Borrowing Strength from the Sclera: Mechanism-Shared Models for Rare Pachychoroid Diseases

Several rare eye diseases, including central serous chorioretinopathy, pachychoroid neovasculopathy and uveal effusion syndrome, appear to share one physical story. A thick, unyielding sclera (the eye's outer collagen wall) narrows the tunnels through which the choroid's large draining veins (the vortex veins) exit the eye. Blood backs up, and the choroid (the vascular layer just inside the sclera) swells. Machine learning for these diseases runs into a familiar wall of its own. Published cohorts contain 9 neovasculopathy eyes or 12 effusion patients. A classifier trained on such data either predicts "healthy" for everyone or memorizes the handful of positives. This note proposes to treat the shared mechanism as the fix for class imbalance. The idea is to learn the scleral-thickness-to-choroidal-swelling relationship from the many ordinary eyes that exhibit it. Each rare disease is then modeled as a small, class-specific shift on that shared curve, so the rare classes need very few parameters of their own.

The specific bet rests on Keidel and colleagues' clinical observation that scleral and choroidal thickness are both elevated in disease but not linearly correlated, which they read as a threshold effect. If that reading is right, a hinge-shaped coupling curve should transfer from abundant eyes to rare ones. Diseases would then be separable from controls with far fewer labeled positives than a generic imbalance-corrected network needs. The claim is falsifiable on two counts: the hinge can be absent, and the rare eyes can fail to sit beyond it.

What exists and what is missing

The mechanical evidence comes from two kinds of study. In porcine eyes, the stiffness of the sclera ringing the cornea predicts eye pressure strongly: Perilimbal sclera mechanical properties reports $R^2=0.74$. Chemically stiffening that tissue alone cuts outflow facility from 1.40 to 0.51 µL/min/mmHg in Effects of Genipin Crosslinking. Together these studies motivate the idea that an enclosing matrix sets vessel outflow. They also constrain this proposal, because they show stiffness can change with no change in geometry, and OCT measures only geometry.

The human studies are small and rely on manual measurement. Scleral thickness as a risk factor for CSC and PNV measures anterior scleral thickness with a single caliper at one spot (403.5 vs 362.5 µm in 23 vs 23 eyes) and finds the non-linear relation described above. Pachychoroid disease and RVO adds 312 eyes with choroidal thickness and pachychoroid labels. Pentraxin-3 and Lipocalin-2 in UES adds 12 effusion patients with very thick choroids. These three papers motivate the proposal: they supply the hypothesized curve and show how scarce the positives are.

On the machine learning side, imbalance is attacked either through the loss or through the data. Unified Focal loss and Balance loss address pixel imbalance. The second paper shows that the many tiny background probabilities summed during training create a false-positive bias that disappears once outputs are thresholded at test time. Red-GAN and GIN address sample imbalance by synthesizing minority cases. GIN goes further and ties latent axes to physical variables. All four learn class structure from the data in hand. None imports a physical relation fitted on other populations, and the generative pair were run on cohorts of hundreds rather than a dozen.

Two choroid tools supply the segmentation starting point. Choroidalyzer segments choroid and vessels, and Bio-Net regularizes choroid segmentation with a thickness prior. Neither segments the sclera, and Choroidalyzer degrades near the optic disc. The gap is therefore twofold. No automated tool measures the containing wall and the contained vessels together. No rare-disease model shares a mechanism across diseases.

Core Claim

Extract a few mechanically meaningful quantities per eye: scleral thickness mapped at many points, choroidal thickness, and the choroid's large-vessel lumen fraction. Fit their coupling as a hinge on all eyes. The resulting classifier for rare pachychoroid diseases should then beat end-to-end imbalance-corrected networks and synthesis-based augmentation when positives number under about ten per class. It should also degrade more slowly as positives are removed.

Concretely, let $T$ be an eye's scleral thickness (a robust summary of the automated map), $E$ its choroidal engorgement (thickness or large-vessel lumen fraction, adjusted for age and refraction), and $\tau$ a breakpoint. The coupling model is

$$E = a + b\,\max(0,\,T-\tau) + \varepsilon,$$

fitted on every available eye, the majority of which are controls and RVO fellow eyes. For disease class $k$, the classifier is

$$\text{logit}\,P(y=k) = \alpha_k + \beta\,\max(0,\,T-\tau) + \gamma\,\hat\varepsilon,$$

where $\hat\varepsilon$ is the residual engorgement unexplained by the hinge. The slope terms $\beta$ and $\gamma$ are shared across CSC, PNV and UES, and only the intercept $\alpha_k$ is class-specific, with intercepts partially pooled in a hierarchical prior. Each rare class therefore spends its few labels on one or two numbers.

Research Program

The first component is a joint wall-and-vessel segmenter. It extends Choroidalyzer's multi-task network with a scleral boundary output on anterior-segment swept-source OCT, along with the choroid region and vessels on enhanced-depth posterior OCT. The scleral inner and outer boundaries and vessel lumina are thin structures that occupy a tiny fraction of each image. The network is therefore trained with Balance loss and compared against Dice plus cross-entropy. Scleral thickness is read at many positions around each rectus insertion rather than one caliper point, which turns a single number into a map.

The second component fits the hinge on the pooled non-rare eyes. Following the published cohorts, these would be the 150 RVO-study controls, the 162 fellow eyes and the 23 matched controls. We test whether a breakpoint model beats a linear one by held-out likelihood. This step is analysis, not engineering, and it is where the mechanism either holds or fails.

The third component is the shared-slope rare-class classifier described above. Uncertainty is propagated from the segmentation into $T$ and $E$ by resampling the soft segmentation outputs.

One dependency must be stated plainly. The rare-disease cohorts in the file are the LMU Munich clinical cohorts, not public benchmarks, so the classification arm requires a data-sharing collaboration or a matched new cohort. Public data covers only the segmentation arm and the loss-reproduction check.

Baselines

The naive reference is subfoveal choroidal thickness alone, thresholded at the value that maximizes Youden's index on training folds. This is what a clinician would do. If the full model cannot beat it, the mechanism adds nothing.

The strongest mechanistic competitor is a two-feature logistic regression on manually measured single-point scleral thickness and choroidal thickness, which is the Keidel et al. measurement protocol turned into a predictor. It sees the same eyes and uses the same cross-validation folds. Its features are the clinicians' caliper values. No reported classification result found on Lacuna for pachychoroid disease prediction from these features; the papers report only group differences.

The strongest machine learning competitor is an end-to-end CNN classifier on the same scans. It is trained with the asymmetric Unified Focal loss and class-balanced sampling, with tuning for Unified Focal loss's weighting and focusing parameters and the learning rate under the same hyperparameter search budget and GPU hours as our full pipeline. For pixel-level imbalance on public data, Balance loss reports 66.8 Dice on KiTS19 kidney tumor versus 59.8 for cross-entropy (Balance loss). We first reproduce that number to within a few points before trusting our implementation.

For the segmentation arm, Choroidalyzer itself is the baseline. It reports choroid vessel Dice of 0.8817 internal and 0.8703 external and region Dice of 0.9789 and 0.9749 (Choroidalyzer). Our extended model must not fall below these on its own tasks while adding the sclera.

Three ablations isolate the key ingredient. The first replaces the hinge with a linear term. The second fits separate slopes per disease instead of shared ones. The third substitutes generic pretrained image embeddings of equal dimension for the mechanical features.

Beating the field means a clear margin over the manual two-feature regression and the focal-loss CNN in sensitivity at 90% specificity and in area under the precision-recall curve under leave-one-patient-out evaluation. The margin must be visible with confidence intervals bootstrapped over patients, and it must widen as positives are subsampled.

Why not just…?

Why not just synthesize virtual pachychoroid patients with a conditional or invertible GAN? GIN and Red-GAN showed gains when hundreds of patients anchored the generator: 168 TAVR patients for GIN, and BraTS and ISIC for Red-GAN. With 9 to 12 positives, a generator can only interpolate between memorized eyes. Its synthetic cases add no information about the class boundary that the shared mechanism does not already encode more cheaply. We include Red-GAN augmentation of the focal-loss CNN as an additional arm. If it matches the hinge model at 9 positives, synthesis is competitive and we will say so.

Why not just train an end-to-end CNN with an imbalance loss? Imbalance losses fix how gradients are weighted, but the model still must learn what makes a sclera "too thick" from the few positive eyes. The subsampling experiment tests this directly. If the CNN's precision-recall area holds up as positives drop from 9 to 3, the mechanism prior is unnecessary.

Why not just run logistic regression on the two caliper measurements? This is the real threat. With so few positives, a two-parameter model may be near-optimal, and our gain would then come only from automation. We settle it with a crossover design: the hinge model on manual features, and the linear model on automated multi-point features. This separates the value of the threshold structure from the value of the scleral map. If neither beats the manual baseline, the contribution reduces to a faster measurement tool. The note would report that rather than claim a classification advance.

Decisive Checks

The first check is whether a hinge beats a linear fit on held-out non-rare eyes. The second is whether rare-disease eyes, which play no part in fitting the curve, fall disproportionately beyond the breakpoint $\tau$. If either fails, the threshold reading of Keidel et al. is wrong for this population and the shared-mechanism classifier should not outperform the baselines.

A third check addresses the known blind spot. Genipin-style stiffening raises outflow resistance without changing thickness (Riesterer et al.), so we expect a subset of diseased eyes with normal scleral thickness that our model misses. We will report its size rather than hide it. The transorbital work of He et al. shows the likely remedy. There, adding Doppler resistance to sheath diameter raised the area under the curve from 0.81 to 0.95, which suggests that a hemodynamic measurement, not more geometry, is the next feature for those missed eyes.