A retinal "aging clock" is a neural network trained to predict a person's chronological age from an eye scan. The difference between predicted and true age, called the age gap, is then used as a biomarker: people whose eyes look older than their years tend to carry more disease and die sooner. Optical coherence tomography (OCT) produces cross-sectional depth images of the retina from light that reflects back out of the tissue. It is an appealing input for these clocks because it shows the deep layers, and the deepest layer, the choroid (the vascular bed under the retina), appears to carry much of the age signal. There is a problem with this. The light that returns from the choroid is the light that survived the most scattering and absorption, so the choroid is exactly where OCT images are most starved for photons. Older eyes also return less light for reasons that have nothing to do with retinal biology, such as clouded lenses, smaller pupils, and poorer fixation. If a clock learns that "dim, noisy deep layers" means "old," then part of the age gap measures optical signal quality rather than tissue aging.
This note proposes to test that possibility directly and to build the tools that separate the two. The test is an intervention we can control. We lower the simulated photon dose of a scan, which changes only how much light was collected and leaves the eye itself untouched. A clock that reads biology should give the same age before and after. A clock that reads signal quality will make the eye look older. We pair this dose intervention with depth-resolved gradient saliency, meaning heatmaps showing which pixels the prediction depends on, summarized along the depth axis. Together they locate where the dose-sensitive part of the age signal lives. We then ask whether the reported link between age gap and systemic comorbidity survives once signal quality is accounted for.
Core Claim
Our central, falsifiable claim is that current OCT age clocks have a measurable dose sensitivity, and that this sensitivity is concentrated in the deep layers their saliency maps emphasize. We define dose sensitivity for a clock $f$ and scan $X$ as
$$S(X) = \frac{\partial\, \mathbb{E}\left[f(T_d(X))\right]}{\partial \log_2 d}\Big|_{d=1},$$
where $T_d$ is a physical photon-thinning operator that reduces the collected light to a fraction $d$ of the original, and the expectation is over shot-noise realizations. In words, $S$ is the number of years the predicted age shifts per halving of collected light. A clock that reads only biology has $S = 0$. The second part of the claim is that some of the reported association between age gap and comorbidity is carried by per-scan optical quality. That link is significant in the multimodal retinal aging clock of Ludwig et al., with Pearson p = 5.22E-12 against the Charlson Comorbidity Index (CCI), a standard score of chronic disease burden. We can test the second part by partialling out depth-resolved signal measures. The third part is constructive. A clock that takes depth-resolved attenuation coefficients as input, and is trained to be invariant across doses, should reach $S \approx 0$ while keeping the comorbidity association that reflects biology.
What is new is the question and the intervention, not a better regressor. Prior clocks have been evaluated only on standard-dose scans. Prior low-light restoration work has never checked whether restoration preserves a residual biomarker.
Relation to Prior Work
The work that motivates this proposal is Ludwig et al.. They combined fundus photographs and OCT B-scans in an Inception-v4 clock. Using Grad-RAM, a regression version of the gradient-weighted class activation map, they found that the network relies on the deep choroid, the optic nerve head, and the nerve fiber layer rather than the macula. They read this as biology, and that may be right. Their scans, however, were all standard-dose clinical acquisitions, and they never varied collected light. Our proposal takes their clock and saliency method as the object of study.
Yu et al. built a fundus-only clock with MAE 2.79 years on more than 133,000 photographs and linked each additional year of gap to 3% higher mortality. We use it as a reference for how accurate a non-depth-resolved clock can be. It is not evidence that OCT is unnecessary.
Two papers supply methods, and one of them creates a tension we must address. Bellemo et al. translated low-penetration spectral-domain OCT (SD-OCT) into swept-source OCT (SS-OCT) quality with a Pix2Pix generator. They report choroidal thickness correlation 0.97, choroidal vascularity index correlation 0.87, and 47.5% expert Turing-test accuracy. Their dataset of 150,784 registered SD/SS pairs from the Singapore Eye Research Institute (SERI) is the only real, non-simulated "same eye, different deep photon yield" contrast in our file, and we use it as such. Shroff et al. warn that below an information-theoretic photon floor, restoration networks produce plausible structure that is not there. Taken together, a generator that reproduces choroidal thickness well could still invent the fine texture a clock reads as age. We test this rather than assume either answer.
Self-supervised low-light denoising, as in DeepCAD-RT, exploits the fact that noise is independent across repeated frames while structure persists; that work reports imaging with tenfold fewer photons. CARE trains restoration on paired low and high exposures with per-pixel uncertainty. We reuse the DeepCAD-RT idea on adjacent OCT B-scans as a denoise-then-predict baseline.
Automatic A-line classification with attenuation coefficients converts OCT depth profiles into depth-resolved attenuation coefficients and reports plaque-type accuracy of 83.47% versus 74.94% on raw intensities. This is our model for a physics-normalized input. ExPeRT argues that gradient saliency is noisy and unfaithful for age regression, and it constrains how we use saliency: we never report a heatmap without checking it against a direct intervention.
Research Program
The first component is a photon-thinning simulator for OCT. OCT is interferometric: its sensitivity in the shot-noise-limited regime scales with collected sample power, and a lower dose raises the noise floor relative to the depth-decaying signal. We will implement $T_d$ on linear-intensity B-scans. The scan is modeled as an exponentially attenuated signal following the Beer-Lambert law, $I(z) = I_0 e^{-\mu_t z}$, sitting on a noise floor. Thinning scales the signal by $d$ and redraws noise at the correspondingly higher relative level. Where the original scans already sit near their floor, as at the choroid-sclera boundary, thinning pushes them below it first. Clinical exports are processed log-intensity images rather than raw spectra, so the simulator is an approximation. We validate it two ways. First, the SERI SD/SS pairs let us check that thinning an SS-OCT scan reproduces the deep-layer statistics of its SD-OCT partner, including the depth profile of contrast and speckle variance. Second, natural variation in device signal-strength indices within the Stanford Research Repository (STARR) cohort gives a real, uncontrolled dose axis.
The second component is the audit. We will train a clock following Ludwig et al.'s recipe on the STARR multimodal cohort, which has 7,750 OCT B-scans from 2,467 patients with age and CCI codes. For each test scan we will estimate $S(X)$ over a grid of doses. We will also compute two kinds of saliency. The first is Grad-RAM. The second is integrated gradients, which averages gradients along a path from a baseline image to the real one; we average it over many shot-noise draws so that saliency, like the prediction, is an expectation under the physics. We collapse each map into a depth profile, attribution mass per retinal layer band, and track how the profile moves as dose falls. The saliency is then checked against a direct intervention. We thin only one depth band at a time and measure the change in predicted age, which gives an occlusion-style ground truth that the gradient profile must agree with before we interpret it. This addresses the ExPeRT objection with an experiment rather than an argument.
The third component is the remedy. We will train a clock whose input is the depth-resolved attenuation-coefficient map rather than raw intensity. Each training batch contains the same scan at several simulated doses, and a consistency penalty requires their predictions to agree. The attenuation input addresses the physics: signal decay with depth is modeled explicitly rather than learned as a cue. The consistency penalty addresses the dose shortcut directly. The two ingredients are ablated separately.
Baselines
The naive reference is the chronological-age mean predictor together with the unmodified raw-intensity clock evaluated on thinned scans with no mitigation. The mean predictor anchors $R^2 = 0$. The unmodified clock shows how bad things are when nothing is done.
The strongest existing method on this data is the Ludwig et al. clock itself. On STARR it reports test MAE 10.63 years ($R^2 = 0.57$) when trained on normal eyes only, and MAE 6.00 years ($R^2 = 0.86$) when trained on all categories with one-hot disease labels. It falls to $R^2 = -1.74$ on AMD eyes when disease context is withheld. We will reproduce both settings on the same splits before any comparison, and we will report our reproduction next to their figures. Beating this baseline does not mean lower MAE at full dose; we would accept MAE within 0.5 years of the reproduced baseline there. It means a dose-sensitivity magnitude at least five times smaller at half and quarter dose, and a CCI association that survives adjustment for signal quality where the baseline's weakens.
The noise-augmentation baseline trains the same raw-intensity architecture on randomly thinned scans without the consistency penalty or attenuation input. It is the simplest competitor to the remedy. It gets the same dose distribution, the same number of epochs, and the same hyperparameter search budget of twenty configurations as our model.
The denoise-then-predict baselines take two forms. One is a self-supervised denoiser trained on adjacent B-scans in the style of DeepCAD-RT. The other, on SERI only, is the Pix2Pix SD-to-SS translator of Bellemo et al.. Each is followed by the frozen full-dose clock. Classical frame averaging of repeated B-scans serves as the physically honest but photon-expensive reference. It is matched in total photons to the low-dose conditions so that averaging does not get extra light for free.
Two ablations isolate the proposal's ingredients: attenuation input without consistency training, and consistency training on raw intensity. As an external reference point we cite the fundus-only clock of Yu et al. (MAE 2.79 years, UK Biobank). We have not found any reported dose-sensitivity result for any retinal age clock in the literature we could search, so no baseline number exists for $S$.
Why not just…?
Why not just train with noise augmentation? This is the most serious threat. It may well drive $S$ toward zero on its own, because a network shown many doses can learn to ignore brightness. The experiment settles it directly: the augmentation baseline and the attenuation-plus-consistency model are compared on the same budget. If augmentation matches on dose sensitivity, cross-device agreement, and the adjusted CCI association, we drop the architectural claim. The contribution then becomes the audit and the finding that augmentation suffices. There is one reason to expect a gap. Augmentation removes sensitivity to simulated dose, but it may still let the clock read real, age-correlated deep-signal loss from media opacity. We check this through the partial-correlation analysis on unthinned STARR scans, where no simulation is involved.
Why not just denoise or enhance the scan first? A strong restorer might return deep layers to full-dose appearance, and on SERI it can be scored against real SS-OCT. The risk named by Shroff et al. is that restoration fills in typical-for-age texture learned from training data, which would pull everyone's age gap toward the population mean. That would look like good MAE while erasing the biomarker. We measure this in two ways. We compare the variance of the age gap on restored versus real SS-OCT, and we compare the gap's correlation with the real-SS-OCT gap for the same eye. High choroidal-thickness agreement, like Bellemo's 0.97, does not answer the question, because thickness is a coarse summary.
Why not just use fundus photographs or buy swept-source hardware? A fundus clock is cheaper and, at 2.79 years MAE (Yu et al.), more accurate on its home cohort. But fundus imaging also depends on light passing through the lens, so it is not immune to the same confound. We will run a fundus clock on STARR and apply the same partial-correlation audit. Swept-source hardware reduces the deep-photon deficit but does not remove it, and it is not what most clinics own. The SERI pairs directly test whether SD- and SS-derived ages of the same eye agree.
Why not just use prototype-based explanations instead of gradients? ExPeRT avoids gradient noise, but it explains a different, specially built model rather than the deployed clocks we are auditing. We use gradients only as a depth-profile summary, and only after the band-thinning interventions confirm them. If gradients and interventions disagree, we report the interventions.
Decisive Experiments
The first check is whether current clocks are dose-sensitive at all. On STARR, the reproduced Ludwig-style clock should show $S > 0$, with thinning the choroid band producing a larger age shift than thinning superficial bands. If $S$ is indistinguishable from zero with confidence intervals under half a year per halving, the central worry is refuted and we report that clocks are robust. That would be a useful negative result.
The second check uses real devices. On the SERI SD/SS pairs, the within-eye difference between SD-derived and SS-derived predicted age should be systematically positive, meaning less deep light yields an older-looking eye, for the raw clock and near zero for the remedy. This depends on SERI exposing age metadata, which the file does not confirm. If age is unavailable, the within-eye difference in predicted age still tests invariance without needing labels.
The third check is whether the biomarker survives. Among STARR normal eyes, we regress CCI on age gap with and without per-scan covariates: deep-to-superficial signal ratio, fitted choroidal attenuation, and device signal strength. We then report how much of the association remains.
The fourth check is the hallucination audit. For restored and translated scans, we report age-gap variance shrinkage and same-eye gap correlation against real SS-OCT.
Success means the remedy holds full-dose MAE near the reproduced baseline, reduces $|S|$ fivefold, closes the SD/SS gap, and keeps a CCI association that no longer depends on signal-quality covariates. Any of these could fail independently, and each failure is informative about which part of the retinal age gap is biology and which is light.
References
- A multimodal retinal aging clock for biological age prediction and systemic health assessment via OCT and fundus imaging
- A cross population study of retinal aging biomarkers with longitudinal pre-training and label distribution learning
- Optical coherence tomography choroidal enhancement using generative deep learning
- Live-cell imaging powered by computation
- Real-time denoising enables high-sensitivity fluorescence time-lapse imaging beyond the shot-noise limit
- Content-aware image restoration: pushing the limits of fluorescence microscopy
- Automatic Classification of A-Lines in Intravascular OCT Images Using Deep Learning and Estimation of Attenuation Coefficients
- Prototype Learning for Explainable Brain Age Prediction