolic-inference strategy for backward reasoning across the visual pathway. This approach is specifically designed to support clinical interpretation of cortical-level visual deficits by reconstructing plausible upstream disruptions in the visual signal trajectory. CFI leverages three core mechanisms: decussation-aware invertible mappings: Discussion-aware invertible mappings, probabilistic transport via latent representations: Probabilistic transport via latent representations, cortical-to-retinal backpropagation through symbolic Jacobians: Cortical-to-retinal backpropagation through symbolic Jacobians.

Schematic representation of the NeuroGraphPath architecture. The visual pathway is modeled as a directed acyclic graph with anatomical nodes (Retina, ON, CH, LGN, VC). Directed edges encode parametric neural transformations. Left and right hemispheric subnetworks are fused at the chiasm via a decussation-aware fusion layer. Each node integrates anatomical position embeddings and anomaly detection heads, supporting interpretable, bidirectional inference.

3.4.1 Decussation-aware invertible mappings

As illustrated in Figure 4, this mechanism addresses the partial crossover at the optic chiasm, which introduces ambiguity in reconstructing retinal activations from cortical projections.

Schematic representation of the Chiasmatic Flow Inversion (CFI) framework, illustrating decussation-aware mappings for reconstructing retinal activations from cortical projections. Forward signal flow propagates visual information from retinal activation zones through the optic chiasm to cortical layers, while invertible mappings enable retrograde reasoning to estimate upstream disruptions. The framework incorporates dual-channel reconstruction to resolve ambiguities introduced by partial crossover at the optic chiasm.

The forward trajectory of visual information is defined from input stimulus s to cortical projection ξ:

where Φ is the composite operator over the visual pathway, and Ψ is the cortical projection layer. The inverse problem estimates a candidate retinal activation such that:

To compute this inverse mapping, a gradient-based symbolic estimator is introduced:

solved iteratively using gradient descent:

To resolve ambiguity, a dual-channel reconstruction is incorporated:

where xdirect assumes uncrossed projection and xcrossed assumes full crossover. The mixing coefficient λ∈[0, 1] is adaptively estimated using a symbolic gating function:

where σ is a sigmoid, and (ω, b) are learned parameters encoding cortical lateralization biases. Symmetry constraints are enforced through a bidirectional consistency term:

ensuring inferred retinal origins align in symmetrical field coordinates.

3.4.2 Probabilistic transport via latent representations

This mechanism propagates uncertainty using latent variables at each node. Let be the latent encoding of region vi inferred via NeuroGraphPath. Uncertainty backflow is computed using the Jacobian of the inverse map:

where is the Jacobian matrix of Φ−1 evaluated at ξ. This enables probabilistic tracing of signal deficits to upstream regions with quantified confidence intervals. To detect lesion origins, a back-inference anomaly function Γ−1 is defined:

where Γ(xi) is the region-level anomaly indicator. This formulation identifies the anatomical locus with maximal posterior anomaly probability. Lesion simulation is supported for what-if diagnostics. Given a symbolic lesion , a perturbed pathway is defined:

and its impact on cortical activity is evaluated:

3.4.3 Cortical-to-retinal backpropagation through symbolic Jacobians

This mechanism reconstructs plausible upstream disruptions in the visual signal trajectory by leveraging symbolic Jacobians. The complete strategy of Chiasmatic Flow Inversion is executed in three stages:

Compute ξ = Ψ(Φ(xR)) for observed or simulated inputs.

Estimate from ξ using gradient inversion and decussation-aware decoding.

Identify Γ−1(ξ) using posterior decoding and lesion simulation for interpretability.

By combining symbolic inversion, probabilistic propagation, and anatomical constraint modeling, this mechanism provides a principled and interpretable approach for retrograde diagnostic reasoning across the visual pathway.

4 Experimental setup4.1 Dataset

Visual Pathway Imaging Dataset (31) focuses on the structural and functional characteristics of the visual pathway, encompassing high-resolution magnetic resonance imaging (MRI) scans of the optic nerve, optic chiasm, optic tract, and visual cortex. The dataset includes labeled volumetric data from both healthy subjects and patients with varying degrees of visual impairment. All images are preprocessed to ensure spatial normalization and intensity correction. Manual annotations of visual pathway components are provided by expert neuro-radiologists. The dataset is valuable for segmentation, disease classification, and anatomical modeling tasks. It enables the development of deep learning models that can assess visual function through anatomical representations of neural structures. The dataset also contains auxiliary information such as demographic metadata, clinical diagnosis, and visual acuity scores, which supports multimodal learning. This resource facilitates a robust evaluation of computer vision algorithms in tasks like structure-function correlation, pathological detection, and 3D reconstruction. The dataset adheres to strict ethical and privacy standards, with de-identified and institutional review board approval. Its public availability has supported several benchmark challenges in neuro-ophthalmic imaging.

OCT Angiography Neuro-Ophthalmic Dataset (32) consists of optical coherence tomography angiography (OCTA) scans capturing retinal and choroidal vasculature with micrometer resolution. It includes scans from patients diagnosed with neuro-ophthalmic disorders such as optic neuritis, ischemic optic neuropathy, and compressive optic neuropathy, as well as from age-matched healthy controls. The dataset is enriched with peripapillary capillary density maps, ganglion cell layer thickness measurements, and vessel skeletonization annotations. Each scan is linked to clinical labels like diagnosis category, disease duration, and severity grade, allowing for supervised learning. The high contrast of vascular structures makes the dataset suitable for tasks like microvasculature segmentation, vessel density quantification, and pathology detection. Data acquisition was performed using standardized imaging protocols across multiple ophthalmic centers, ensuring consistency. This dataset facilitates the development and validation of algorithms that integrate retinal vascular features into neuro-ophthalmic disease models. It also serves as a reference for correlating structural and perfusion deficits in the context of visual function loss.

AI Enhanced MRI Vision Analysis Dataset (33) integrates conventional structural MRI with AI-generated synthetic modalities that highlight pathology-relevant features. The dataset includes raw T1-weighted and T2-weighted scans along with derived quantitative maps such as lesion probability heatmaps, contrast-enhanced overlays, and visual pathway saliency projections. The synthetic modalities are generated using a pre-trained generative adversarial network trained on a large-scale ophthalmic dataset. Ground truth lesion masks, visual field test results, and functional MRI overlays are provided for a subset of cases. The dataset covers a wide range of neuro-ophthalmic diseases including demyelinating diseases, tumors, and hereditary optic neuropathies. It is intended for research on multi-channel image fusion, lesion localization, and progression modeling. All scans are co-registered to a common anatomical space to enable voxel-level comparisons. The dataset provides a testing ground for the integration of deep learning models with synthetic enhancement tools to boost diagnostic performance in low-contrast scenarios.

Multimodal Neuro-Ophthalmic Imaging Dataset (34) comprises a collection of paired imaging modalities including fundus photography, OCT, MRI, and visual evoked potentials (VEP). The dataset is curated to capture cross-modal representations of neuro-ophthalmic conditions such as papilledema, optic disc drusen, and compressive lesions. Each subject's record includes synchronized data acquired within a narrow clinical window to preserve temporal alignment. The dataset enables multimodal learning by aligning spatial features from different imaging techniques with functional outcomes. It contains expert-labeled segmentations, disease classification labels, and electrophysiological responses to visual stimuli. Standardized preprocessing is applied to all modalities, including intensity normalization, noise reduction, and anatomical co-registration. The inclusion of visual electrophysiology offers an additional functional layer that complements structural assessments. The dataset supports advanced fusion models and has been used to demonstrate improvements in classification accuracy through cross-modal interaction. It offers significant utility for tasks such as modality-specific feature extraction, multi-input neural network training, and longitudinal monitoring of disease progression. The term “AI-Enhanced MRI" refers to the combination of conventional structural MRI with AI-generated synthetic modalities that emphasize pathology-related features. These include lesion probability maps, visual saliency projections, and contrast-enhanced overlays derived from a pretrained generative model trained on ophthalmic imaging data. The synthetic outputs are co-registered with native MRI volumes and serve as additional input channels for model training, enabling improved sensitivity to subtle or diffuse visual pathway lesions.

All datasets followed institutional IRB-approved protocols. Subjects were included if they were aged 18–80, had confirmed neuro-ophthalmic diagnoses, and completed all imaging protocols. Exclusion criteria included prior ocular surgery (except cataract), media opacities affecting image quality, or neurological disorders unrelated to the visual pathway. The Visual Pathway Imaging Dataset consists of 122 subjects (69 pathological, 53 healthy controls), yielding 244 MRI scans. The OCT Angiography Dataset includes 108 subjects (42 optic neuritis, 28 ischemic optic neuropathy, 38 controls), with 216 OCTA scans. The AI Enhanced MRI dataset covers 96 subjects with a total of 192 MRI volumes, annotated for lesion type (tumor, demyelination, hereditary neuropathy). The Multimodal Neuro-Ophthalmic Imaging Dataset includes 84 subjects and 672 synchronized recordings (84 × 4 modalities × 2 eyes). Disease breakdown includes papilledema (22), optic disc drusen (17), compressive lesions (21), and controls (24). For co-registration, OCTA and fundus photos are aligned using vessel-based rigid registration. MRI is skull-stripped and co-registered to OCT volumes using mutual information-based affine alignment. VEP signals are time-locked to visual stimulus presentations, and synchronized with imaging sessions using a shared timestamp protocol and physiological markers. This protocol ensures multimodal data fidelity and consistency for joint model training and evaluation.

The synthetic modalities are generated using a GAN trained on paired MRI and OCT/OCTA datasets with expert-delineated lesions. Outputs include lesion probability heatmaps, saliency overlays, and approximated perfusion-deficit maps. Each channel encodes a distinct diagnostic perspective and is spatially aligned with structural MRI volumes.

4.2 Experimental details

All experiments are conducted using PyTorch 2.1 on a workstation equipped with dual NVIDIA RTX A6000 GPUs, each with 48 GB VRAM, and an Intel Xeon Gold 6330 CPU with 256 GB RAM. The models are trained using the AdamW optimizer with a weight decay of 1e−4 and an initial learning rate of 2e−4, which is decayed using a cosine annealing schedule. Batch size is set to 16 for 2D inputs and 4 for volumetric data. Training runs for 200 epochs with early stopping applied based on validation loss stagnation over 15 epochs. All experiments are repeated three times with different random seeds to ensure statistical robustness, and the average performance is reported.

For the input data, all images are resampled to a uniform resolution of 256 × 256 pixels for 2D modalities and 128 × 128 × 64 voxels for 3D scans. MRI volumes are skull-stripped and normalized to zero mean and unit variance. OCT and fundus images are enhanced using CLAHE and denoised with a non-local means filter. VEP waveforms are normalized across subjects and interpolated to a fixed time scale of 500 ms. In multimodal cases, modalities are spatially co-registered using mutual information maximization. During training, data augmentation is applied including random horizontal and vertical flips, elastic deformations, intensi