Adaptive optics cameras for the eye measure and cancel the eye's optical blur. This lets clinicians see single capillaries and vessel walls about 2 µm across, which are the same calibre of vessel that fails in cerebral small vessel disease. These instruments correct only the phase of light, meaning how far each part of the wavefront lags behind the rest. The eye also changes light's polarization in ways that vary across the pupil. The cornea and lens are birefringent, meaning they delay one polarization relative to the other, and the macula partly absorbs one polarization, an effect called diattenuation. Vectorial adaptive optics corrects phase and polarization together, using stacked liquid-crystal spatial light modulators (SLMs, pixelated devices that shift phase locally) and a deformable mirror. On the bench it has recovered structure that phase-only correction leaves hidden. It has never been used in a living eye. One reason is that its controller searches through many modulator settings by trial and error, and each setting costs one modulator refresh. Meanwhile the eye moves on roughly 10 ms timescales.
We propose to borrow the trick that made radio propagation modelling fast. That field trains a network offline on expensive physics simulations so that, at run time, a few measurements map directly to the answer. Here the network is trained on a simulator of polarized light passing into the eye and back out, through cornea, lens and retina. At run time it turns a handful of probe frames, taken at deliberately chosen modulator settings, into a full phase-plus-polarization correction. It also reports whether the eye's diattenuation has pushed the correction past what is physically recoverable. The claim is narrow and testable. In a live eye the delay in vectorial correction comes from the number of measurements, not from computing. A network trained on simulated eyes with a structured output can reach the vectorial precision of a brute-force scan with an order of magnitude fewer probe frames. That would make polarization-corrected imaging of retinal microvessels practical. The retinal-cerebral link sets the target: reproducible measurements of vessel walls and capillaries. It is not something this project will itself prove.
Where this sits relative to prior work
Three papers supply the hardware and the metrics. Active compensation of extrinsic polarization errors using adaptive optics corrected spatially varying polarization errors using an imaging polarimeter and an SLM. That works when the polarimeter can see the distorted beam directly. In the eye it cannot, because the only accessible light has crossed the anterior eye twice and has been partly depolarized by the retina. Vectorial adaptive optics for advanced imaging systems built a wide-field vectorial correction module from two SLMs and a deformable mirror, and defined two scores. Vectorial precision $P$ measures how close the corrected polarization state is to the target. Vectorial uniformity $U$ measures how consistent that state is across the field. In air, precision went from under 1% to over 95% and uniformity from 77.1% to 91.2%. However, it optimizes by a coarse scan of SLM grey levels while watching pupil intensity behind an analyzer, and it was tested only on air, a wave plate and a crystal. We reuse its module and its $P$ and $U$ scores, and replace its scan. A reconfigurable arbitrary retarder array as complex structured matter showed that cascading three SLMs and a deformable mirror gives four controllable quantities per virtual pixel. That flexibility is exactly why blind search grows so expensive.
Using optical skyrmions to assess vectorial adaptive optics capabilities in the presence of complex aberrations constrains what we can promise. Its probes are skyrmion beams, whose polarization pattern covers every possible polarization state. It shows that retardance is reversible, but strong diattenuation pushes signal below the detector noise floor. Beyond that boundary no correction, applied before or after the aberration, recovers the lost information. We therefore build a correctability output into the estimator rather than claiming universal correction.
The method is taken from radio-map surrogates. RadioUNet trained a U-Net on ray-tracing output and predicted pathloss maps in 1–10 ms instead of about 100 s. Its sparse-measurement variant shows how a few real measurements can calibrate a simulation-trained network. The First Pathloss Radio Map Prediction Challenge found that every surrogate lost accuracy on an unseen city. We take that as a warning that a network trained on simulated eyes will fail on eyes outside its training range. We therefore hold out eye types explicitly, as the challenge held out Istanbul. EM-GANSim shows that adding physics losses keeps such surrogates physically consistent. We use a stronger version of that idea by building the physics into the network's output.
The closest learned-correction work is Coordinate-based neural representations for computational adaptive optics in widefield microscopy (CoCoA). For each image stack, it solves jointly for the sample and a phase-only aberration, and it has worked in vivo in the mouse brain. CoCoA differs from our proposal in three ways. It assumes scalar light, it runs a per-sample optimization offline, and it only reconstructs images in software. Our method is vectorial, amortized into a single forward pass, and drives hardware in real time.
On the retinal side, AOSLO-net and Deep-Learning Image Stabilization for Adaptive Optics Ophthalmoscopy show what phase-only adaptive optics ophthalmoscopes already resolve and how quickly the eye moves. The second paper motivates the 10 ms budget. DeepRETStroke and RETFound show that fundus photographs at 10–15 µm resolution already carry cerebrovascular signal. These papers motivate wanting cleaner microvascular measurements. They also caution against claiming that higher resolution adds predictive value, since we will not test that here.
Core Claim
We model the single-pass pupil of the eye as a field of 2×2 Jones matrices, the standard description of how a medium changes the amplitude, phase and polarization of light. At pupil coordinate $u$ we factor it as
$$J(u) = e^{i\phi(u)}\, D(u)\, R(u),$$
where $\phi(u)$ is the scalar phase, $R(u)$ is a reversible retarder (a polarization rotation with no energy loss), and $D(u)$ is a diattenuator (polarization-dependent loss). The light we detect has passed this pupil twice. On the way back, the retina has depolarized it by an unknown fraction $\rho$.
Our estimator receives $K$ probe frames. Each frame is a pupil- or image-plane intensity recorded behind an analyzer while the modulators apply a known probe state. The estimator outputs low-order coefficients for $\phi$, $R$ and $D$ in a smooth basis, and a correctability score. The correction applied is the inverse of $\phi$ and $R$, plus whatever partial compensation of $D$ is possible.
We claim three things. The latency is set by $K$, because each liquid-crystal refresh takes milliseconds while a forward pass is negligible. A network trained in simulation, whose output is constrained to this factorization, reaches the precision of a coarse scan at much smaller $K$. And the factorized output, not the network's capacity, is what keeps it accurate on held-out eye types.
Research Program
The first component is a simulator of polarized light making the double pass through the eye, playing the role that ray tracing plays for RadioMapSeer. Corneal retardance is drawn from lamellar models with an optic axis that varies across the pupil. Lens birefringence is scaled to mimic ageing and nuclear cataract. Henle-fibre retardance and macular diattenuation are added at the retina, together with a depolarization fraction and a scalar phase drawn from ocular Zernike statistics. Each eye is propagated through the actual modulator stack of the vectorial adaptive optics module. The simulated detector has realistic shot and read noise and a dynamic-range threshold matching the skyrmion paper's intensity-filtered analysis. This lets us label the correctability boundary in simulation.
The second component is the probe design and the estimator. We optimize the $K$ probe states jointly with a convolutional-transformer encoder, similar in shape to the radio surrogates, that maps $K$ frames to factorized coefficients. Training minimizes the predicted loss of vectorial precision after correction, rather than errors in the coefficients themselves. A separate output head predicts whether the post-correction intensity-filtered skyrmion number would stay above a threshold. Training includes sweeps over $K$ from 3 to 32.
The third component closes the loop. The estimator drives the module, and residual drift is tracked by re-estimating every few frames. Image registration uses a cheap correlation tracker, because the UECO stabilization paper shows that deep feature trackers are far too slow for live use.
Baselines
The naive reference is phase-only adaptive optics, with the deformable mirror corrected and the SLMs left flat. This is how current AOSLO systems behave. It tells us how much residual polarization aberration a real eye leaves behind. If that residual is small, the project should stop.
The strongest existing method is the coarse-scan sensorless optimizer of Vectorial adaptive optics for advanced imaging systems. It reported precision above 95% and uniformity rising from 77.1% to 91.2% in air. We found no reported result on Lacuna for its loop time or probe count, or for any biological or double-pass setting. We will therefore run it ourselves on the same simulated eyes, the same bench model eye and the same subjects, with the same noise, the same modulators, and its scan resolution tuned for the best precision. We count the probe frames and wall-clock time it uses. Beating it means matching its final precision and uniformity to within two percentage points while using at least ten times fewer probe frames. Alternatively, it means exceeding its precision at equal frame count.
A second strong competitor is a model-based modal sensorless method built without learning. It is the vectorial analogue of the classical quadratic-fit modal approach. Each low-order phase, retardance and diattenuation mode is biased up and down, and a quadratic is fitted to a transmission metric to find the optimum. It receives the same mode basis and the same probe budget as our network. This baseline isolates whether learning adds anything beyond choosing a good basis.
As an offline accuracy ceiling, we extend CoCoA's per-sample optimization to Jones pupils, giving it the same forward model as our simulator and unlimited iterations. We do not expect to beat it on accuracy. We expect to approach it at a small fraction of its time.
Ablations remove the factorized output and let the network predict unconstrained SLM maps directly. They also drop the diattenuation term, train on single-pass instead of double-pass physics, and replace the learned probes with random ones. If the unconstrained network generalizes to held-out eyes as well as the factorized one, our explanation for robustness is wrong.
Why not just…?
Why not just use the phase-only adaptive optics ophthalmoscopes that already resolve capillaries? This is the most serious objection, because AOSLO-net already segments microaneurysms at a Dice score of 0.7816 on phase-only images. The only evidence that polarization errors hide structure after phase correction comes from inanimate targets in the vectorial adaptive optics paper. Whether the human cornea and lens leave enough spatially varying retardance across a 6–7 mm pupil to matter is untested. The first experiment below measures this directly, and the proposal stops if the effect is small.
Why not just use the coarse scan or a modal quadratic search, with no learning? For a few modes in air this could genuinely suffice. The modal baseline is included to settle the question. The reason to expect learning to help is the double pass and the depolarization. These mix the inbound and outbound polarization effects and flatten the transmission metric, which breaks the assumption of one smooth quadratic peak per mode. If the modal method matches our network at equal $K$, we will report that the contribution is the eye-specific mode basis and the simulator, not the network.
Why not just measure the pupil polarization directly with an imaging polarimeter and invert the Jones matrices in closed form? That method, from the extrinsic-compensation paper, needs access to the distorted beam on one side of the medium. In the eye, only light that has crossed the anterior segment twice and been partly depolarized by the retina is available. Closed-form inversion of that light has no unique answer without a model. We include a polarimeter-plus-inversion variant on the bench model eye to show where it breaks.
Why not just deconvolve polarization effects in software after capture? The skyrmion study shows that diattenuated states pushed below the noise floor cannot be recovered after the fact. Software-only methods such as CoCoA are also scalar and run offline. Our offline vectorial CoCoA ceiling quantifies how much correction applied in hardware before detection gains over this route.
Why not just keep using fundus photographs, which already predict silent brain infarction? DeepRETStroke reports an AUROC of 0.797 for MRI-confirmed silent infarction from fundus images. We do not claim to beat it. Our outcome is the repeatability of capillary and vessel-wall measurements, which fundus photographs cannot resolve. Any predictive comparison is left to later cohort studies.
Proposed Experiments
The decisive first check takes place in a physical model eye and in 20 to 30 adults spanning ages 25 to 80, including people with early lens opacity. A pupil-plane polarimetric camera is added to the detection path of an AOSLO, alongside the vectorial module. We measure $P$ and $U$ after phase-only correction. The go criterion is a median residual precision loss above 10 percentage points in older eyes, together with a visible change in vessel-wall contrast in at least some subjects. Below that threshold, polarization correction is not worth its hardware cost in the eye.
In simulation, we train on young-to-middle-aged eye parameters and test on held-out older and cataract-like parameter ranges, mirroring the unseen-city split of the radio challenge. We report precision, uniformity, probe count, wall-clock time and correctability-flag calibration for every baseline and ablation. On the bench model eye, built from real birefringent and diattenuating elements, we measure closed-loop update rate against the 10 ms budget. That budget is set by the 100 Hz motion reported in the UECO stabilization paper and by the 1–10 ms inference range of RadioUNet. We separate SLM settle time from compute time.
In vivo, in the go-cohort subjects, we image parafoveal arterioles and capillaries under phase-only and vectorial correction in two sessions a week apart. The primary outcomes are the test-retest intraclass correlation of lumen calibre and wall thickness, and vessel-wall contrast. AOSLO-net-style segmentation is retrained on each condition, and its Dice score against expert tracing is reported relative to the published 0.7816. The result that would justify the retinal-cerebral motivation is markedly better repeatability of wall measurements in older eyes under vectorial correction. Those are the subjects in whom fundus-based cerebrovascular models are most used and where media aberrations are worst.
References
- Active compensation of extrinsic polarization errors using adaptive optics
- Vectorial adaptive optics for advanced imaging systems
- A reconfigurable arbitrary retarder array as complex structured matter
- Using optical skyrmions to assess vectorial adaptive optics capabilities in the presence of complex aberrations
- RadioUNet: Fast Radio Map Estimation with Convolutional Neural Networks
- The First Pathloss Radio Map Prediction Challenge
- EM-GANSim: Real-time and Accurate EM Simulation Using Conditional GANs for 3D Indoor Scenes
- Coordinate-based neural representations for computational adaptive optics in widefield microscopy
- AOSLO-net: A deep learning-based method for automatic segmentation of retinal microaneurysms from adaptive optics scanning laser ophthalmoscope images
- Deep-Learning Image Stabilization for Adaptive Optics Ophthalmoscopy
- A deep learning system for detecting silent brain infarction and predicting stroke risk
- A foundation model for generalizable disease detection from retinal images