Two-Band Ratio Attenuation for Retinal Proxies of Cerebral Small Vessel Disease

Retinal scans are increasingly read as a window onto the brain's small blood vessels. Most methods, though, measure geometry (how dense the capillaries look) or let a large neural network find whatever texture predicts disease. Both are fragile in the same way. Cataracts, defocus and poor scan quality dim and distort the image, and patients with cerebrovascular disease and dementia tend to have worse scans. A model can therefore learn the artifact instead of the disease. Atmospheric scientists solved a structurally identical problem decades ago with differential absorption lidar (DIAL). DIAL fires two nearby laser wavelengths through the same air and takes the ratio of the two returns. Everything the two wavelengths share cancels: unknown particle reflectivity, telescope losses, haze between the instrument and the target. What remains is the property that differs between them.

We propose to apply this ratio trick to optical coherence tomography (OCT), the depth-resolved infrared imaging used in both eye clinics and catheter-based artery imaging. One acquisition's spectrum is split into two sub-bands. Comparing how fast each sub-band's signal fades with depth gives a measurement of tissue light loss that is unaffected by anything sitting in front of the retina. The method is validated first inside arteries, where intravascular OCT comes with expert tissue labels. It is then tested on whether the measurement adds information about MRI-confirmed small vessel disease beyond what capillary density already provides, while staying stable when a patient's cataract is removed.

Core Claim

A depth-resolved "differential extinction" computed from two sub-bands of a single OCT scan has two properties. It does not depend on the constant backscatter-to-attenuation ratio that current OCT attenuation methods assume. It is also mathematically blind to depth-independent losses in front of the tissue, such as lens opacity, tear film, and overall signal strength. On that basis we claim that retinal deep-vascular-slab differential extinction explains variance in cerebral small vessel disease that OCT angiography (OCTA) vessel density does not, and that it does so without inheriting the scan-quality confound.

Why the Lidar Analogy Is Exact Enough to Use

Single-wavelength elastic lidar returns a signal that depends on two unknowns at every range, backscatter and extinction. It can only be inverted by assuming their ratio, the "lidar ratio," and that ratio varies widely across particle types (A study on the aerosol extinction-to-backscatter ratio with combination of micro-pulse LIDAR and MODIS over Hong Kong). The field breaks this deadlock in three ways: by observing the same layer from multiple angles (Comparison of Scanning LiDAR with Other Remote Sensing Measurements and Transport Model Predictions for a Saharan Dust Case), by using Raman returns from nitrogen whose density is known (Raman LIDARs for the atmospheric calibration along the line-of-sight of the Cherenkov Telescope Array), or by differential absorption. These papers supply the conceptual toolkit. We reuse only the third route, because Raman-like reference scatterers do not exist in the retina, and the weak-signal problem that forces Raman systems onto 1.8 m telescopes is far worse at retinal-safe light powers.

OCT has inherited the elastic-lidar problem unchanged. Automatic Classification of A-Lines in Intravascular OCT Images Using Deep Learning and Estimation of Attenuation Coefficients converts intravascular OCT depth scans into attenuation maps using a single-scattering model. Fibrolipidic-versus-fibrocalcific accuracy rises from 74.94% to 83.47%, but the model assumes a constant backscatter-to-attenuation ratio along the beam, which is exactly the lidar-ratio assumption. That paper both motivates the work (physical normalization helps) and defines the weakness we target.

Write the signal in sub-band $k \in \{1,2\}$ at depth $z$ as $I_k(z) = C_k(z)\,T_k\,\beta_k(z)\exp(-2\int_0^z \mu_k)$. Here $C_k(z)$ is the instrument's known depth response (focus and sensitivity roll-off, calibrated on a reference phantom), $T_k$ is the depth-independent transmission of everything in front of the retina, $\beta_k$ is backscatter, and $\mu_k$ is extinction. After dividing out $C_k$, the DIAL estimate is

$$\Delta\mu(z) \equiv \mu_1(z)-\mu_2(z) = -\tfrac{1}{2}\frac{d}{dz}\ln\frac{I_1(z)}{I_2(z)} + \tfrac{1}{2}\frac{d}{dz}\ln\frac{\beta_1(z)}{\beta_2(z)}.$$

The transmissions $T_1$ and $T_2$ vanish under the derivative. The last term is zero wherever the spectral shape of backscatter is constant within a tissue layer, which is a much weaker assumption than a constant lidar ratio. The approximation fails at layer boundaries, where scatterer size changes. We handle this the way Separation of absorption and scattering profiles in spectroscopic optical coherence tomography using a least-squares algorithm does: a regularized least-squares inversion with outlier rejection and Laplacian smoothing, applied piecewise within segmented layers. That paper supplies the inversion method and also two hard constraints. First, absorption errors stay under 5% only above about 14 dB SNR. Second, scatterer retrieval is markedly less accurate than absorber retrieval.

We are explicit about what $\Delta\mu$ measures. In the near-infrared bands of clinical swept-source OCT, hemoglobin absorption contrast is weak. $\Delta\mu$ will therefore mostly reflect the wavelength slope of scattering, which tracks scatterer size, plus a contribution from water absorption near 1 µm. This is the quantity we propose to test. We do not claim to measure lipid or oxygen.

Research Program

The program runs in three stages. Each stage must pass before the next is trusted.

In the first stage, we build the two-band estimator and characterize it on phantoms modeled on the Rohrbach design. These are layered silica microbeads of 0.16, 0.33 and 0.8 µm with a near-infrared dye. We add one manipulation: calibrated diffusers and neutral-density filters placed in front of the phantom, to mimic cataract and low signal strength. We map retrieval error against sub-band separation, sub-band width (splitting the spectrum coarsens depth resolution) and SNR. This yields an operating envelope that states which retinal slabs are thick enough to measure. Arteriolar walls are about 10–20 µm thick and will probably fall below the split-band resolution. The target is therefore slab-level extinction in the superficial and deep vascular complexes and the perivascular tissue column beneath large arterioles, not the wall itself.

The second stage is ground-truth validation in arteries. Intravascular OCT is the one setting where tissue identity is labeled at depth-scan resolution, including the plaque classes of the A-line dataset and the acute intimal tears and mural thrombi that Assessment of acute radial artery injury after distal transradial access for coronary intervention: an optical coherence tomography study found in nearly half of transradial patients. The published A-line dataset appears to contain processed intensity images, so we will request the raw spectral interferograms. If they were not retained, we will acquire ex vivo human radial and coronary segments with matched histology on a swept-source system that stores raw data. This stage asks whether the ratio method separates tissue types that single-band attenuation confuses. That is a necessary condition for trusting it in the eye, where no tissue labels exist.

The third stage is the retinal application. We acquire swept-source OCT and OCTA with raw spectra retained in an older-adult cohort with brain MRI, following the protocol of the West China small vessel disease study (Lower Retinal Arteriolar Density Is Associated With Higher Cerebral Small Vessel Disease Burden: An Optical Coherence Tomography Angiography Study). We add an acute-stroke arm with large artery atherosclerosis and small artery disease subtypes, as in Differences in retinal microvasculature between large artery atherosclerosis and small artery disease: an optical coherence tomography study, plus healthy controls, which that study lacked. Vessel and slab segmentation uses OCTA-Net trained on ROSE and artery/vein labels from OCTA-500. The downstream predictor is deliberately small. Following the evidence in State-of-the-art retinal vessel segmentation with minimalistic models that a 68,000-parameter network suffices on standard vessel tasks, the predictor is a regression on a handful of slab-wise $\Delta\mu$ summaries plus density. Any gain should then be attributable to the physical measurement rather than to model capacity.

Baselines

The naive reference point is clinical risk factors alone (age, blood pressure, diabetes, smoking). For context, the only number on Lacuna is for silent brain infarction rather than white matter disease: clinical traits reach an AUC of 0.633 there (A deep learning system for detecting silent brain infarction and predicting stroke risk).

The strongest existing retinal-structure method on this exact question is OCTA vessel density and fractal dimension from OCTA-Net, as used in the West China small vessel disease cohort. That study reports a correlation of r = −0.17 between arteriolar density and periventricular white matter hyperintensity, r = −0.20 with basal ganglia perivascular spaces, and r = −0.14 with composite burden (Lower Retinal Arteriolar Density Is Associated With Higher Cerebral Small Vessel Disease Burden). For subtype discrimination, the corresponding result is deep vascular complex density separating large artery from small artery disease at p = 0.022 (Differences in retinal microvasculature between large artery atherosclerosis and small artery disease). We recompute these features on our own scans with the same segmentation model, so both arms see identical eyes and identical preprocessing. Beating this baseline means a statistically significant increase in cross-validated explained variance for white matter hyperintensity and composite small vessel disease burden when $\Delta\mu$ is added to density and clinical covariates.

The strongest learned baseline is a retinal foundation model. RETFound is a masked-autoencoder model pretrained on 1.6 million retinal images and reports a 3-year heart-failure AUC of 0.799 from OCT. We fine-tune it on our OCT volumes with the same folds and a hyperparameter search budget matched to the small model. A head-to-head comparison is included because, given enough data, it could match us on accuracy. Our claim against it concerns robustness, tested below. DeepRETStroke works from 2D fundus photographs and cannot see the deep slab. We report it for orientation only, with its published silent-infarct AUC of 0.797, and do not treat it as a matched competitor.

The ablation that isolates the key ingredient is the same pipeline using single-band depth-resolved attenuation, the Liu-model approach of the intravascular A-line paper, computed from the full spectrum. Everything else is identical: segmentation, slabs, regressor and folds. In the artery stage, the reference numbers for this ablation are 83.47% two-way and 53.47% four-way accuracy, against 74.94% and 47.68% for raw intensity (Automatic Classification of A-Lines). These numbers come from their data, not ours, so we re-run their method on our raw-spectrum arterial data and compare on equal footing. A second ablation shuffles sub-band assignment, comparing two random halves of the spectrum with similar centers. This checks that any gain comes from spectral difference and not from noise averaging.

Why not just…?

Why not just compute single-band attenuation coefficients on retinal OCT? This is the most serious competitor. It already removes overall scaling, because the depth-resolved estimator is normalized by the signal integrated below each point. It also inherits the constant-ratio assumption, which fails wherever scatterer size changes, and capillary rarefaction and perivascular remodeling plausibly change exactly that. If the single-band ablation matches $\Delta\mu$ on both the arterial labels and the MRI endpoints, the two-band machinery is not worth its resolution and SNR cost, and we will report that and recommend the simpler method. The artery stage settles this before any retinal claim is made.

Why not just use OCTA density, add a scan-quality covariate, or fine-tune a foundation model? Density's reported correlations with small vessel disease are modest (r between −0.14 and −0.20). Association Between Dementia and Optical Coherence Tomography Scan Quality shows that scan quality is itself tied to the disease state. Adjusting for a quality index therefore risks removing real signal along with artifact, while leaving it in lets the model learn the artifact. A fine-tuned foundation model has no mechanism that separates the two. The cataract-surgery experiment below directly measures how much each method's output moves when only the media change.

Why not just use visible-light OCT oximetry, which is already differential absorption in the retina? Visible Light Optical Coherence Tomography does exploit the absorption difference between oxygenated and deoxygenated hemoglobin. It also faces shallow penetration, retinal light-safety limits and slower acquisition, and it has not been linked to small vessel disease. It targets oxygen in blood rather than tissue extinction around the deep vascular slab. Our method runs on standard near-infrared swept-source systems and could be added to scans clinics already take. Where visible-light hardware is available, it would complement our measurement rather than replace it.

Proposed Experiments and Decisive Checks

The phantom check passes if $\Delta\mu$ recovers bead-layer contrast with error near the 4% mean reported by Rohrbach et al. and changes by less than its own repeatability when an upstream diffuser cutting signal by several dB is inserted. Over the same insertion, single-band attenuation should show a systematic shift at the layer boundaries.

The artery check passes if $\Delta\mu$-augmented classification beats single-band attenuation on four-way plaque labels and on acute-injury labels in the raw-spectrum arterial data, under identical networks and splits.

The retinal check has two parts, and both must pass. The first is incremental explained variance over density and RETFound for periventricular and deep white matter hyperintensity, perivascular spaces and composite burden, together with discrimination of large artery from small artery stroke against healthy controls. The second is a natural experiment: patients scanned before and two to four weeks after cataract surgery. The retina is essentially unchanged over that window and only the optical media are cleared. The pre/post change in $\Delta\mu$ should be no larger than ordinary test-retest variation. Shifts in density, raw intensity and the RETFound score should be measurably larger.

If $\Delta\mu$ improves prediction but moves after surgery, it is measuring media rather than tissue and the claim fails. If it is stable but adds nothing over density, it is a clean measurement of the wrong quantity. Only the combination of stability and added prediction supports the core claim.

References

  1. A study on the aerosol extinction-to-backscatter ratio with combination of micro-pulse LIDAR and MODIS over Hong Kong
  2. Comparison of Scanning LiDAR with Other Remote Sensing Measurements and Transport Model Predictions for a Saharan Dust Case
  3. Raman LIDARs for the atmospheric calibration along the line-of-sight of the Cherenkov Telescope Array
  4. Automatic Classification of A-Lines in Intravascular OCT Images Using Deep Learning and Estimation of Attenuation Coefficients
  5. Separation of absorption and scattering profiles in spectroscopic optical coherence tomography using a least-squares algorithm
  6. Assessment of acute radial artery injury after distal transradial access for coronary intervention: an optical coherence tomography study
  7. Lower Retinal Arteriolar Density Is Associated With Higher Cerebral Small Vessel Disease Burden: An Optical Coherence Tomography Angiography Study
  8. Differences in retinal microvasculature between large artery atherosclerosis and small artery disease: an optical coherence tomography angiography study
  9. ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model
  10. OCTA-500: A retinal dataset for optical coherence tomography angiography study
  11. State-of-the-art retinal vessel segmentation with minimalistic models
  12. A deep learning system for detecting silent brain infarction and predicting stroke risk
  13. A foundation model for generalizable disease detection from retinal images
  14. Association Between Dementia and Optical Coherence Tomography Scan Quality
  15. Visible Light Optical Coherence Tomography: Technology and Biomedical Applications