Retina: Structure, Function & Applications

• Retina is a laminated neural tissue that converts photons into voltage changes through layered, parallel, nonlinear processing.
• It integrates photoreceptors, bipolar, and ganglion cells to create multiplexed visual computations such as center-surround antagonism and motion sensitivity.
• Advanced imaging methods and biomimetic designs leverage retinal structure to diagnose CNS disorders and guide prosthetic innovations.

The retina is a laminated neural tissue of the central nervous system and, at the same time, an electro-optical device that converts photons into voltage changes and spike trains while performing substantial preprocessing before signals are sent to the brain. Across the cited literature, it is treated not as a passive camera-like sensor but as a layered, parallel, nonlinear processor; as the only part of the CNS that allows direct optical access in vivo; and as a quantitative target for imaging, modeling, organoid engineering, and biomimetic system design (Salbaum et al., 2022, Pfäffle et al., 2018).

1. Anatomical substrate and layered organization

The retina is described as a multilayer circuit composed of photoreceptors, horizontal cells, bipolar cells, amacrine cells, ganglion cells, and glial cells. In the vertebrate organization summarized in the cited work, photoreceptors occupy the outer retina, synaptic interactions between photoreceptors, horizontal cells, and bipolar cells occur in the outer plexiform layer, and synaptic interactions between bipolar cells, amacrine cells, and ganglion cells occur in the inner plexiform layer; ganglion-cell axons then form the optic nerve (Chiquita, 2018, Meister, 28 Sep 2025). The tissue itself is reported as about $150\text{–}300\,\mu m$ thick depending on location and disease state, and it is optically challenging because its neuronal, vascular, and glial layers are mostly transparent and therefore weakly reflecting (Laforest et al., 2017).

This architecture is functionally partitioned into parallel pathways. The cited sources emphasize ON and OFF channels, transient and sustained channels, and multiple ganglion-cell output classes specialized for direction selectivity, object motion sensitivity, looming detection, and related feature channels (Chiquita, 2018). A 2025 synthesis further formalizes this organization as a cell-type-based circuit template in which photoreceptors provide input, bipolar cells define the main parallel subunits, horizontal and amacrine cells shape lateral and inhibitory interactions, and ganglion cells provide the output readout (Meister, 28 Sep 2025). This supports the view that retinal anatomy is not merely descriptive morphology; it is the substrate for a multiplexed computation that begins before any signal reaches thalamus or cortex.

2. Computational principles and receptive-field organization

A recurring claim in the literature is that the retina is a parallel image processor and a nonlinear computational system rather than a simple relay of luminance (Chiquita, 2018). In one explicit formulation, the “standard model” is a multi-channel linear-nonlinear cascade in which each bipolar-cell type applies a spatiotemporal filter $K_i$ to the stimulus $S(\vec x,t)$,

$$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$

and ganglion cells pool rectified subunit outputs with weights $w_{ji}$,

$$G_{j}(\vec x, t) = \sum_i \sum_{\vec x'} w_{ji}\left(\vec x' - \vec x\right)\, N_{ji}\left( B_{i}(\vec x', t)\right),$$

with optional amacrine-cell channels added when explicit inhibition is required (Meister, 28 Sep 2025). Within this framework, parameter changes in kernels, weights, rectification, delays, and inhibition can reproduce center-surround antagonism, nonlinear spatial summation, texture sensitivity, direction selectivity, object motion sensitivity, looming sensitivity, motion anticipation, and spike-latency coding (Meister, 28 Sep 2025).

Efficient-coding accounts and end-to-end deep-network models provide complementary interpretations of the same receptive-field geometry. In a hierarchical efficient-coding model, sparse PCA is used as a retinal/LGN analogue and yields center-surround receptive fields under constraints of reconstruction, sparse connectivity, and equalized activity, while subsequent sparse coding or ICA stages produce oriented V1-like filters (Shan et al., 2013). In an anatomically constrained deep CNN, a reduced number of neurons at the retinal output, interpreted as an optic-nerve bottleneck, is sufficient to make retinal outputs center-surround-like and downstream layers oriented; the same work argues that retinal outputs become more nonlinear and lossy feature detectors when downstream cortex is simple, and more linear and faithful encoders when downstream cortex is more complex (Lindsey et al., 2019). This suggests that the classical opposition between “efficient coding” and “feature extraction” is, in these models, parameter-dependent rather than absolute.

3. Dynamical, predictive, and probabilistic descriptions

The retina has also been analyzed as a high-dimensional, structured, non-autonomous dynamical system driven by non-stationary and spatially inhomogeneous visual scenes (Cessac, 2020). In that description, the outer plexiform layer and inner plexiform layer are treated as coupled dynamical subsystems; receptive fields are represented by spatiotemporal convolutions; retinal waves in development are modeled through starburst amacrine-cell dynamics with bifurcations; and ganglion-cell spike trains are analyzed using generalized linear models, Gibbs measures, and linear-response theory (Cessac, 2020). These approaches place retinal function within nonlinear dynamics, bifurcation theory, and ergodic/statistical theory rather than within a purely static receptive-field formalism.

Experiments on bullfrog retina directly support a predictive component. Using mutual information between a stochastic pulse-train stimulus and retinal firing rates, the cited study found that for sufficiently regular stimuli the peak of $I_m(S,R,\delta t)$ can move from negative to positive $\delta t$, meaning that retinal responses become most informative about future stimulus events rather than only past ones (Chen et al., 2016). The same work reports that predictive information is strongest when stimulus information rate is low enough, that the relevant timescale is roughly $100\text{–}200\,\text{ms}$, and that the retina can distinguish time series generated by an Ornstein–Uhlenbeck process from those generated by a hidden Markovian process (Chen et al., 2016). A plausible implication is that temporal regularity is exploited locally in retinal circuitry rather than delegated entirely to downstream areas.

A more speculative 2025 proposal recasts the retina as a probabilistic measurement device. In that model, photon arrivals are Poisson, activation thresholds fluctuate as $$\theta_i(t) \sim \mathcal{N}(\bar{\theta}, \Delta \alpha^2)$$, and ganglion-cell spikes emerge from stochastic threshold crossings after layered retinal processing. The paper formalizes this with an uncertainty relation,

$K_i$0

where $K_i$1 is threshold variability and $K_i$2 is temporal uncertainty in first-response timing (Taranath et al., 30 Jul 2025). The authors explicitly do not claim that the brain is a quantum computer; they instead propose a classical but quantum-inspired framework in which retinal computation is probabilistic and time-sensitive (Taranath et al., 30 Jul 2025). This remains a proposal rather than a settled consensus, but it exemplifies current attempts to formalize intrinsic retinal variability as computation rather than nuisance noise.

4. In vivo imaging and measurement of retinal structure and function

Because the retina is directly accessible optically, it has become a central site for non-invasive CNS imaging. Phase-sensitive full-field swept-source OCT has been used to measure intrinsic optical signals from photoreceptors and ganglion cells in living human retina, with optical path-length changes extracted from phase shifts according to

$K_i$3

In the cited measurements, cone outer segments elongated by more than $K_i$4, rod outer segments by more than $K_i$5, and ganglion-cell signals reached about $K_i$6 after approximately $K_i$7; the ganglion-cell response was roughly ten-fold smaller than the photoreceptor response and required sub-pixel co-registration, correction of lateral and axial motion, dynamic phase referencing, background subtraction, and averaging to become visible (Pfäffle et al., 2018). The same work reconstructed a functional wiring diagram by showing that ganglion-cell activation maps are laterally shifted relative to photoreceptor activation maps, with the largest displacement about $K_i$8 at $K_i$9 temporal from the fovea and about $S(\vec x,t)$0 at $S(\vec x,t)$1 from the fovea (Pfäffle et al., 2018).

A complementary family of methods uses transscleral oblique illumination and pupil-based detection to image transparent retinal cells with high phase contrast. One implementation reconstructs quasi-quantitative phase images from differential phase contrast,

$S(\vec x,t)$2

and reports an effective cutoff proportional to $S(\vec x,t)$3, that is, twice the numerical aperture expected for coherent imaging through the pupil alone (Laforest et al., 2017). A related method, transscleral optical phase imaging, imaged retinal cells from the retinal pigment epithelium to the nerve and vascular layers in eleven healthy volunteers without pupil dilation, over a $S(\vec x,t)$4 field of view, within a maximum time of $S(\vec x,t)$5; it reported an average RPE cell area of $S(\vec x,t)$6 and an average RPE density of $S(\vec x,t)$7 between $S(\vec x,t)$8 and $S(\vec x,t)$9 eccentricity (Laforest et al., 2019). Together, these methods show that structural, functional, and cellular-scale retinal imaging now span photoreceptor dynamics, ganglion-cell function, RPE mosaics, vascular-associated cells, and multi-layer morphometry.

5. Retina as a CNS biomarker window

The retina is repeatedly presented as an especially valuable CNS biomarker site because it is anatomically part of the central nervous system, shares neuronal circuitry, a specialized immune response, and the blood-retina barrier, yet can be imaged non-invasively (Pfäffle et al., 2018). The cited literature specifically notes associations between retinal abnormalities and glaucoma, age-related macular degeneration, multiple sclerosis, Parkinson’s disease, and Alzheimer’s disease, with particular emphasis on ganglion-cell involvement and the possibility that functional changes may precede obvious structural degeneration (Pfäffle et al., 2018). This has motivated a shift from purely structural ophthalmic imaging toward multimodal retinal phenotyping and functional assays.

Recent machine-learning work extends this logic to long-horizon risk prediction. REVEAL aligns color fundus photographs with individualized disease-risk narratives and reports prediction of incident AD and incident dementia on average $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$0 and $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$1 years before diagnosis, respectively; for incident AD, the full model with group-aware contrastive learning achieved AUROC $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$2, balanced accuracy $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$3, F1 $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$4, and MCC $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$5, while for incident dementia it reached AUROC $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$6, balanced accuracy $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$7, F1 $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$8, and MCC $$B_{i}(\vec x, t) = \int_{\vec x', t'} K_i\left(\vec x' - \vec x, t' - t\right) S \left( \vec x', t' \right)\, d^2x' \, dt',$$9 (Leem et al., 20 Apr 2026). These results are modest in absolute terms, but they support the claim that retinal morphometry and systemic risk factors contain complementary information.

At the same time, the biomarker literature is not uniformly affirmative. In APP/PS1 mice imaged with multi-contrast OCT, abnormal structural properties and phase-retardation signals were observed, but total, inner, and outer retinal thickness measures showed no statistically significant differences between transgenic and wildtype animals; hyperreflective foci and depolarizing abnormalities were also qualitatively similar in both groups, and retinal $w_{ji}$0 candidates were rare (Harper et al., 2019). This cautions against treating every retinal abnormality as disease-specific. A balanced reading of the current literature is therefore that the retina is a powerful biomarker substrate, but specificity depends strongly on modality, disease context, and the joint modeling of imaging with other clinical variables.

6. Models, organoids, and biomimetic design targets

Retinal theory now spans molecular to organ scales. A multi-scale simulation models phototransduction biochemistry, calcium dynamics, neuronal electrical coupling, and spherical eye geometry within a generalized bi-domain or multi-domain PDE framework, with membrane voltage $w_{ji}$1, non-uniform finite differences in spherical coordinates, and adaptive variable-step BDF2 time integration (Abuelnasr et al., 2023). The same model reproduces desensitization to repeated light stimuli, photoreceptor a-wave behavior, calcium buffering, and an interplay between photoreceptor gap junctions and inner-segment, but not outer-segment, calcium concentration; applications proposed in the paper include analysis of retinal calcium imaging, electroretinograms, visual prosthetics, and ephaptic coupling (Abuelnasr et al., 2023). In parallel, retina organoids provide a stem-cell-derived three-dimensional system in which eye-field patterning, optic-vesicle-like morphogenesis, layered organization, and the emergence of light-sensitive, synaptically connected photoreceptors can be studied experimentally (Salbaum et al., 2022). Mechanical biophysics is already quantifiable in these systems: AFM measurements cited in the review reported RPE stiffness around $w_{ji}$2 versus about $w_{ji}$3 for neural retina in mouse retinal organoids (Salbaum et al., 2022).

The retina also serves as a template for engineered computation and as a perceptual benchmark. A low-light image-restoration network explicitly modeling cone, rod, horizontal-cell, bipolar-cell, and ganglion-cell interactions uses two depthwise-convolution residual modules, only $w_{ji}$4 learnable parameters, and reports average SSIM $w_{ji}$5 on the LOL dataset, versus $w_{ji}$6 to $w_{ji}$7 for other methods, while emphasizing low computational overhead and biological interpretability (Ming et al., 2022). A separate image-coding framework models the retina as a scalable transform-plus-quantizer-plus-spike encoder and introduces triangular dither noise before leaky integrate-and-fire quantization to whiten reconstruction error and decorrelate it from the input (Masmoudi et al., 2012). In display engineering, the retina becomes an explicit design target: assuming an $w_{ji}$8 pupil diameter and about $w_{ji}$9 million photoreceptor cells, one paper estimates an ideal pixel size of about $$G_{j}(\vec x, t) = \sum_i \sum_{\vec x'} w_{ji}\left(\vec x' - \vec x\right)\, N_{ji}\left( B_{i}(\vec x', t)\right),$$0, corresponding to about $$G_{j}(\vec x, t) = \sum_i \sum_{\vec x'} w_{ji}\left(\vec x' - \vec x\right)\, N_{ji}\left( B_{i}(\vec x', t)\right),$$1 PPI, and then reports a reflective “Retina Electronic Paper” based on WO$$G_{j}(\vec x, t) = \sum_i \sum_{\vec x'} w_{ji}\left(\vec x' - \vec x\right)\, N_{ji}\left( B_{i}(\vec x', t)\right),$$2 nanodisc meta-pixels down to about $$G_{j}(\vec x, t) = \sum_i \sum_{\vec x'} w_{ji}\left(\vec x' - \vec x\right)\, N_{ji}\left( B_{i}(\vec x', t)\right),$$3 in size, $$G_{j}(\vec x, t) = \sum_i \sum_{\vec x'} w_{ji}\left(\vec x' - \vec x\right)\, N_{ji}\left( B_{i}(\vec x', t)\right),$$4 PPI, $$G_{j}(\vec x, t) = \sum_i \sum_{\vec x'} w_{ji}\left(\vec x' - \vec x\right)\, N_{ji}\left( B_{i}(\vec x', t)\right),$$5 frame rate, about $$G_{j}(\vec x, t) = \sum_i \sum_{\vec x'} w_{ji}\left(\vec x' - \vec x\right)\, N_{ji}\left( B_{i}(\vec x', t)\right),$$6 reflectance, and about $$G_{j}(\vec x, t) = \sum_i \sum_{\vec x'} w_{ji}\left(\vec x' - \vec x\right)\, N_{ji}\left( B_{i}(\vec x', t)\right),$$7 optical contrast (Santosa et al., 5 Feb 2025). This suggests that the retina is now used not only as an object of neuroscience, but also as a quantitative limit for near-eye displays and other artificial visual systems.