Known-Answer Vessel Probes for Adapting OCTA Capillary Segmentation Across Scanners

Optical coherence tomography angiography (OCTA) photographs the retina's smallest blood vessels without dye. It works by detecting where the signal changes between repeated laser scans, because those changes come from moving blood cells. The capillaries it reveals are often only one or two pixels wide, and their density and branching are being studied as biomarkers for diabetic retinopathy, Alzheimer's disease and cardiovascular disease. Segmentation models trained on one scanner degrade on another, and relabeling capillaries for every new device is slow expert work. The usual way to use unlabeled scans from a new device is self-training, where a "teacher" model labels those scans and a "student" learns from its labels. This fails in a specific way on capillaries. On the new device the teacher misses faint vessels, calls them background, and the student learns to miss them too. Nothing in the unlabeled data reveals which "background" pixels are really missed vessels.

We propose inserting known answers into the unlabeled target scans. Procedurally generated capillary networks, whose labels are exact, are drawn into patches of real scans from the new device. They are rendered using that device's own background texture, noise and vessel intensity statistics, estimated from the scan itself. These probes do two jobs. They give the student exact supervision on thin vessels in the target device's appearance. They also measure how often the current model misses vessels of each width on that device, and that estimate is used to down-weight the teacher's "background" labels wherever missed capillaries are likely. The resulting method needs one small labeled source set, unlabeled single scans from each new device, and no image registration.

Core Claim

Under scanner shift, most of the capillary error that self-training accumulates comes from false-negative pseudo-labels on thin vessels. We claim that synthetic vessels embedded in target-device scans, rendered from that device's own statistics, can both correct this error and measure it. Our specific prediction is a measurable gain in thin-vessel recall and topology (fewer broken capillaries) on unseen devices and wider fields of view. The comparison is against the same teacher-student loop without probes, against probes rendered with generic noise, and against training purely on synthetic data.

Relation to Prior Work

Three existing lines each hold one piece of this. Synthetic OCTA angiographs for vessel segmentation without human annotations grows vessel trees with a fast space-colonization algorithm, a branching procedure that grows vessels toward randomly placed attractor points. It then makes the synthetic images look real through handcrafted noise, adversarial contrast perturbation or CycleGAN translation. We reuse its generator as our source of exact labels. That paper, however, trains only on synthetic images for a fixed target style and adapts nothing at test time. We instead insert synthetic vessels into real target pixels, inside a self-training loop.

MiDSS handles the case of labels from one site and unlabeled data from several shifted sites. It builds "intermediate domains" by copy-pasting regions between labeled and unlabeled images and by mixing low-frequency Fourier amplitudes, which carry global brightness and style. Its copy-paste cuts through vessels, leaving severed trees with mismatched labels. It was also only tested on solid structures such as the optic cup, prostate and heart ventricles. Our probes are complete, self-consistent vessel graphs, so they carry no seam-induced label error.

GrInAdapt adapts an OCTA-500 model to AI-READI devices by registering multiple views of the same eye and fusing them into pseudo-labels. It reports registration failure in about a quarter of subjects and cannot work on a standalone scan. We impose neither requirement.

Two further papers shape the design as motivation. D2Net separates vessel content from instrument noise with two encoders. It is fully supervised, and its authors report false-positive "overshooting" on wider fields of view. ErrorNet shows on fundus photographs that explicitly modeling segmentation breaks improves cross-dataset Dice. Together these tell us that noise streaks and broken vessels are the failure modes to target.

We take one method component directly. clDice supplies a differentiable centerline loss, reported to add under 10% training time on 1024×1024 images. We apply it on probe regions, where the centerline is exactly known, rather than on noisy pseudo-labels.

One benchmark constrains our choices. Benchmarking retinal vessel models for cross-dataset generalization finds contrast degradation hurts more than disease shift. This is why the probes are rendered with each scan's own contrast, not a generic noise model.

Research Program

Probe renderer. For each unlabeled target scan, the current teacher first separates confident vessel pixels from confident background. From the confident thin vessels we collect intensity distributions stratified by estimated radius, using a distance transform on the skeleton. From the background we extract the texture residual left after smoothing: speckle, low-flow haze and horizontal motion streaks. A synthetic capillary mesh is then drawn into a chosen patch. Its intensities are sampled from the target's radius-matched vessel distribution, and it is modulated multiplicatively by the patch's own residual, so the target's noise passes over the probe.

Choosing where probes go. Probes are placed in patches of the target scan, and the local vessel density of the synthetic tree is matched to the teacher's estimate for that region. The foveal avascular zone, a small central area normally free of vessels, is excluded. Its location is found with the source-trained foveal segmentation task from OCTA-500.

Guarding against a shortcut. The obvious risk is that the student learns to spot inserted vessels rather than vessels. We monitor this with a separate small classifier trained to tell probe patches from untouched ones. Its AUC is reported, and the renderer is changed if the AUC exceeds a preset level such as 0.7.

Probe-calibrated pseudo-labels. During training, the student's predictions on probes give a running estimate of recall by device and vessel radius, $r_d(\rho)$: the fraction of true probe vessels of radius $\rho$ that the teacher detects on device $d$. Teacher-labeled background pixels are then weighted by

$$w_d(x) = 1 - (1 - r_d(\rho^\ast))\, m(x),$$

where $\rho^\ast$ is the thinnest radius class, and $m(x) \in [0,1]$ is a local vesselness score from an untrained Frangi filter, meaning how tube-like the neighborhood of pixel $x$ looks. Where probes show the teacher misses thin vessels on this device, tube-like "background" contributes little loss, so the student is not taught to erase possible capillaries. Probe regions themselves are supervised with Dice plus clDice against their exact labels. The rest of the image uses the weighted teacher labels and ordinary teacher-student consistency.

Baselines

The naive reference is a U-Net trained on the small labeled source set with standard photometric augmentation, applied to each target unchanged. It anchors how large the shift is. The same network trained on all available target labels gives a supervised upper bound.

The strongest annotation-free competitor is synthetic-only training with CycleGAN translation, S+GAN. It reports Dice 0.842±0.003 on the Giarratano dataset against 0.907±0.002 for a model trained on target labels. We run it with the published generator, one CycleGAN per target device, and the same backbone, iteration budget and hyperparameter search as our method. Our method, which uses a few source labels plus unlabeled target scans, should close a meaningful share of that 0.065 gap. A gain under 0.01 Dice with no gain in thin-vessel recall would mean the probes add nothing.

The strongest mixed-domain semi-supervised competitor is MiDSS. It reports 87.66% Dice on Fundus against an 88.46% upper bound, but has no reported result on Lacuna for OCTA. We rerun its public recipe on our splits with the identical backbone, and also give it the clDice loss so that loss is not our only advantage.

A plain teacher-student baseline, an exponential-moving-average Mean Teacher with clDice and Fourier amplitude randomization, isolates the probes' contribution. This is effectively GrInAdapt's adaptation stage without its registration-based fusion. The GrInAdapt paper itself reports about 4% average Dice gain on AI-READI, which we cite for scale but cannot reproduce fairly on single scans.

For supervised context on multi-device data, D2Net reports 63.48% Dice on FOCA and 83.83% AUC on unseen ROSE-H. It uses more labels than our setting, so it serves as context, not a head-to-head comparison.

Two ablations isolate the key ingredient. The first renders probes with S+GAN's handcrafted noise model instead of target-derived statistics. The second keeps the target-derived probes but drops the recall-based weighting $w_d$. Every method gets the same labeled images, unlabeled pool and GPU-hour cap, and three seeds.

Why not just…?

Why not just train on synthetic data translated by a GAN, as S+GAN does? This is the real threat. Our experiments settle it directly: S+GAN and our method are run per target device under matched budgets. Our expectation of an advantage rests on two points. The GAN must learn the device's appearance from unpaired data, and GAN translation can hallucinate or erase thin vessels. Our probes, by contrast, inherit real target pixels. If S+GAN matches us on capillary recall, the self-training half of the method is unnecessary, and we would report that.

Why not just run MiDSS with a topology loss? Its copy-paste cuts vessels at patch seams. On capillary networks this creates label errors exactly on the one-to-two pixel structures we care about. Running MiDSS with clDice as a baseline tests whether that matters in practice. We additionally measure Betti-number error (counts of broken pieces and spurious loops) near seam locations.

Why not just label five target images per device? This is cheap and honest. We include a few-shot supervised fine-tuning curve, 1, 3 and 5 labeled target images, and report how many labels our unlabeled method is worth. OCTA-500 reports steep gains up to 120 training samples, which suggests five labels leave substantial headroom. That is an expectation we test, not assume.

Why not just use a Frangi filter for the unlabeled targets? The S+GAN paper reports that Frangi and optimal-oriented-flux filters fall well short of learned models on noisy OCTA. We use Frangi only as a soft prior inside $w_d$ and report it standalone for completeness.

Proposed Experiments

Protocol. The labeled source is 20 OCTA-500 3×3 mm scans with capillary labels. The unlabeled targets are OCTA-500 6×6 mm scans, which test field-of-view shift, plus ROSE-1, ROSE-2, FOCA and Giarratano, with their labels held out. AI-READI Cirrus, Maestro2 and Triton scans enter the unlabeled pool. They are evaluated only if reference maps are available to us.

Metrics. We report Dice, clDice, Betti-0 and Betti-1 errors, and recall stratified by vessel radius. For ROSE, whose capillaries are annotated only as centerlines, we report centerline recall and precision within a one-pixel tolerance. This also reduces sensitivity to the annotation-convention mismatch that led the S+GAN authors to radius-filter their labels. For ROSE-1 we additionally report AUC, to compare against OCTA-Net's 0.9453 on the superficial vascular complex from ROSE.

Decisive checks. The first check validates the recall estimate. On targets that have labels, the recall-by-radius curve estimated from probes should rank-correlate with the true curve measured against held-out labels across devices and training checkpoints. If it does not, the calibration mechanism is unfounded, whatever the final Dice. The second check tests the headline claim. The gain over the matched Mean Teacher should concentrate in the thinnest radius class and in Betti-0 error, most strongly on the 6×6 mm and FOCA shifts. A uniform gain across widths would point to generic augmentation, not to fixing false-negative pseudo-labels. The third check tests the shortcut risk. The probe-detector AUC should stay low, and target-derived probes should beat handcrafted-noise probes. If handcrafted probes do as well, the target-statistics rendering is not the reason the method works. The fourth check is clinical. Vessel density and fractal dimension computed from our segmentations should agree more closely with those from held-out expert labels than the baselines' do, because these are the quantities downstream clinical studies actually use.