Two older adults can have the same amount of white-matter damage on a brain MRI, visible as bright patches called white matter hyperintensities, and yet one stays sharp while the other declines. This mismatch between how damaged the brain looks and how well it works is a long-standing problem. MRI measures of lesion load explain only part of cognitive performance. We propose testing one specific explanation: the state of the small blood vessels feeding the damaged tissue. Damage sitting on top of a thinned-out microvascular bed may cost more cognition than the same damage in a well-perfused brain. The retina is the only place those small vessels can be photographed directly and cheaply.
The concrete idea is to convert each retinal photograph into a graph of vessel segments and branch points. From that graph we compute how much of the fine, distal branching survives, with an explicit correction for branches that are merely invisible because of cataract or blur. We then test whether this "distal topology" index changes the slope linking white-matter damage to cognition, rather than simply adding another predictor of cognition. Abdominal fat depots measured by MRI enter as the upstream exposure that may drive the vessel loss. The claim is falsifiable, and it comes with a built-in test of whether the graph machinery is needed at all.
Core Claim
Earlier retinal work asks whether retinal features predict cognition or brain lesions. We ask whether they explain the residual: the part of cognition that lesion burden does not account for. Writing $C_i$ for a cognitive score of person $i$, $W_i$ for white matter hyperintensity volume, $R_i$ for the retinal distal-topology index (higher means more rarefied), and $Z_i$ for covariates, the central model is
$$C_i = \alpha + \beta W_i + \gamma R_i + \delta\, W_i R_i + \theta^\top Z_i + \varepsilon_i .$$
The hypothesis is $\delta < 0$: the same lesion volume costs more cognition when distal retinal branching is depleted. This mirrors the reserve pattern reported by Corpus Callosum Volume and Interhemispheric Transfer in Multiple Sclerosis, which found that callosal volume tracked transfer time in patients ($r=-0.66$) but not in controls, so structure only mapped to function once a reserve was exhausted. A non-zero $\delta$ is a different and stronger statement than a main effect $\gamma$. Generic confounders such as age shift scores, but it is harder for them to bend the lesion-to-cognition slope.
The secondary claim is that visceral fat, rather than BMI, relates to the distal-topology index. If it does, part of the moderation should run through it. This connects the vessel finding to a modifiable exposure.
What Prior Work Does and Does Not Settle
Two retinal studies map retinal features to brain lesions but never to cognition, and they motivate the vascular link. Retinal Biomarkers Discovery for Cerebral Small Vessel Disease in an Older Population used summary morphometrics (vessel width, tortuosity, fractal dimension) in the Lothian Birth Cohort 1936. It found that arteriolar tortuosity, branching coefficients and fractal dimension relate to hyperintensities, but those features were selected in only about 23% of bootstrap trials. That weakness is part of why we target a sharper feature. Lower Retinal Arteriolar Density Is Associated With Higher Cerebral Small Vessel Disease Burden reported modest correlations between OCT-angiography vessel density and lesion subtypes (around $r=-0.17$). Neither study builds a connected vessel graph, and neither reads out cognition.
The main constraint is the negative result in Comparison of the additive value of blood-based and retinal biomarkers to clinical features in predicting cognitive decline. In that study, 50 standard retinal features lowered five-year decline prediction from AUROC 0.691 to 0.680, while blood markers raised it to 0.748. We take this seriously. It suggests retinal features are a poor additive predictor of decline. Our design responds in two ways. It changes the question from additive prediction to moderation of a lesion effect, which fits Sim et al.'s own observation that retinal features track small vessel disease while blood markers track atrophy. It also changes the representation.
The graph method comes from Predicting Stroke through Retinal Graphs and Multimodal Self-supervised Learning. That work turned UK Biobank vessel segmentations into junction-and-segment graphs and trained a graph attention network (a neural network that passes messages along graph edges, weighting neighbours) for stroke. It reached 71.92% AUROC against 71.54% for a ResNet-50 on raw images. That margin is too small to justify graphs on accuracy alone, so we do not rest the proposal on accuracy. The graph earns its place by making distal connectivity an explicit, countable quantity. Losing that connectivity is exactly the failure that Human Treelike Tubular Structure Segmentation identifies as the main bottleneck for treelike structures. For reconnecting fragmented centerlines we borrow the cost-guided path-search idea from Circle of Willis Centerline Graphs, which joined broken skeleton pieces with A* search over a learned cost.
On body composition, Body roundness index and mental health linked a waist-and-height index to worse white-matter diffusion metrics in UK Biobank, but used no vascular measurement. Silhouette images enable estimation of body fat distribution showed that body outlines predict visceral fat ($R^2=0.885$) and the visceral-to-subcutaneous ratio far better than anthropometric formulas (0.553 against 0.171 in males). We use that result as the fallback when direct MRI fat measures are missing. Where they exist, we use the measured values.
Research Program
The first component is a quality-aware distal topology index. Vessels are segmented and thinned to centerlines. Branch points and endpoints become nodes. Gaps are bridged by cost-guided path search, and a learned classifier, as in road-graph extraction (Towards Satellite Image Road Graph Extraction), decides which candidate links are real. The central design choice is how to treat a missing branch. Cataract and blur in older eyes, which caused over 30% exclusion in the Lothian study, can make fine branches disappear. That would mimic rarefaction and correlate with age.
We therefore score each terminal region by local contrast and sharpness. A branch absent where the image is clear counts as lost. A branch absent where the image is degraded counts as unobserved and is down-weighted. The index $R_i$ is the quality-weighted deficit of terminal branches and branching generations, relative to an expectation conditioned on optic-disc-centred field position. We validate the correction by synthetically blurring clear images. Under blur, the raw count should fall while the corrected index stays stable.
The second component is the moderation analysis in the UK Biobank participants who have baseline retinal photographs, a later brain MRI with hyperintensity volume, cognitive tests, and abdominal MRI fat measures. Using the earlier retinal visit gives temporal ordering. The size of this overlap is a feasibility check we must run first; the file does not report it. Covariates are age, sex, blood pressure, diabetes, smoking, education and image-quality score. Cognitive outcomes are reaction time, fluid intelligence and processing-speed measures.
The third component is the body-composition arm. We regress $R_i$ on visceral fat, subcutaneous fat and their ratio, controlling for BMI and waist circumference. We then test whether the visceral-fat-to-cognition association flows partly through $R_i \times W_i$. External replication of the moderation term would use the Lothian Birth Cohort 1936, conditional on access to its cognitive data. We would also attempt a head-to-head with the Sim et al. memory-clinic protocol if access allows.
Baselines
The naive reference is the lesion-only model: cognition regressed on hyperintensity volume and covariates, with no retinal input. Every retinal model must add explained variance or a significant interaction beyond it. No reported result for this specific UK Biobank model was found on Lacuna, so we compute it ourselves on identical folds.
The classical competitor is summary morphometry of the VAMPIRE type, as used by Sim et al. and the Lothian study: vessel calibre, tortuosity, fractal dimension and branching coefficients. Each is entered in turn as $R_i$ in the same interaction model. The reported number this design is up against is Sim et al.'s clinical-plus-retinal AUROC of 0.680 against a 0.691 clinical baseline, with blood markers at 0.748. In any replication on that protocol, beating it means exceeding 0.691 with the topology index while keeping the same Elastic Net setup.
The strongest learned competitor is a retinal foundation-model embedding from RETFound (A foundation model for generalizable disease detection from retinal images). We reduce it to a scalar by fitting it to the lesion-adjusted cognitive residual inside training folds, then use that scalar as $R_i$. We also run Huang et al.'s graph attention encoder in the same way. Huang et al. report 71.92% stroke AUROC for that encoder, but no reported result was found on Lacuna for either model on cognitive moderation. All learned baselines get the same folds, covariates and hyperparameter search budget as our index.
Two ablations isolate the key ingredient. The first is the topology index without the quality correction. The second is an image-quality score alone entered as $R_i$. Beating the baselines means a larger and replicated $|\delta|$, and more out-of-sample variance explained in cognition beyond lesions. The quality-only control must show no comparable interaction.
Why not just…?
Why not just use RETFound or a CNN on the raw fundus image? This is the real threat. A large image model may absorb distal-branch information implicitly. Huang et al. found only a 0.4-point AUROC gap between graphs and images for stroke. If the RETFound-derived scalar matches our index on $\delta$, we will report that the graph is not needed for this question.
We expect the index to hold up for two reasons. It is fit without the outcome, so it cannot overfit cognition. Its quality correction blocks the age-cataract shortcut that a learned image model can exploit. The blurred-image experiment and the quality-only control test both points directly.
Why not just use VAMPIRE summary features like Sim et al.? Fractal dimension already summarises branching complexity. However, it mixes vessel loss with image degradation and with large-vessel geometry, and its link to lesions was unstable in Lothian bootstrap selection (about 23%). The head-to-head in the same interaction model settles this. If fractal dimension gives the same $\delta$, the cheaper feature wins.
Why not just use BMI, waist circumference or the body roundness index instead of MRI fat depots? For the vessel arm, waist-based indices may be enough. The silhouette study shows they are much worse at the visceral-to-subcutaneous ratio but not useless. We fit all of them and report whether visceral fat explains $R_i$ beyond BMI and waist. If it does not, the body-composition arm reduces to a covariate, and the core moderation claim stands independently.
Why not just add blood biomarkers, which already beat retina in Sim et al.? Blood markers track atrophy and neuronal injury, not microvascular supply. Where available, they enter as covariates. The test is whether $\delta$ survives their inclusion, which it should if the two measure different pathways.
Decisive Checks
The proposal succeeds only if three things hold. The quality-corrected distal topology index must show a negative interaction with hyperintensity volume on processing speed in UK Biobank, and that interaction must survive adjustment for age, vascular risk, image quality and the strongest baseline. The same sign must replicate in an external older cohort. The quality-only control must show no such interaction.
The proposal fails informatively in two cases. If $\delta$ is null for every retinal representation, retinal vessels do not explain lesion-cognition discordance. That would extend Sim et al.'s negative result from prediction to moderation. If RETFound or fractal dimension matches the graph index, the field should use the simpler tool. Either outcome is worth knowing. A positive result would give a cheap eye-based readout of which high-lesion patients are most at risk, linked to a measurable and modifiable fat depot.
References
- Corpus Callosum Volume and Interhemispheric Transfer in Multiple Sclerosis
- Retinal Biomarkers Discovery for Cerebral Small Vessel Disease in an Older Population
- Lower Retinal Arteriolar Density Is Associated With Higher Cerebral Small Vessel Disease Burden: An Optical Coherence Tomography Angiography Study
- Comparison of the additive value of blood-based and retinal biomarkers to clinical features in predicting cognitive decline
- Predicting Stroke Through Retinal Graphs and Multimodal Self-supervised Learning
- Human treelike tubular structure segmentation: A comprehensive review and future perspectives
- Circle of Willis centerline graphs: A dataset and baseline algorithm
- Body roundness index and mental health in middle-aged and elderly adults: a prospective cohort study
- Silhouette images enable estimation of body fat distribution and associated cardiometabolic risk
- Towards Satellite Image Road Graph Extraction: A Global-Scale Dataset and A Novel Method
- A foundation model for generalizable disease detection from retinal images