Retinal photographs predict stroke and brain small-vessel damage with surprising accuracy, but the models that do this learn statistical shortcuts from human eyes. Those shortcuts cannot be carried over to the rats, mice and pigs in which cerebrovascular drugs are actually tested. We propose a different approach. First, turn each segmented retinal vessel tree into an electrical-circuit model of blood flow, in which every vessel segment is a resistor whose value is set by its length and radius. Then summarize the solved flow with a small set of dimensionless numbers: ratios in which absolute vessel size, blood viscosity and heart rate cancel out. The hypothesis is that these numbers mean the same thing physiologically in a mouse and in a person. If so, a risk score fitted on human eyes paired with brain MRI could be applied unchanged to an animal's retina. Preclinical studies would then gain a human-calibrated readout of cerebral microvascular injury.
This matters because the preclinical cerebrovascular pipeline is visibly broken. A meta-analysis of 145 animal ischemia studies, 89% in rats, found enormous reported effect sizes alongside heterogeneity above $I^2 = 80\%$, strong publication bias, and wholesale failure in Phase III trials (Treatment of acute cerebral ischemia using animal models: a meta-analysis). Part of the problem is that there is no common unit for comparing a rodent's vascular state with a patient's. The retina is the one vascular bed that can be imaged non-invasively, repeatedly, and with the same optical instruments in every species. What it lacks is a species-neutral way to read the image.
Core Claim
Raw retinal morphometrics are not comparable across species. Vessel density, calibers and fractal dimension all shift with eye size, and rodents lack a macula entirely (Inhibition of AMD-Like Pathology With a Neurotrophic Compound in Aged Rats and 3xTg-AD Mice). Hemodynamic quantities computed on the vessel graph can be made scale-free, however.
We model each segment $e$ as a Poiseuille resistor with conductance $g_e = \pi r_e^4 / (8 \mu L_e)$, where $r_e$ is the segment radius, $L_e$ its length and $\mu$ blood viscosity. We fix pressure at the arteriolar root near the optic disc and at the venular outlets, and solve Kirchhoff's current law for the node pressures and segment flows. From the solution we report four descriptors:
- the network's effective resistance divided by that of a Murray's-law tree with the same root radius and terminal count;
- the coefficient of variation of flow across terminal capillary regions;
- a redundancy ratio, comparing effective resistance with the resistance of the single shortest path, in the spirit of the effective resistance distance used in brain networks;
- the mean deviation of bifurcation radius exponents from Murray's cube law.
Viscosity and absolute length scale cancel in every one of these ratios. Steady Poiseuille flow is a defensible approximation in retinal microvessels across mammals, because pulsatile inertial effects are small at these calibers regardless of heart rate. That assumption is itself tested below rather than taken for granted.
The claim is falsifiable in two ways. First, after normalization, healthy animals of different species should overlap on these descriptors far more than they do on raw density or fractal dimension. Second, a mapping from descriptors to cerebral white-matter injury, fitted only in humans, should rank animals by measured brain injury without being refitted.
Relation to Prior Work
Three groups of work motivate the proposal and supply its components.
The first establishes the eye–brain signal but in non-transferable forms. The OCTA study of 246 older adults found that lower arteriolar density tracks periventricular white matter hyperintensities ($r = -0.17$). Lower venular density tracks deep hyperintensities ($r = -0.17$), and reduced arteriolar branching complexity also tracks periventricular hyperintensities ($r = -0.18$) (Lower Retinal Arteriolar Density Is Associated With Higher Cerebral Small Vessel Disease Burden). This motivates separating arterioles from venules, but that study uses scalar area fractions and models neither connectivity nor flow. DeepRETStroke, a large vision transformer pre-trained on 895,640 photographs, detects MRI-confirmed silent brain infarcts at AUC 0.797, against 0.633 for clinical traits (A deep learning system for detecting silent brain infarction and predicting stroke risk). RETFound does similar systemic prediction from 1.6 million human images (Zhou et al.). Both are powerful within human cohorts, but they are trained exclusively on human eyes. A review of fundus foundation models also notes their limited interpretability (Wei Dai et al.). We do not try to beat them on human AUC. We aim at the cross-species question they cannot address.
The second group supplies the method. AFR-Net treats structural brain connectivity as conductance and solves Kirchhoff equilibria with functional demand, raising Parkinson's classification on PPMI to AUC 90.38% from 71.88% (Uncovering Latent Communication Patterns in Brain Networks via Adaptive Flow Routing). Graph signal processing decomposes brain signals over a structural graph Laplacian (A Graph Signal Processing Perspective on Functional Brain Imaging). Both work on 82–400-region neural parcellations with abstract weights. We reuse the Kirchhoff machinery but give it literal physical meaning on vessels, where conductance comes from measured geometry. AFR-Net's $O(N^3)$ dense solver is a stated limitation. Retinal vessel graphs are near-planar and sparse, so sparse Cholesky or conjugate-gradient solves handle tens of thousands of segments.
The third group constrains the design. Rodents lack a macula and fovea (Wei et al.), so we restrict all descriptors to the peripapillary region around the optic disc, which every species has. Small-rodent hemodynamics operate at roughly tenfold higher heart rates than human hemodynamics. Large animals such as Tibetan minipigs under descending aortic constriction reproduce human-like pressure-overload phenotypes (A Porcine Model of Heart Failure With Preserved Ejection Fraction Induced by Chronic Pressure Overload). That makes them a natural middle rung between rodent and human, and the constriction raises pressure upstream, including in the head vessels. Segmentation is also a constraint. On FIVES, U-Net-class models reach a Dice similarity coefficient of about 0.90, more complex models do not help, and broken vessel topology is the main failure (Benchmarking Retinal Blood Vessel Segmentation Models for Cross-Dataset and Cross-Disease Generalization). A resistor network with gaps gives wrong flows, so topology preservation is a prerequisite, not a refinement.
Research Program
The pipeline has four stages.
Segmentation. A U-Net trained with a centerline-Dice (clDice) topology loss on FIVES segments OCTA superficial-plexus and fundus images. Arteries and veins are labeled, and a gap-bridging step reconnects short breaks.
Graph construction. The segmentation is skeletonized into a graph whose edges carry estimated radius and length. On fundus photographs, where capillaries are invisible, arteriolar terminals are joined to nearby venular terminals through a lumped capillary resistance. Because this lumped resistance is a modeling choice, its value is swept and descriptor stability reported.
Solving and summarizing. Kirchhoff flow is solved and the four dimensionless descriptors are computed, separately for arteriolar and venular subtrees where possible.
Calibration. A low-capacity model, either penalized linear or monotone, maps descriptors plus age and blood pressure to white-matter injury in humans. It is then frozen.
We also quantify how segmentation errors propagate. Each descriptor's variance is measured under synthetic edge deletions matched to the gap statistics observed on FIVES. Any descriptor that is more sensitive to segmentation noise than to disease is dropped before outcome analysis.
Baselines
The naive reference is a clinical-trait model (age, sex, blood pressure, diabetes, smoking) predicting white-matter injury. For silent infarcts, the closest reported number is AUC 0.633 for clinical traits in DeepRETStroke. For white-matter hyperintensity volume in UK Biobank there is no reported result found on Lacuna, so we compute it ourselves on the same splits.
The morphometric baseline is arteriolar and venular vessel density plus box-counting fractal dimension. These are computed from the identical segmentations and fed to the identical calibration model, with the reported correlations of $r = -0.14$ to $-0.20$ (Lower Retinal Arteriolar Density) as the expected human-only strength.
The strongest existing method for human prediction is a foundation-model embedding. We use RETFound features with a linear probe, and DeepRETStroke where weights are obtainable, given the same labeled training subset and hyperparameter search budget as our calibration model. Its reported strongest relevant figure is SBI AUC 0.797 (DeepRETStroke).
Two ablations isolate the key ingredient. The first keeps the graph and descriptors but replaces $r^4/L$ conductance with topology-only or $r^2/L$ weights, removing the physics. The second keeps the physics but uses raw, un-normalized flows and resistances, removing dimensionlessness.
Success has two tiers. On humans, we must be no worse than the morphometric baseline and within a few AUC points of the foundation probe, while offering arteriolar and venular decomposition. On the cross-species tests below, we must beat every baseline. In practice the foundation probe and the raw morphometrics are expected to fail there, and the comparison that counts is against species-standardized morphometrics.
Why not just…?
Why not just use RETFound or DeepRETStroke embeddings? Within humans they will probably match or exceed us, and we say so plainly. The question here is different: their features are learned on human pixel statistics, and nothing ensures that an embedding direction meaning "white-matter damage" in a human fundus means anything in a rat fundus, which has different optics, pigmentation and no macula. Experiment 3 applies the frozen human probe to animal images directly. We expect it to fail, and if it does not, the case for physics weakens considerably.
Why not just z-score density and fractal dimension within each species? This is the real threat. It is cheap, and it could remove scale differences as effectively as physical normalization. Its weakness is that it needs a large healthy reference group per species and per age, and it assumes disease shifts each metric by the same number of standard deviations in every species. Our descriptors make no such assumption. Experiments 2 and 3 run this baseline head to head. If species-standardized morphometrics transfer as well as the flow descriptors, we will report that the hemodynamic model adds nothing beyond normalization.
Why not just train a graph neural network on the vessel graph? A learned graph model uses the same topology without imposing physics. Like the foundation probe, it fits human-specific statistics and will not be dimensionless unless it is forced to be. The conductance ablation tests the same question more cleanly. If topology-only weights perform like $r^4/L$ weights, the physical model is not earning its place.
Why not just stay in human biobanks? Human cohorts cannot test interventions. The point of a cross-species readout is to score drugs in animals on a human-calibrated axis. This program does not argue that animal models are reliable, only that their outcomes need a common unit before their reliability can be judged.
Proposed Experiments
Experiment 1 (human validity). We compute descriptors on UK Biobank fundus photographs from participants with brain MRI, and on an OCTA cohort with paired MRI of the kind used in the 246-subject study. We report partial correlation with white-matter hyperintensity volume and with periventricular and deep Fazekas grades, adjusting for age and blood pressure, and AUC for the top quartile of hyperintensity volume. A decisive check is whether arteriolar descriptors track periventricular injury and venular descriptors track deep injury, reproducing the sub-compartment pattern of the OCTA study.
Experiment 2 (invariance). We acquire peripapillary OCTA in healthy mice, rats, minipigs and humans across age bands. For each descriptor and each baseline metric we report the fraction of variance explained by species. The prediction is that the dimensionless descriptors show a small species fraction while raw density and caliber show a large one.
Experiment 3 (zero-refit transfer, decisive). We apply the frozen human-calibrated score to animals with induced chronic hypertension: rodent hypertension models and the minipig aortic-constriction model (Tan et al.), each against sham controls. We then measure white-matter injury by MRI and histology. The primary endpoint is Spearman correlation between the frozen score and measured brain injury, pooled across species with a single mapping, compared against the frozen foundation probe and species-standardized morphometrics.
A secondary check tests the steady-flow assumption. Time-resolved dynamic OCT velocities (Time-Resolved Dynamic Optical Coherence Tomography for Retinal Blood Flow Analysis) are compared with solver-predicted segment flows in each species. Agreement that degrades in rodents would localize where heart-rate differences break the model.
Success in Experiment 3 would give preclinical cerebrovascular studies a common outcome unit, which the 145-study meta-analysis never had. Failure would show that cross-species divergence in the retina is not merely a matter of scale.
References
- Treatment of acute cerebral ischemia using animal models: a meta-analysis
- Inhibition of AMD-Like Pathology With a Neurotrophic Compound in Aged Rats and 3xTg-AD Mice
- Lower Retinal Arteriolar Density Is Associated With Higher Cerebral Small Vessel Disease Burden: An Optical Coherence Tomography Angiography Study
- A deep learning system for detecting silent brain infarction and predicting stroke risk
- A foundation model for generalizable disease detection from retinal images
- An Efficient and Interpretable Foundation Model for Retinal Image Analysis in Disease Diagnosis
- Uncovering Latent Communication Patterns in Brain Networks via Adaptive Flow Routing
- A Graph Signal Processing Perspective on Functional Brain Imaging
- A Porcine Model of Heart Failure With Preserved Ejection Fraction Induced by Chronic Pressure Overload Characterized by Cardiac Fibrosis and Remodeling
- Benchmarking Retinal Blood Vessel Segmentation Models for Cross-Dataset and Cross-Disease Generalization
- Time-Resolved Dynamic Optical Coherence Tomography for Retinal Blood Flow Analysis