Brain MRI only produces usable images when the person lies still for many minutes. People who cannot do this are dropped from the data: young children, autistic children, people with movement disorders, frail or cognitively impaired older adults. A growing set of models predicts brain blood-vessel damage from a photograph of the back of the eye, the retina, which takes about a second to capture. These models learn their brain labels from MRI, so they learn only from people who managed to stay still. Those are likely the people with the least brain disease.
This proposal treats that stillness requirement as a labeling bias inside the retina-to-brain model, and builds a correction that does not need MRI from the excluded group. The correction works because three things are available for everyone, including people whose scans were rejected: a retinal photo, clinical traits, and later recorded outcomes such as stroke.
Core Claim
Retinal models of cerebral small vessel disease (SVD), meaning damage to the brain's tiniest vessels that shows on MRI as white-matter bright spots and enlarged fluid spaces around vessels, are trained on a sample filtered by whether the person could stay still. We claim three things:
- This filter is not random. Motion intolerance tracks the same frailty, cognitive decline and motor impairment that accompany SVD. The labels are therefore "missing not at random," and a model trained only on complete cases will under-predict disease burden and be miscalibrated in exactly the people who cannot be scanned.
- Because the retinal image exists for everyone, the probability of yielding a usable scan can be estimated from retina and clinical traits and used to reweight training.
- The remaining bias, which depends on the unseen brain state itself, can be bounded with outcomes that never required stillness, such as incident stroke.
What is new is not a better retinal encoder. The new part is the observation that the oculomic literature inherits neuroimaging's head-fixation problem through its labels, plus a correction and an evaluation protocol designed around that fact.
Where This Sits
The motivating evidence comes from the neuroimaging side. Pediatric and autism cohorts show high scan rejection and need slow mock-scanner habituation (Papadelis et al.; Chen et al.). ALS studies fall back on resting-state protocols because motor tasks are infeasible in the scanner (Luo et al.). Primate work often trades motion for anesthesia, which changes the very physiology being measured (Milham et al.). None of these papers asks what happens downstream when such filtered cohorts become label sources for other models; that is our question.
The method comes from the retinal side. RETFound is a masked-autoencoder foundation model pretrained on 1.6 million retinal images that transfers to systemic-disease prediction. DeepRETStroke uses MRI-confirmed silent-infarct soft labels to detect silent brain infarction from fundus photos and to predict stroke. The OCTA small vessel disease study links retinal arteriolar density to periventricular white-matter lesions in 246 older adults (r = −0.17). We reuse these encoders and labels. The difference is that all three treat MRI-labeled participants as the population, while we treat them as a selected subsample of it.
Singh et al. show that retinal imaging works even in premature infants. This supports the premise that the retinal input is far less selective than the MRI label.
Two papers constrain the design. The dynamic OCTA neurovascular-coupling study shows that functional retinal tests catch disease earlier than static images. However, its protocol needs 45 minutes of dark adaptation and repeated chin-rest imaging, so dynamic retinal testing would reintroduce a stillness filter. We therefore restrict the input to single static captures. Similarly, fNIRS decoding and HD-DOT escape head-fixation but measure only the outer cortex, so they cannot supply deep white-matter labels.
Research Program
The work has three components.
Audit. In cohorts where participants were invited to MRI and their outcome was logged (usable scan, motion-degraded scan, aborted, declined), we measure how much the retina–SVD relationship and model calibration shift as the motion-exclusion threshold tightens. The threshold is set by a structural image-quality or motion index.
Correction. Let $x$ be the retinal image, $z$ the clinical traits, $y$ the MRI-derived SVD burden, and $R$ an indicator that a usable scan exists. We fit a propensity model $\pi(x,z) = P(R=1 \mid x,z)$ on everyone invited. The retinal predictor $f$ is then trained with the weighted loss
$$\mathcal{L}(f) = \sum_{i:\,R_i=1} \frac{\ell\big(f(x_i), y_i\big)}{\pi(x_i,z_i)},$$
where $\ell$ is the regression or soft-label loss. Motion-degraded but retained scans enter through a noise model: label variance grows with the motion index rather than the scan being discarded. This weighting removes bias only if selection depends on $x$ and $z$ alone.
Outcome anchoring. For the remaining dependence on unseen $y$, we use incident stroke and dementia, which are recorded regardless of scan tolerance. If the corrected model's predicted burden has the same hazard relationship with stroke in the never-scanned group as in the scanned group, the correction is behaving. A divergence quantifies what the propensity model misses. We report sensitivity bounds that vary how strongly selection depends on $y$ beyond $x$ and $z$.
Baselines
Naive reference. The do-nothing reference is a clinical-traits model (age, blood pressure, glycemia, lipids). It costs nothing extra, and on silent infarction it reached AUC 0.633 in DeepRETStroke. It is fit on the same complete-case training set, and its calibration is reported in the excluded stratum.
Strongest existing method. This is a DeepRETStroke/RETFound-style foundation model fine-tuned on complete cases only, which is current practice. DeepRETStroke reports silent-brain-infarction AUC 0.797 against the 0.633 clinical baseline. We initialize from public RETFound weights, fine-tune on the same split with the same hyperparameter-search budget as our method, and evaluate on the same strata. For whole-burden white-matter regression in the excluded group, we found no reported result on Lacuna.
Simplest fix. "Include everyone, add motion as a covariate" retains degraded scans and conditions on the motion index. It gets the same encoder and tuning budget.
Ablations. Two ablations isolate our ingredients: weighting without the motion-dependent noise model, and noise model without weighting.
What counts as beating the baselines. In the excluded and high-motion strata, the method must show lower calibration error and less under-prediction of burden. It must also maintain a closer stroke-hazard ratio across scanned and unscanned groups. And it must not lose more than a pre-registered small margin of AUC in the clean stratum.
Why not just…?
Why not just use motion-robust or faster MRI to scan the excluded people? This is the correct long-run fix and a real threat to the premise. Prospective motion correction and sedation-free fast protocols would shrink the excluded group. However, they do nothing for the large existing cohorts that already trained today's retinal models. Sedation also alters physiology, a problem Milham et al. document in primates. Where any such rescue scans exist, we use them as the best available ground truth for the excluded stratum rather than competing with them.
Why not just add the motion index as a covariate and keep all scans? It handles degraded labels but not absent ones. People who declined or aborted the scan have no label to condition on. The covariate baseline is run directly, and if it matches the full method in the outcome-anchored test, we will report that the simple fix suffices.
Why not just rely on clinical traits, which every excluded person already has? DeepRETStroke shows retinal models clearly beat traits on silent infarction (0.797 vs 0.633). The open question is whether that advantage survives in the excluded group. The traits baseline is evaluated stratum by stratum to answer it.
Why not use wearable brain imaging to label the excluded group? fNIRS has limited depth and is vulnerable to scalp blood flow (Cao et al.). It cannot see periventricular or deep white matter, where the OCTA study locates the retinal associations.
Proposed Experiments
Semi-synthetic test. The first decisive check takes clean-scan participants with known burden and imposes a synthetic exclusion rule whose probability rises with true burden and with frailty markers. Every method is trained on the survivors and scored on the dropped participants, where ground truth is known. This shows whether weighting recovers calibration and how its error grows as selection depends more on $y$ itself.
Real cohorts. The second check uses cohorts pairing fundus or OCT with brain MRI and linked outcomes, with invitation and scan-quality logs. Candidates are the UK Biobank imaging subsample used in RETFound's evaluations and, if access allows, the DeepRETStroke multi-country cohorts and the West China SVD cohort. Results are stratified by scan status. The 246-person OCTA cohort serves mainly as a check that retinal arteriolar density still tracks periventricular lesions after reweighting.
Outcome-anchored test. The third check compares the predicted-burden-to-stroke hazard ratio between scanned and never-scanned participants, under each method.
Interpretation. The idea fails if complete-case models are already well calibrated in high-motion and excluded strata. That would mean stillness selection is ignorable for oculomics, a useful negative result worth publishing. The idea succeeds if the bias is present and the correction shrinks it on the outcome test without harming the clean stratum.
References
- Editorial on emerging neuroimaging tools for studying normal and abnormal human brain development
- The Development of a Practical Artificial Intelligence Tool for Diagnosing and Evaluating Autism Spectrum Disorder: Multicenter Study
- Patterns of Spontaneous Brain Activity in Amyotrophic Lateral Sclerosis: A Resting-State fMRI Study
- Accelerating the Evolution of Nonhuman Primate Neuroimaging
- A foundation model for generalizable disease detection from retinal images
- A deep learning system for detecting silent brain infarction and predicting stroke risk
- Lower Retinal Arteriolar Density Is Associated With Higher Cerebral Small Vessel Disease Burden: An Optical Coherence Tomography Angiography Study
- Deep learning-based prediction of cardiopulmonary disease in retinal images of premature infants
- Reversed Neurovascular Coupling on Optical Coherence Tomography Angiography Is the Earliest Detectable Abnormality before Clinical Diabetic Retinopathy
- Brain Decoding Using fNIRS
- Mapping brain function underlying naturalistic motor observation and imitation using high-density diffuse optical tomography