EyeNexus: Multifaceted Eye-Centric Platforms

• EyeNexus is a multifaceted term describing diverse, eye-centered computational platforms that integrate medical image analysis, optical wearables, and VR systems.
• It employs advanced methodologies such as U-Net segmentation, ensemble learning with self-attention, and metasurface optics for precise ocular analysis.
• The varied implementations enhance clinical screening, personalized auto-capture, and low-latency VR streaming, adapting dynamically to technical and network constraints.

EyeNexus is a name applied in recent technical literature to several distinct systems and platform concepts spanning medical image analysis, ophthalmic instrumentation, near-eye display optics, attention-aware smart eyewear, and gaze-driven VR streaming. The most concrete uses are an end-to-end OCT-based eye-disease prediction web application for classifying Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), Drusen, and Normal (Naik et al., 2023), and a later gaze-driven, bandwidth-aware foveated streaming system for VR cloud gaming that combines real-time eye tracking with spatial compression and video encoding control (Wu et al., 15 Sep 2025). In adjacent technical syntheses, the same name is also used for envisioned platforms built around metasurface contact-lens displays, dual-modality ophthalmic screening devices, orthogonal-view OCT, and attention-aware moment auto-capture (Lan et al., 2019, Peng et al., 13 Apr 2025, Mazlin et al., 2022, Chang et al., 2021). This multiplicity is essential to understanding the term: EyeNexus does not denote a single canonical architecture across the cited literature.

1. Scope, nomenclature, and research contexts

The literature associates EyeNexus with several technically different artifacts. In one context it is an ophthalmic AI web application for OCT triage; in another it is a VR systems contribution with explicit latency, bitrate, and quality-control mechanisms; in others it functions as an envisioned integration label for near-eye optics or portable screening hardware (Naik et al., 2023, Wu et al., 15 Sep 2025, Lan et al., 2019, Peng et al., 13 Apr 2025).

Context	Defining characteristics
OCT disease prediction	Raw OCT upload, custom U-Net segmentation, InceptionV3/Xception ensemble with self-attention
Near-eye AR concept	Passive hologram metasurface on the cornea, pixel-by-pixel retinal projection
Attention-aware smart eyewear	Temporal visual attention network, gaze-triggered scene analysis, on-device auto-capture
Dual-modality ophthalmic screening	Coaxial fundus imaging and refraction measurement with Dense-U-Net pupil segmentation
VR cloud gaming	Real-time gaze-driven FSC and FVE with network-adaptive foveation control

A common misconception is to treat these appearances as a single unified product line. The available record does not support that interpretation. Instead, the name is used across multiple research narratives: as the actual name of a web application (Naik et al., 2023), as the title of a VR streaming system (Wu et al., 15 Sep 2025), and as a platform label in technical mappings that adapt independent optics or wearable-computing papers into an EyeNexus-oriented design space (Lan et al., 2019, Peng et al., 13 Apr 2025, Chang et al., 2021, Mazlin et al., 2022, Wirth-Singh et al., 2024). This suggests that EyeNexus functions less as a single fixed system than as a reusable umbrella designation for eye-centric computational platforms.

2. OCT-based disease prediction application

In its most explicitly medical form, EyeNexus is the end-to-end web application proposed in “Eye Disease Prediction using Ensemble Learning and Attention on OCT Scans” for automated prediction of four retinal categories—CNV, DME, Drusen, and Normal—from raw OCT images (Naik et al., 2023). The stated problem scope is screening and triage of OCT B-scans with a focus on early detection and rapid, accessible decision support. It is not positioned as a definitive diagnostic instrument; rather, it aids clinicians and patients by flagging likely disease categories for timely follow-up.

The pipeline begins with user registration, login, and raw OCT upload. Images are stored in secured Firebase storage, encrypted. Preprocessing comprises conversion to arrays, normalization of pixel intensities, resizing to $496\times496$ for U-Net segmentation and $299\times299$ for classification, and a data-augmentation stage referenced in the system diagram but not enumerated in the text (Naik et al., 2023). The segmentation module is a custom U-Net that denoises the raw OCT and produces a binary semantic mask of “affected” versus “not affected” regions, emphasizing hyper-reflective zones and affected inner retinal layers. These segmented images are then passed to the classifier.

The U-Net architecture follows the standard contracting and expansive pattern. The encoder uses repeated convolution-plus-pooling blocks with image size reduction $$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$, while the decoder uses $2\times2$ up-convolutions, copy-and-crop skip connections, concatenation at each scale, and a final $1\times1$ convolution for per-pixel prediction (Naik et al., 2023). The paper reports training accuracy and Intersection-over-Union but does not explicitly state the loss function used during U-Net training. It provides, for context, standard Dice, Dice-loss, BCE, combined BCE-plus-Dice, and IoU formulations, while noting that the exact loss used in EyeNexus’s U-Net is not specified.

The classification stage is an ensemble of ImageNet-pretrained InceptionV3 and Xception backbones with self-attention applied to each model’s feature maps. For a feature map $X$, the attention mechanism uses $f(x)=W_f x$, $g(x)=W_g x$, and $h(x)=W_h x$, with attention weights $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$ and attended output $299\times299$0 (Naik et al., 2023). The paper explicitly notes that it applies dot-product attention without the standard scaling term $299\times299$1. Attended feature maps are combined with original features through element-wise addition, but no multi-head attention, normalization layers, or deeper residual attention stacks are described. The ensemble “combines the predictions” from the two models, although the paper does not specify whether the aggregation occurs at the logit or probability level, nor any weighting scheme.

The reported dataset provenance is the Kermany et al. 2018 OCT dataset with 207,103 images across CNV, DME, Drusen, and Normal. Because of hardware limitations, classifier training uses a balanced 2,000-image subset with 500 images per class and evaluation uses 968 balanced test images with 242 per class (Naik et al., 2023). Segmentation is trained on 32 manually annotated OCT images, 8 per class, with an 80/20 train/test split. The U-Net is trained for 70 epochs; training/validation curves for the ensemble are shown across 50 epochs, but optimizer, learning rate, batch size, and related training hyperparameters are not reported.

Performance is reported at two levels. For segmentation, the U-Net reaches 94.69% training accuracy and 84.078% IoU on the test split from the 32 annotated cases (Naik et al., 2023). For classification, the individual self-attention models reach 90% accuracy for Xception, 86% for InceptionV3, 84% for DenseNet201, 58% for ResNet50, and 55% for VGG16 on the 968-image test set. The deployed InceptionV3-plus-Xception ensemble with self-attention reaches 96.69% overall accuracy, with precision $299\times299$2, recall $299\times299$3, and F1-score $299\times299$4 (Naik et al., 2023). The paper compares this to Singh et al. 2021 at 91.4%, Pin et al. 2021 at 91.6%, and Lu et al. 2018 at 95.9%, listing the proposed EyeNexus ensemble at 96.6–96.7%. Per-class precision/recall/F1, ROC-AUC, confusion matrices, calibration metrics, and external validation are not reported.

Operationally, the web application uses a Flask backend, Firebase authentication and storage, and an HTML/CSS frontend with Home, About, Contact, Register, and Login pages (Naik et al., 2023). The report page returns the predicted category and logs the image and date. Firebase encryption and mandatory authentication are described, but handling of protected health information, audit logging, and formal compliance attestations are not discussed.

3. Ophthalmic imaging hardware and screening-device interpretations

In a separate EyeNexus-oriented hardware interpretation, the platform is envisioned as a portable ophthalmic screening device that integrates non-mydriatic fundus photography and objective refractive error measurement on one coaxial optical axis (Peng et al., 13 Apr 2025). The cited system uses a long-pass dichroic mirror that reflects 400–760 nm and transmits 830–1300 nm, thereby reflecting $299\times299$5 nm NIR and $299\times299$6 nm visible light for fundus alignment and acquisition, while passing $299\times299$7 nm for refraction. A shared broadband industrial camera, the MV-CS013-60GN at $299\times299$8, serves both modalities, though the prototype mechanically detaches the fundus beam-splitter/eyepiece assembly during refraction because sensitivity at 980 nm is about 28% of peak (Peng et al., 13 Apr 2025).

The alignment and focusing loop is heavily algorithmic. A Dense-U-Net segments background, sclera, iris, and pupil from NIR anterior-segment images, after gamma correction and histogram equalization. Connected-component analysis and Canny-based edge analysis localize the pupil center, and the chinrest motor centers the pupil relative to the image center (Peng et al., 13 Apr 2025). The Dense-U-Net has approximately 0.25M parameters and is trained on OpenEDS with 8,916 train, 2,403 validation, and 1,440 test images for 150 epochs using cosine annealing from 0.01 to $299\times299$9. The composite loss is

$$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$0

with $$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$1, $$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$2, $$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$3, and $$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$4, where $$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$5 increases linearly with epochs and saturates at half the total epochs (Peng et al., 13 Apr 2025). Reported segmentation performance is F1 $$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$6, mIoU $$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$7, and Euclidean Distance Error for pupil localization of approximately 2.8 px.

The same device performs refraction measurement with a narrow-divergence 980 nm annular beam that forms a ring-pattern reflection. After Gaussian denoising, adaptive binarization, centroid extraction, radial probing, skeletonization, and ellipse fitting, the system uses the empirical linear relation

$$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$8

to link ring thickness diameter $$496\times496 \rightarrow 248\times248 \rightarrow 124\times124 \rightarrow 62\times62 \rightarrow 31\times31$$9 to diopter $2\times2$0 (Peng et al., 13 Apr 2025). The reported refractive estimation error is approximately 3–5% versus calibration eyes. The acquisition workflow is explicitly sequential—NIR alignment, automated pupil centering, passive autofocus, 980 nm refraction, and then $2\times2$1 nm fundus capture—though both modalities share the same optical axis.

A further ophthalmic-imaging interpretation relevant to EyeNexus derives from compact orthogonal view OCT for cell-resolution imaging of the human trabecular meshwork and cornea (Mazlin et al., 2022). That instrument combines en face time-domain full-field OCT (TD-FF-OCT) with cross-sectional spectral-domain OCT (SD-OCT). TD-FF-OCT uses an 850 nm LED with $2\times2$2 nm, NA $2\times2$3, lateral resolution 1.7 $2\times2$4m, axial resolution 7.7 $2\times2$5m in cornea, and a $2\times2$6 mm field. SD-OCT uses a 940 nm superluminescent diode with $2\times2$7 nm, lateral resolution approximately 10 $2\times2$8m, axial resolution approximately 7 $2\times2$9m, and a B-scan field of approximately $1\times1$0 mm lateral $1\times1$1 mm axial (Mazlin et al., 2022). The full device footprint is $1\times1$2 cm.

Its principal EyeNexus relevance is that it resolves uveal meshwork pores, corneoscleral meshwork pores, trabecular beams, internal elastic fibers, ciliary body tendons, and trabecular cell nuclei in vivo (Mazlin et al., 2022). Reported in vivo measurements include uveal pore sizes of approximately $1\times1$3 by $1\times1$4, corneoscleral pore sizes of approximately $1\times1$5 by $1\times1$6, beams approximately $1\times1$7 thick with approximately $1\times1$8 core fibers and approximately $1\times1$9 spacing, ciliary body tendons approximately $X$0 thick and longer than $X$1, and nuclei-like bright oval spots approximately $X$2 (Mazlin et al., 2022). The paper states that the device is capable of detecting blockage of trabecular pores, directly targeting early glaucoma pathophysiology rather than downstream retinal damage.

Taken together, these hardware-oriented uses of EyeNexus point toward a portable, AI-assisted ophthalmic screening platform rather than a purely software application. This is an interpretation rather than an identity claim: the fundus/refraction system, the orthogonal-view OCT instrument, and the OCT web classifier originate from different works. A plausible implication is that the name EyeNexus has been used to frame an integrated diagnostic stack encompassing screening-grade imaging, automated alignment, and downstream AI analysis.

4. Near-eye optical architectures and augmented-reality interpretations

One EyeNexus interpretation centers on a near-eye augmented-reality platform built around a metasurface-enabled contact lens display that projects imagery directly to the retina while leaving the natural scene largely intact (Lan et al., 2019). The underlying demonstration places a passive hologram metasurface on the cornea and illuminates it with a pixelated, see-through display embedded in or beneath the lens. In the cited demonstration, a $X$3 metasurface centered on the pupil sits above a $X$4 pixel array with 50 $X$5m pitch, occupies roughly 1% of a typical pupil area, and is illuminated by a coherent green source at 543 nm (Lan et al., 2019). The optical logic is pixel-by-pixel projection: each input pixel is mapped by a static phase-only hologram to a prescribed retinal location.

The phase-only transmission is modeled as $X$6, and propagation may be analyzed with Fresnel diffraction: $X$7 with $X$8 (Lan et al., 2019). The demonstrated metasurface uses Pancharatnam–Berry geometric phase with silicon nanobeams acting as local half-wave plates, so that the output acquires the phase $X$9, where $f(x)=W_f x$0 is the nanobeam orientation. Phase quantization into eight levels from $f(x)=W_f x$1 to $f(x)=W_f x$2 with step $f(x)=W_f x$3 was reported as sufficient for the demonstration (Lan et al., 2019). The device exploits the fact that the eye cannot focus on structures at the corneal plane, but can focus or otherwise perceive the wavefront-shaped virtual image produced by the metasurface.

The geometric limits are explicit. The eye’s focal length is taken as approximately $f(x)=W_f x$4 mm. The demonstrated retinal patch is about $f(x)=W_f x$5, corresponding to an effective field of view of roughly 17°, and the $f(x)=W_f x$6 pixel layout yields about 1.7° per pixel—described as very coarse but sufficient for icons and alerts (Lan et al., 2019). The same analysis notes that scaling to $f(x)=W_f x$7 pixels would approach about 10 arcmin per pixel. Diffraction efficiency for visible PB-phase metasurfaces is described as commonly achieving 60–90% conversion into the desired cross-polarized order when $f(x)=W_f x$8 and $f(x)=W_f x$9, with a residual zero-order removable by a cross-polarized analyzer (Lan et al., 2019).

A different but related optical route relevant to EyeNexus is the meta-doublet eyepiece for wide-field, large-aperture near-eye optics (Wirth-Singh et al., 2024). That system uses two transmissive metasurfaces, MS1 and MS2, with propagation phase rather than geometric phase, patterned as SiN pillars of height approximately 750 nm and period $g(x)=W_g x$0 nm on quartz (Wirth-Singh et al., 2024). The $g(x)=W_g x$1 cm-aperture configuration has a 21.0 mm entrance aperture, 60° full field of view, 15.0 mm eye relief, 5.4 mm pupil diameter, 15.17 mm effective focal length, NA $g(x)=W_g x$2, and 35.7 mm total track length (Wirth-Singh et al., 2024). Measured normal-incidence transmission is 70% through MS1 alone, 83% through MS2 alone, and 63% through both optics combined. At 20° off-axis, the 2 cm design retains about 64% of on-axis PSF intensity (Wirth-Singh et al., 2024).

These two optical strands imply different EyeNexus display paradigms. The metasurface contact-lens display prioritizes imperceptibility and corneal integration, while the meta-doublet eyepiece prioritizes wide field of view, large aperture, and a conventional eyewear geometry (Lan et al., 2019, Wirth-Singh et al., 2024). The sources themselves do not unify them. A plausible implication is that EyeNexus has been used as a conceptual bridge between corneal-display and eyepiece-based near-eye optics, rather than as a settled optical product class.

5. Attention-aware smart eyewear and personalized auto-capture

Another EyeNexus-associated interpretation derives from MemX, an attention-aware smart eyewear system designed for personalized moment auto-capture (Chang et al., 2021). In that mapping, EyeNexus adopts a temporal visual attention (TVA) network that combines gaze history, eye-movement phase likelihood, and lightweight scene features to decide whether to record a compact video snippet. The proposed sensor stack includes an inward-facing eye camera such as a Sony IMX291 at $g(x)=W_g x$3, a forward-facing world camera such as a Logitech B525 at $g(x)=W_g x$4, and on-device compute on an embedded SoC, prototyped on Jetson Xavier NX and targeting Ambarella CV28 (Chang et al., 2021).

The algorithmic structure is explicit. Historical gaze positions $g(x)=W_g x$5 are converted into $g(x)=W_g x$6 Gaussian heatmaps according to

$g(x)=W_g x$7

and these constitute $g(x)=W_g x$8 (Chang et al., 2021). Scene features $g(x)=W_g x$9 are extracted from $h(x)=W_h x$0 by a shallow MobileNetV2 using the first two blocks to yield a $h(x)=W_h x$1 representation. The network concatenates $h(x)=W_h x$2, $h(x)=W_h x$3, and their element-wise product $h(x)=W_h x$4, applies one convolutional layer Conv0 with 56 output channels, and a fully connected layer FC0 reducing to two logits. Historical eye-movement phase likelihood $h(x)=W_h x$5 is appended to the FC0 input. The decision function is given as

$h(x)=W_h x$6

Training uses YouTube-VIS together with synchronized eye data from 30 participants, a 70/20/10 train/validation/test split, Adam, an initial learning rate of 0.01, 10% decay every 30 epochs until 100 epochs, and frozen MobileNetV2 layers for the first 30 epochs to avoid overfitting (Chang et al., 2021). The baseline eye-tracking-only trigger defines potential attention as the majority of gaze positions falling within a region of area $h(x)=W_h x$7 or smaller during a time period $h(x)=W_h x$8, with $h(x)=W_h x$9. Precision, recall, and average precision are reported using their standard definitions (Chang et al., 2021).

Empirically, the TVA-based method reaches precision 87.50%, recall 86.40%, AP 93.65%, and energy savings 70.61%, while the eye-tracking-alone method reaches precision 40.52%, recall 12.70%, AP 35.93%, and energy savings 76.97% (Chang et al., 2021). The VIS-based method is more accurate at precision 98.08% and recall 90.26%, but yields 0.00% energy savings and runs at about 0.6 FPS on Xavier NX, whereas TVA exceeds 30 FPS (Chang et al., 2021). Pilot studies report 86.36% average energy savings compared with a record-everything baseline, and overall moment-of-interest capture accuracy of 96.05% across 11 in-field pilot studies (Chang et al., 2021).

Within EyeNexus usage, this work is important not because the original paper is titled EyeNexus, but because it supplies a concrete design template for gaze-gated wearable perception. It situates EyeNexus in a broader class of resource-constrained, on-device, attention-aware systems where gaze is used not merely for interaction, but for computational triage.

6. Gaze-driven VR cloud gaming system

The most explicit non-medical instantiation of the term is “EyeNexus: Adaptive Gaze-Driven Quality and Bitrate Streaming for Seamless VR Cloud Gaming Experiences” (Wu et al., 15 Sep 2025). This system addresses a distinct problem from the medical and optical usages: low-latency delivery of cloud-rendered VR frames over variable wireless and mobile networks. Its core contribution is the joint use of gaze-driven spatial compression (FSC) and gaze-driven video encoding (FVE), coupled to a network-aware control loop that adapts the foveation extent in response to delay gradients and feedback timeouts (Wu et al., 15 Sep 2025).

The server side includes a virtual HMD, FSC implemented as a Direct3D 11/HLSL pixel shader, FVE through NVENC H.264 in CONSTQP mode, and a quality/resolution manager that monitors delay gradients and generates macroblock-level QP maps (Wu et al., 15 Sep 2025). The client side includes OpenXR eye tracking at 200 Hz, Android MediaCodec decoding, OpenGL/GLSL foveated spatial decompression, and stereo display on a Meta Quest Pro at a 72 FPS target. The base quantizer is $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$0, macroblocks are $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$1, and gaze is projected from per-eye yaw/pitch using

$a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$2

FSC uses ROI sizes $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$3, $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$4 and compression ratios $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$5, $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$6 (Wu et al., 15 Sep 2025).

Foveated video encoding uses a Gaussian quantization-offset field centered on the gaze macroblock: $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$7 with $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$8, where $a_{i,j}=\mathrm{softmax}(f(x_i)^\top g(x_j))$9 controls the effective foveation radius (Wu et al., 15 Sep 2025). The delay increment per frame is

$299\times299$00

and an AIMD-style controller updates $299\times299$01 with $299\times299$02, $299\times299$03, $299\times299$04, and bounds $299\times299$05 (Wu et al., 15 Sep 2025). Motion-to-photon latency is decomposed as

$299\times299$06

The reported results are substantial. Under WiFi, mean MTP latency is 69.4 ms for EyeNexus versus 77.7 ms for GCC, 78.0 ms for ALVR, 79.3 ms for FovOptix, and 80.9 ms for CGFVE (Wu et al., 15 Sep 2025). Under mobile 5G emulation, EyeNexus reaches 69 ms versus 74 ms for FovOptix, 77 ms for GCC, 98 ms for ALVR, and 237 ms for CGFVE, corresponding to latency reduction up to 70.9% versus CGFVE (Wu et al., 15 Sep 2025). Reported perceptual visual-quality improvement is up to 24.6%, and an IRB-approved user study with 14 participants reports playability and visual-quality improvements of up to 48%, while eliminating reported motion sickness in the study (Wu et al., 15 Sep 2025). EyeNexus also yields the smallest bitrate imbalance under both mobile and WiFi conditions.

This VR streaming EyeNexus is technically unrelated to the OCT triage application except for its reliance on eye tracking and spatially nonuniform processing. Their coexistence under the same name underscores the semantic breadth of the term.

7. Limitations, validation status, and interpretive issues

Across the cited record, EyeNexus-related systems are uneven in maturity. The OCT web application has balanced internal test results and a secure Flask/Firebase deployment, but no external validation cohort, no per-device analysis, no prospective clinical study, no calibration or uncertainty estimation, and no detailed PHI-compliance pathway (Naik et al., 2023). The dual-modality fundus/refraction prototype reports mIoU $299\times299$07, EDE $299\times299$08 px, and refractive MAE below 5%, but also notes limitations from commercial spherical lenses, low 980 nm sensor sensitivity, and the need for expanded calibration and improved stray-light control (Peng et al., 13 Apr 2025). Orthogonal-view OCT demonstrates unique cell-resolution visualization of the trabecular meshwork, yet alignment to the trabecular region is time-consuming and multiple scattering through the sclera limits certain views (Mazlin et al., 2022).

The near-eye optical concepts also remain constrained. The contact-lens metasurface display demonstrates only a $299\times299$09 pixel array over roughly 17° field of view, with coherent-light, alignment, dispersion, speckle, safety, and manufacturability challenges explicitly identified (Lan et al., 2019). The meta-doublet eyepiece achieves a 21.0 mm entrance aperture and 60° full field at 633 nm, but remains narrowband and exhibits angle-dependent throughput reduction, with large-area fabrication constrained by lithographic resolution (Wirth-Singh et al., 2024). MemX-derived smart eyewear reaches greater than 30 FPS and strong energy savings, but still has privacy concerns, motion-compensation requirements, and domain-specific trade-offs in the threshold parameters $299\times299$10 and $299\times299$11 (Chang et al., 2021). The VR streaming EyeNexus, while quantitatively strong, still lists rapid saccades, gaze noise, extreme network jitter, stereo discrepancies, and client-side resource limits as open issues (Wu et al., 15 Sep 2025).

From an encyclopedic standpoint, the principal interpretive issue is ontological rather than algorithmic. EyeNexus is not a singular research object with one provenance, one codebase, or one validation lineage. It is a label attached to at least two fully distinct systems and several additional proposed integrations. This suggests that the term is best understood as a recurring name for eye-centered computational platforms whose unifying motifs are gaze, optical alignment, perceptual prioritization, and clinically or experientially relevant decision support.