Can the Retina Tell Us What Brain Free Water Means?

Diffusion MRI measures how freely water moves inside brain tissue. One of its most useful outputs, the "free-water fraction", estimates how much of each voxel's signal comes from water that moves without obstruction, outside cells. In older adults this fraction rises in white matter tracts, the brain's cabling, and it tracks cognitive decline. Nobody can say what the rise means biologically. It could come from inflammation, from tissue loss, or from fluid leaking out of damaged small blood vessels. The authors of the strongest study of this signal say plainly that diffusion MRI alone cannot tell these apart (Free-water metrics in medial temporal lobe white matter tract projections relate to longitudinal cognitive decline).

The retina offers a second, independent view. It is embryologically part of the brain, and it is the only place where small blood vessels and unmyelinated nerve fibers can be photographed directly. We propose using retinal photographs as an outside witness, with one constraint: the photograph is split into two regions, the vessels and the surrounding neural tissue. We then ask whether features taken only from the vessels predict brain free water specifically. They should do so more strongly than they predict measures of nerve-fiber density. Features from the neural tissue should show the reverse pattern. If this crossed pattern appears, it is evidence that part of the free-water signal is microvascular. The test is statistical, can be pre-registered, and runs on data that already exist.

Core Claim

A population-level test can partly resolve the non-specificity of diffusion free-water metrics. The test asks whether they share variance with a directly imaged microvascular bed, the retinal vessels, but not with a directly imaged neural bed, the retinal ganglion-cell and nerve-fiber layers. The two retinal beds must be separated by segmentation, meaning pixel-level delineation of vessels versus background. Whole-image retinal models mix the two: Effects of Hypertension, Diabetes, and Smoking on Age and Sex Prediction from Retinal Fundus Images showed that the non-vessel background alone carries strong age and sex information (sex AUC above 0.96 after vessels were digitally removed). A whole-fundus embedding correlating with brain free water would therefore say nothing about vessels.

The contribution is an estimand, not a better predictor. The estimand is a cross-validated incremental variance explained, measured separately for each retinal region paired with each diffusion compartment. It is estimated under pre-registered covariates and held-out acquisition sites. The prediction that could falsify the idea is a double dissociation. Retinal vessel features should explain more variance in tract free water than in tissue-specific anisotropy or neurite density. Retinal neural features, especially thickness of the ganglion cell–inner plexiform layer (GCIPL), should explain more variance in visual-pathway neurite and anisotropy measures than in free water.

Relation to Prior Work

Two studies motivate the gap.

Archer et al. separated free water from tissue anisotropy in 319 older adults. They showed that tract free water reflects baseline cognition, but they could not attribute it to a cause.

Eye-brain connections revealed by multimodal retinal and brain imaging genetics in the UK Biobank paired retinal imaging with brain MRI in about 6,000 people and linked GCIPL thickness to posterior thalamic radiation integrity. Its white matter measures were single-tensor quantities such as fractional anisotropy (FA) and mean diffusivity. Those measures cannot separate edema from axon loss, as documented by Favre et al. and Alexander's feasibility study. Its fundus features were whole-image CNN principal components. We keep its cohort and replace both ends. On the brain side we use compartment-resolved outcomes. On the retinal side we use predictors restricted to one region.

Three lines of work constrain the design.

  • Vessel-only proxies without dissociation. DeepRETStroke (A deep learning system for detecting silent brain infarction and predicting stroke risk) and Lu et al. show that whole retinal photographs predict cerebrovascular outcomes. Lu et al. use Mendelian randomization to argue that retinal microvascular change causally precedes stroke. Both treat brain injury as a binary event, and neither isolates vessels. We take the vascular causal story as a prior, but test it against diffuse microstructure rather than infarcts.
  • Compartment models in cerebrovascular disease. Shi et al. found that the isotropic fraction from Neurite Orientation Dispersion and Density Imaging (NODDI), a three-compartment diffusion model, detected chronic ischemic damage in moyamoya disease better than mean diffusivity. They had only 15 patients and no peripheral vascular measurement. This supports the use of compartment metrics as outcomes.
  • Masking methods. We reuse soft masking rather than hard cropping. Bai's mask-weighted global average pooling applies a smoothed organ mask as weights before pooling, with gains of 0.9% sensitivity and 2% specificity over unweighted pooling. Lesion-centric latent phenotypes applies the same idea inside a segmentation encoder. Gare et al. show that unmasked models latch onto irrelevant anatomy, with AUC rising from 0.6553 to 0.7742 after masking. The field also warns that binary masks discard the tissue just outside the boundary. We therefore treat perivascular band width as an explicit experimental variable instead of fixing it.

The reproducibility literature supplies the evaluation discipline.

  • The screening mammography meta-repository models containerized, multi-dataset evaluation with bootstrap intervals.
  • The "Dice paradox" of Shakya et al. shows that overlap scores can hide failures on thin structures. This matters for vessel masks, which are nearly all thin structure.
  • MEDS-DEV shows how fixed task definitions let other institutions rerun an analysis without pooling data.

Research Program

Outcomes. On the brain side, we fit a free-water elimination model and a NODDI model to the UK Biobank multi-shell diffusion data. We extract tract-averaged free water, tissue FA, neurite density index (NDI), orientation dispersion index and NODDI isotropic fraction. Tracts are defined with atlas templates of the kind Archer et al. used. The pre-specified primary tracts are the medial temporal projections from that study and the posterior thalamic radiation from the eye-brain study. Brain ROIs stay atlas-based on purpose. The segmentation question in this proposal lives on the retinal side.

Retinal regions. We segment vessels in each fundus photograph and derive three soft weight maps:

  • a vessel map;
  • the vessel map dilated by a perivascular band of width $w$ pixels;
  • the complementary parenchyma map.

Each map weights the final-layer patch features of a frozen retinal foundation encoder, RETFound (Zhou et al.), before pooling. This gives vessel, perivascular and parenchyma embeddings from the same network. A fixed encoder means any difference between regions comes from the mask, not from separate training. OCT supplies GCIPL and nerve-fiber-layer thicknesses as a second, interpretable neural readout.

Mask quality. Following the Dice paradox, we judge vessel segmentation mainly by agreement of derived calibers and branch counts with a reference grading on a manually checked subset. Overlap scores are secondary.

Estimand. For retinal region $r$ and diffusion outcome $y$, let $\Delta R^2_{r,y}$ be the cross-validated gain in variance explained when region-$r$ features are added to a covariate model. Covariates are age, sex, blood pressure, diabetes, smoking, head motion, intracranial volume, white matter hyperintensity volume, imaging site, and the interval between retinal and brain acquisitions. The primary contrast is

$$D = \left(\Delta R^2_{\text{vessel},\,FW} - \Delta R^2_{\text{vessel},\,NDI}\right) - \left(\Delta R^2_{\text{parenchyma},\,FW} - \Delta R^2_{\text{parenchyma},\,NDI}\right).$$

Here $D>0$ means vessel features are relatively specific to free water and parenchymal features relatively specific to neurite density. The same form is used with tissue FA in place of NDI. Intervals come from a subject-level bootstrap. The covariate set, the tract list, $D$ and its sign are pre-registered before any retinal–brain model is fit.

Baselines

Naive reference. The do-nothing reference is the covariate-only model. Its $\Delta R^2$ is zero by construction, and its absolute $R^2$ shows how much of tract free water age and vascular risk factors already explain. Everything else is measured against it.

Classical retinal vessel calibers. The strongest competitor is the classical, non-learned measure: central retinal artery and vein equivalents (CRAE and CRVE) and vessel tortuosity. These come from an automated vessel analyzer of the SIVA-DLS type used by Cheung et al., who found small but dose-dependent caliber changes, for example CRVE of 216.7 versus 215.5 µm, in 10,362 children. Calibers are already vessel-specific and interpretable. They enter the same covariate model, cross-validation folds and bootstrap as the learned features. Our vessel embeddings count as beating them only if their free-water $\Delta R^2$ is higher with a bootstrap interval excluding zero. If calibers match them, we report the dissociation using calibers alone, and the learned component is dropped.

Whole-image foundation embedding. The strongest existing learned approach is the whole-fundus RETFound embedding, pooled without masks. It stands in for the practice of DeepRETStroke and the eye-brain study. For reference, the eye-brain study reports that adding retinal traits to MRI raised fluid intelligence prediction from r = 0.314 to r = 0.344. No reported result was found on Lacuna for predicting diffusion free water, NDI or tissue FA from retinal images, so the whole-image baseline number is produced in this study. Whole-image embeddings use the same frozen encoder and the same ridge regression with the same nested tuning grid as the masked arms.

Isolating the key ingredient. The ablation that isolates region separation swaps the vessel mask for a random mask of equal area and equal spatial smoothness. A band-width sweep with $w$ from zero to a wide perivascular margin tests whether signal lives in the vessels or just outside them. A shuffled-pair control reassigns retinas across participants within age–sex strata and calibrates the null distribution of $D$.

Why not just…?

Why not just correlate standard vessel calibers with free water? This is the real threat, and it is why calibers are a full baseline arm rather than a strawman. Calibers compress the vascular bed into two or three numbers and miss capillary-scale texture. The learned vessel embedding may capture more. If the caliber arm yields the same $D$, that is itself the scientific result, and it is cheaper to replicate. The proposal survives because its claim is the dissociation, not the feature extractor.

Why not just use the whole-image foundation model, which already predicts silent infarcts at AUC 0.797? A whole-image correlation with free water cannot be attributed to vessels. The parenchymal background alone encodes age with $R^2 = 0.92$ in healthy eyes (Kim et al.), and age drives free water. The whole-image arm may predict more total variance. That would not answer the question. The comparison between the whole-image arm and the summed regional arms shows how much of the whole-image signal is attributable to each region.

Why not just use MRI markers of small-vessel disease, such as white matter hyperintensity volume, to disambiguate free water? Those markers come from the same scanner session and share its motion, partial-volume and site confounds. They also detect vascular injury only once it is visible as a lesion. We include hyperintensity volume as a covariate, so any retinal vessel signal we report is beyond it. If the vessel arm adds nothing once hyperintensities are included, we will report that the retina is redundant for this purpose.

Proposed Experiments

The primary analysis uses UK Biobank participants with paired fundus, OCT and multi-shell diffusion data, the roughly 6,000-person overlap described by the eye-brain study. Models are trained on three imaging sites and evaluated on the held-out fourth. All four rotations are reported, and the site-averaged $D$ is the headline number. The pipeline ships as a container with a fixed task specification, in the spirit of the mammography meta-repository and MEDS-DEV. External cohorts can then compute the same $D$ without sharing images. Such replication is a deliberate next step, because no second paired retina–diffusion cohort appears in our sources.

Decisive checks.

  • The pre-registered $D$ must be positive with a bootstrap interval excluding zero on held-out sites.
  • It must exceed the 95th percentile of the shuffled-pair and random-mask null distributions.
  • The reversed contrast must also hold in the posterior thalamic radiation: GCIPL predicts NDI or tissue FA more than free water. Anatomy predicts this direction from the visual pathway.
  • The band-width sweep tells us where the signal sits. A peak at $w>0$ would suggest perivascular rather than luminal involvement, which is relevant to the leakage hypothesis Archer et al. could not test.

Secondary analysis. A secondary analysis in participants with repeat diffusion imaging asks whether baseline vessel features predict change in free water more than change in tissue FA. This mirrors Archer's finding that tissue FA, not free water, tracked memory decline.

If $D$ is indistinguishable from zero under adequate power, that is a useful negative result. It would mean retinal microvasculature does not preferentially index the isotropic compartment, and the free-water signal is more likely inflammatory or degenerative than vascular.

References

  1. Free-water metrics in medial temporal lobe white matter tract projections relate to longitudinal cognitive decline
  2. Effects of Hypertension, Diabetes, and Smoking on Age and Sex Prediction from Retinal Fundus Images
  3. Eye-brain connections revealed by multimodal retinal and brain imaging genetics in the UK Biobank
  4. Widespread white matter microstructural abnormalities in bipolar disorder: evidence from mega- and meta-analyses across 3033 individuals
  5. Axon radius measurements in vivo from diffusion MRI: a feasibility study
  6. A deep learning system for detecting silent brain infarction and predicting stroke risk
  7. Phenotypic screening and genetic insights for predicting major vascular-related diseases using retinal imaging
  8. Neurite orientation dispersion and density imaging reveals microstructural damage in moyamoya disease: a study based on tract- and gray-based spatial statistics
  9. Computed tomography–based COVID–19 triage through a deep neural network using mask–weighted global average pooling
  10. Lesion-Centric Latent Phenotypes from Segmentation Encoders for Breast Ultrasound Interpretability
  11. The Role of Pleura and Adipose in Lung Ultrasound AI
  12. Meta-repository of screening mammography classifiers
  13. Benchmarking Encoder-Decoder Architectures for Biplanar X-ray to 3D Bone Shape Reconstruction
  14. MEDS Decentralized, Extensible Validation (MEDS-DEV) Benchmark: Establishing Reproducibility and Comparability in ML for Health
  15. A foundation model for generalizable disease detection from retinal images
  16. Influence of secondhand smoke exposure on the retinal vasculature of children in Hong Kong