Unmixing Projection Artifact from Real Capillary Loss in Multi-Device Retinal Angiography

Retinal angiography with optical coherence tomography (OCTA, a camera that maps blood flow in each depth layer of the retina without dye) is one of the cheapest ways to look at small vessels that develop from the same tissue as the brain. It is a candidate screening tool for dementia and cerebral small vessel disease in clinics that cannot offer amyloid PET or lumbar puncture. The literature, however, cannot agree on which retinal layer carries the signal. One lightweight classifier reports the deep capillary layer as most informative for early dementia. A segmentation study finds mild cognitive impairment mainly in the superficial layer, with deep loss only in established Alzheimer's disease. A meta-analysis finds no significant deep-layer parafoveal loss at all. The deep layer is also where a well-known measurement error is concentrated. Light passing through superficial vessels creates a ghost copy of them in the deep-layer image, and the strength of that ghost depends on the device and the scan size.

This note proposes to treat that conflict as a source-apportionment problem of the kind environmental chemists solve with positive matrix factorization (PMF). PMF splits each air or soil sample into non-negative contributions from several emitters. Here, each eye's measured layer-by-sector vessel densities are modeled as a non-negative sum of biological capillary components plus a technical projection component. The projection component is pinned to a known physical shape: deep-layer density proportional to superficial density in the same sector, with a device-specific ratio. PMF also weights every entry by its own measurement uncertainty. That weighting lets noisy wide-field scans contribute instead of being pooled naively or dropped. The claim is testable: if the deep-layer Alzheimer's signal is partly ghost, it should shrink and become consistent across devices once the projection component is removed. If it is real, it should survive.

Core Claim

Additive projection artifact is a genuine non-negative linear mixture. That is the narrow setting where PMF's physical assumptions hold, unlike the loose analogy of "site as a pollution source." Fitting a constrained, uncertainty-weighted factorization to multi-device OCTA feature matrices should produce layer-specific disease effects that agree across devices and scan sizes, where unadjusted features do not. It should also give an interpretable, low-compute screening score that loses little accuracy relative to deep classifiers.

Why This Is Not Already Done

The closest work splits along two lines. The first line separates device noise inside a neural network. D2Net uses a dual-stream encoder with a divergence penalty to separate vessels from instrument artifacts at the pixel level. It reaches 63.48% Dice on FOCA and 83.83% zero-shot AUC on ROSE-H, but it degrades on 6×6 mm scans and yields no explicit, per-patient amount of artifact. We reuse a segmentation stage of this kind as the upstream measurement, not as the competitor. The second line measures standardized features and tests them one at a time. Deep segmentation of OCTA extracts 12 geometric parameters with OCTA-Net and FAZ-Net and compares groups variable by variable. The SVD study regresses arteriolar and venular density on MRI burden; for example, it reports r = −0.17 with periventricular white matter lesions. Neither line models how one measured column contaminates another.

Two papers motivate the work. The meta-analysis shows that 3×3 mm scans give lower heterogeneity than 6×6 mm scans, and that the deep parafoveal effect is null (P = 0.19). PolarNet+ names the deep layer as most sensitive. On the method side, the receptor-modeling studies supply the technique. These include PMF of smelter soils, which resolved four sources from only 32 samples, and ME-2 constrained factorization, which pins factor ratios to known values. To our knowledge from the file, none of this has been applied to ophthalmic data.

Research Program

Each scan is first segmented with an OCTA-Net-style pipeline. We compute vessel area density, length density and fractal dimension for the superficial, deep and composite inner layers in each sector of the standard nine-sector macular grid. This gives a scans-by-features matrix $X$ pooled across devices and scan sizes. We fit

$$\min_{G \ge 0,\, F \ge 0} \; Q = \sum_{i,j} \left( \frac{X_{ij} - \sum_k G_{ik} F_{kj}}{\sigma_{ij}} \right)^2,$$

where $G_{ik}$ is how much of component $k$ is present in scan $i$, and $F_{kj}$ is component $k$'s profile over features $j$. The term $\sigma_{ij}$ is the uncertainty of entry $ij$. We estimate it from repeat-scan variability where available, and otherwise from segmentation confidence and scan size, so that 6×6 mm entries count less.

One component is constrained in the ME-2 style. Its deep-layer loadings must equal $\alpha_d$ times its superficial loadings sector by sector, where $\alpha_d$ is a ratio estimated separately for each device $d$. This encodes the physics of projection: ghost density appears in the deep layer where superficial vessels are. The remaining components are unconstrained and are read as biological capillary reservoirs.

We will treat rotational ambiguity, the known weakness of PMF, the way EPA PMF does. We choose the number of components by comparing $Q$ to its expected value and by bootstrap stability, and we report displacement-based uncertainty on every profile. Fractal dimension is not additive, so a sensitivity analysis restricts the factorization to the density columns, which are additive and non-negative.

The output is a small set of per-eye contributions. Downstream analyses use these contributions, plus deep-layer density with the projection component removed, in place of raw columns. These analyses are disease group effects, MRI correlations, and a logistic screening score with a few dozen parameters.

Data

The primary analysis needs OCTA with dementia labels from more than one device. The ROAD and ROMCI cohorts used by PolarNet+, with 1,671 participants, are the target, and access requires collaboration with their holders. The West China cohort from the SVD study provides paired 3T MRI for 246 adults. Device and scan-size variation for estimating projection without disease labels comes from OCTA-500, which has both 3×3 and 6×6 mm subsets, FOCA, and the ROSE sets.

A secondary, smaller arm asks whether the same weighted factorization finds a center component in fundus-photo features from AutoMorph on UK Biobank. The UK Biobank bias study reported a 10% AUC gap between assessment centers despite identical hardware there. We treat this arm as exploratory, because no physical mixing law is known for center effects.

Baselines

The naive reference is the analysis most papers run. We take raw layer-sector densities, test AD and MCI against controls with site entered as a covariate, and fit a logistic screening score on the raw features. This is the approach of Deep segmentation of OCTA, and it is what produces the conflicting layer conclusions.

The strongest feature-level competitor is residualization. We regress deep density on same-sector superficial density within each device and keep the residual, optionally after per-site mean-and-variance harmonization. This is a serious threat because it removes the same shared variance with no factorization. No reported result was found on Lacuna for this baseline on dementia OCTA.

The strongest end-to-end competitor is PolarNet+, which reports AUC 88.69% for AD, 88.02% for MCI, and 86.55% on external sets with 3.92M parameters. We will retrain it on our exact splits rather than quote it. We will also fine-tune RETFound on OCTA projections with the same splits and an equal hyperparameter-search budget. All methods share patient-level splits and a held-out device. Each gets the same number of tuning trials.

The two ablations isolate the key ingredients. The first is unweighted factorization with all $\sigma_{ij}$ equal. The second is factorization without the projection constraint, which is plain non-negative matrix factorization. A PCA baseline checks whether non-negativity matters at all.

Success has two parts. On consistency, between-device heterogeneity ($I^2$) of the standardized deep-layer AD effect should fall below that of both the naive and residualized baselines, with non-overlapping bootstrap intervals. On screening, held-out-device AUC should come within three points of the retrained PolarNet+, using a model small enough to run on a clinic laptop. Beating PolarNet+ on AUC is not required for the claim.

Why not just…?

Why not just regress the deep layer on the superficial layer? This is the real threat. Residualization removes all shared variance, including genuine superficial and deep capillary loss that happen together in disease. The constrained factorization removes only variance with the projection shape and the device-specific ratio. A semi-synthetic test settles which is right. We inject known deep-only and joint-layer dropout into control eyes' feature vectors at realistic magnitudes and measure recovery bias. If residualization recovers injected joint-layer dropout as well as the factorization does, we will drop the factorization and report residualization as sufficient.

Why not just fine-tune a foundation model? RETFound and DeepRETStroke may match or beat our AUC; the latter reports SBI detection AUC 0.797. But a probability score cannot say whether the deep layer carries real signal. The UK Biobank bias study also shows that end-to-end models absorb center shifts, and that standard fixes "level down" performance. We run the foundation model as a baseline and compare its held-out-device drop against ours.

Why not just use PCA or unconstrained factor analysis? Components with mixed signs cannot represent an additive ghost, and they have no place to impose the projection ratio. The PCA ablation measures this directly. We concede that PMF has its own rotational freedom, and we report bootstrap profile stability so that readers can judge it.

Decisive Checks

Three results would decide the question. The first is whether the fitted ratio $\alpha_d$ is larger on 6×6 mm than on 3×3 mm scans, and differs by device, in the direction implied by the meta-analysis's heterogeneity pattern. If it does not, the projection component is not capturing projection. The second is whether the deep-layer AD effect, after removing projection, becomes consistent across devices. Whether it shrinks toward the meta-analytic null or stays the size PolarNet+ implies, this directly arbitrates the three-way literature conflict. The third uses the MRI cohort: projection-removed arteriolar and venular contributions should keep or strengthen their reported associations with white matter lesions, while the projection component should show none. A projection component that correlates with brain lesions would mean it has absorbed biology, and the model would be rejected.