The retina is an outgrowth of the brain. Its small blood vessels share developmental origin and regulation with the brain's small vessels, so retinal images carry weak but real signals about cerebral small vessel disease, the diffuse damage seen on brain MRI as white matter hyperintensities (bright patches) and enlarged perivascular spaces. The datasets that could quantify this link are badly incomplete, and in a structured way. Brain MRI exists for a small fraction of people with retinal images. Different cohorts use different retinal devices: fundus photographs in some, optical coherence tomography (OCT, a depth-resolved scan) or its blood-flow variant OCT angiography in others. In older adults, a large share of images cannot be measured because of cataract. One Scottish study discarded 32% of participants for this reason. Those excluded are older and sicker, which is exactly the group in which the eye–brain signal matters most.
We propose to treat this as an imputation problem at the level of measured features rather than pixels. The missing values come in whole blocks: "this person has no MRI," "this cohort has no OCT-A," "this image was too blurred for vessel width." We adapt the discriminator "hint" from adversarial imputation so that it works on blocks instead of single entries. The discriminator is the network that judges whether a filled-in value looks real. The adapted hint makes it check each imputed block against the blocks that were actually observed for the same person, and it only asks about pairs of blocks that are observed together somewhere in the data. The output is multiple imputed copies of each person's missing features, so that people excluded for cataract, missing MRI, or the wrong device can be kept in eye–brain analyses with honest uncertainty.
Core Claim
In pooled retinal–brain cohorts, missingness has three features: it covers whole modalities, it depends on patient state, and it varies by cohort design. Under these conditions, an adversarial imputer whose hint reveals missingness block by block and respects which block pairs are ever jointly observed will do two things better than chained-equation and random-forest imputation. It will recover retinal–cerebral association estimates closer to what full data would give, and it will produce better-calibrated intervals for held-out brain markers. The gain should grow with the nonlinearity of the eye–brain coupling and with the fraction of missingness driven by image quality. We also state when the method should not win: when the couplings are close to linear and missingness is driven only by observed covariates. In that regime we expect MICE to tie, and we will report that.
Why Not the Obvious Version
The obvious move at this intersection is to synthesize a brain FLAIR volume from a fundus photograph with a GAN, treating the MRI as a missing block. We reject it. Correlations between retinal vessel density and white matter lesion burden are around $r=-0.17$ (Lower Retinal Arteriolar Density…), so a retina cannot determine where lesions sit in the brain. Any synthesized volume would be mostly hallucination around a scalar. When scalar endpoints are the goal, direct prediction already works: DeepRETStroke detects MRI-confirmed silent brain infarcts at AUC 0.797 from photographs (DeepRETStroke). The closest imputation work stays within one organ. Unified Multi-Modal Image Synthesis synthesizes missing MRI sequences from other MRI sequences. Modality-Agnostic Style Transfer imputes regional PET and MRI features in ADNI (a multi-site Alzheimer's imaging study) using gradient-reversal disentanglement. Neither crosses organs. Neither handles missingness that depends on patient state. Neither uses a hint, the ingredient that gives GAIN its guarantee of recovering the true conditional distribution.
Research Program
The data representation is a per-person vector split into blocks. The blocks are: fundus morphometry (arteriolar and venular calibers, fractal dimension, tortuosity, branching coefficients, VAMPIRE-style as in Retinal Biomarkers Discovery for SVD); a low-dimensional principal-component summary of RETFound embeddings; OCT layer thicknesses (retinal nerve fiber layer, ganglion cell layer); OCT-A vessel densities where available; clinical risk factors; and brain small vessel disease scalars (periventricular and deep white matter hyperintensity volume, perivascular space burden). An image-quality score is always observed and enters as a covariate. Low-quality images are not dropped. Each morphometric feature is masked individually, depending on whether it can be measured at that quality level. Caliber often survives blur that destroys fractal dimension, so this recovers part of the 32% loss at the feature level.
The method reuses GAIN's generator and discriminator and modifies the hint. In GAIN, the generator fills missing entries from observed values, the mask, and noise. The discriminator predicts, entry by entry, which values were observed, and it is told the true mask status of all but a random subset of entries. Our version reveals the mask of every block except one held-out block $b$. It samples $b$ only among blocks that are observed together with at least one other revealed block in some training row. The discriminator is scored on its prediction for block $b$:
$$ \mathcal{L}_D = -\,\mathbb{E}\big[\,m_b \log D_b(\tilde{x}, h) + (1-m_b)\log(1-D_b(\tilde{x}, h))\,\big], $$
Here $\tilde{x}$ is the completed vector, $h$ is the block hint, $m_b\in\{0,1\}$ records whether block $b$ was observed, and $D_b$ is the discriminator's probability that block $b$ is real. The generator adds GAIN's reconstruction loss on observed entries.
The restriction on which blocks can be held out is our answer to a theory gap. GAIN's uniqueness proof assumes entries are missing completely at random. The conditional between two blocks that are never observed together is simply not identified. We will state and check the weaker claim: the hinted game constrains only conditionals supported by jointly observed pairs, which in UK Biobank means the 6,454 people with both retinal and brain imaging (Eye-brain connections). For pairs outside that support, the output must fall back to the unconditional marginal, and we will flag it. Repeated noise draws give multiple imputations. These are pooled with Rubin's rules, which combine within-imputation and between-imputation variance, so that downstream association estimates carry the added uncertainty.
Baselines
The naive reference is complete-case analysis: drop anyone with a missing block, as the Lothian study did. There, Elastic Net selected retinal predictors of white matter hyperintensities in only about 23% of bootstrap trials, compared with up to 99.6% for perivascular spaces (Retinal Biomarkers Discovery for SVD).
The strongest statistical competitors are multiple imputation by chained equations (MICE) with nonlinear terms and MissForest, run on exactly the same feature blocks, quality covariate, and number of imputations. Each gets a tuning budget matched to the adversarial model in wall-clock hours. MICE is what an epidemiologist would actually use, which makes it the competitor to beat. The only head-to-head numbers in the file are on unrelated tabular data: GAIN reached downstream AUROC 0.9711 against MissForest's 0.9597 on UCI News (GAIN). There is no reported result found on Lacuna for any imputer on retinal–brain feature data.
Two ablations isolate the key ingredients. The first is standard entry-level GAIN on the same vectors, which tests whether block hints matter at all. The second is block-hint GAIN without the joint-support restriction, which tests whether that restriction prevents confident nonsense on unidentified block pairs.
As a strong learned alternative, we run AugMask diffusion imputation (AugMask) under the same budget.
Beating the baselines requires the following on semi-synthetic masks with known truth. Our method must show lower bias in the retinal–cerebral regression coefficients relative to full-data estimates. It must show nominal 90% interval coverage for the brain block, within ±3 points of 90%. It must avoid any loss in the imputation error of the brain block relative to MICE. All of this must hold across at least three missingness mechanisms. Winning on error alone does not count.
Why not just…?
Why not just use MICE with multiple imputation? This is the real threat. MICE is transparent and widely trusted, and if the eye–brain couplings are near-linear it may match us. The Lothian finding that linear models capture white matter signal only weakly is suggestive but does not settle it: that could reflect a weak signal rather than nonlinearity. Our semi-synthetic experiments vary coupling nonlinearity on purpose. If MICE ties on real UK Biobank masks, we will report that the block hint matters only when missingness depends on quality, and we will recommend MICE otherwise.
Why not just predict brain markers directly from images, as DeepRETStroke does? For risk scores alone, direct prediction is the right tool, and we include a RETFound-finetuned predictor as the reference for held-out brain scalars. It does not solve the problem we target. It trains only on people with MRI and gives no answer for people whose images are unmeasurable. It also does not produce completed data with uncertainty for estimating associations across pooled cohorts that use different retinal devices.
Why not use SMIL-style prototypes or diffusion, since GANs fail at 90% missingness? SMIL reports GAN imputation at 56.5% accuracy against its own 60.7% on CMU-MOSI with 10% text (SMIL), and our MRI block is similarly scarce. We read this as evidence against unrestricted generation. SMIL, however, outputs classifier features, not imputed values usable for association estimates, so it answers a different question. Diffusion is included directly as the AugMask baseline. If it wins, we adopt the block-hint support restriction as a masking rule for diffusion instead.
Decisive Experiments
In UK Biobank's paired retinal–brain subset, we mask the brain block for 90% of people. We mask morphometric features through three mechanisms: missing completely at random; quality-driven missingness in which dropout depends on age and the observed quality score; and a not-at-random mechanism in which dropout also depends on the true feature value, mimicking cataract co-occurring with vascular aging. We then compare coefficient bias, interval coverage, and error against the unmasked data.
Cross-modality pooling is tested within UK Biobank by withholding the OCT block from one half and the fundus block from the other. Each must be imputed from the blocks observed alongside it. This simulates pooling a fundus-only cohort with an OCT-only cohort.
External checks use two cohorts. On LBC1936, we test whether including the 32% previously excluded participants, with feature-level quality masks, stabilizes selection of white matter predictors beyond the 23% bootstrap rate. On the West China OCT-A cohort, we test whether imputed estimates stay in the direction of the reported $r\approx-0.17$ associations.
The result that would falsify the proposal is block-hint GAIN failing to beat MICE on coefficient bias under the quality-driven mechanism, the setting it was designed for.
References
- 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
- Unified Multi-Modal Image Synthesis for Missing Modality Imputation
- Modality-Agnostic Style Transfer for Holistic Feature Imputation
- GAIN: Missing Data Imputation using Generative Adversarial Nets
- Retinal Biomarkers Discovery for Cerebral Small Vessel Disease in an Older Population
- A foundation model for generalizable disease detection from retinal images
- Eye-brain connections revealed by multimodal retinal and brain imaging genetics
- AugMask: Training Diffusion Models on Incomplete Tabular Data via Stochastic Augmentation and Masking
- SMIL: Multimodal Learning with Severely Missing Modality