Source-linked AI summary

In Defense of OCTA: The Reconstruction-Utility Gap in OCT-to-OCTA Synthesis

Michael Chertok, Alon Tiosano, Orly Gal-Or, Lior Kramarski, Einav Baharav Shlezinger, Irit Bahar, Lior Wolf

arXiv:2608.15626v1cs.LG

TL;DR

OCT-to-OCTA synthesis is typically judged by reconstruction similarity, but the key question is whether synthetic images preserve OCTA’s vascular measurements. Using a frozen real-OCTA segmenter to probe XOCT and TransPro, the paper finds fine-capillary collapse, fabricated detail, and absent neovascular lesions despite strong reconstruction fidelity.

  • Problem

    OCT-to-OCTA synthesis lacks evidence that reconstruction similarity preserves the downstream vascular measurements OCTA is acquired to provide.

  • Method

    A frozen real-OCTA segmentation network probes XOCT and TransPro, stratifying downstream performance by vascular structure and using matched-blur and train-on-synthetic controls.

  • Results

    Both synthesizers collapse fine capillaries; XOCT falls from 0.798 → 0.635 Dice, five times the large-vessel loss, while neither reproduces neovascular lesions.

  • Takeaways & Limitations

    Reconstruction fidelity is not clinical utility, so OCTA-to-OCTA models should be judged by whether they support the vascular measurements OCTA is acquired to make.

  • Takeaways & Limitations

    The capillary result uses one OCTA engine, and Dice is only a surrogate for clinical reading.

Abstract

from arXiv · show

Optical coherence tomography angiography (OCTA) images retinal blood flow, giving capillary-perfusion and foveal-avascular-zone biomarkers that grade diabetic-retinopathy ischemia. Because OCTA hardware is less common than structural OCT, recent work synthesizes it from OCT, reporting strong reconstruction (3D PSNR > 31 dB, SSIM > 0.9). We ask not whether the synthetic image looks similar, but whether it supports the measurements OCTA is acquired for. A frozen real-OCTA segmenter, applied as a probe to two synthesizers (XOCT, TransPro), shows downstream Dice falling with structural fineness: large vessels survive (0.862 -> 0.831) while the fine capillary network collapses (0.798 -> 0.635, five times the large-vessel loss; paired Wilcoxon p < 1e-3), TransPro worse throughout. A matched-blur control shows this detail is fabricated, not blurred. Retrained on a private Spectralis dataset, neither synthesizer reproduces the neovascular lesion (qualitative, n=3). Reconstruction fidelity is not clinical utility; we establish downstream-task fidelity as the evaluation OCT-to-OCTA synthesis needs.

1 Introduction

OCT-to-OCTA synthesis can achieve strong reconstruction fidelity while failing to preserve the fine capillary detail needed for downstream vascular measurements. This section introduces a frozen real-OCTA segmentation probe showing that reconstruction quality does not establish clinical utility.

  • Motivation and problem: PSNR above 31 dB in 3D and SSIM above 0.9 indicate increasing reconstruction fidelity, but do not show that synthetic OCTA supports downstream vascular measurements.OCTA provides depth-resolved vascular maps and FAZ biomarkers, while structural OCT is more widely available, motivating synthesis from OCT alone.
  • Evaluation approach: A frozen real-OCTA segmentation network probes utility by measuring structure-stratified quality on synthetic OCTA from XOCT and TransPro.The probe is training-free and reusable for scoring synthesizers.
  • Results: 0.798 → 0.635 Dice for XOCT shows collapse of the fine capillary network, with five times the large-vessel loss; TransPro degrades every target more.Both synthesizers collapse finer structure despite matched reconstruction fidelity, while XOCT preserves large vessels better.
  • Controls: A matched-blur control shows that missing capillaries are fabricated rather than blurred.The control localizes the failure to synthesis content rather than simple loss of image sharpness.
  • Controls: A train-on-synthetic control shows that vascular information is absent, not merely affected by domain shift.This control supports interpreting the downstream segmentation failure as missing vascular information.

2 Related Work

OCT-to-OCTA synthesis has advanced from early flow-map generators and 2D GANs to 3D, layer-aware models, but evaluations still emphasize pixel fidelity and on-image proxies. Prior medical-imaging work supports downstream task performance as the discriminating measure, motivating a frozen real-OCTA segmenter as the measuring instrument.

  • OCT-to-OCTA synthesis: OCT-to-OCTA synthesis progressed from deep-learning flow-map generators and 2D conditional GANs to 3D volumetric models with layer-aware supervision.TransPro adds frozen vessel-segmentation guidance and projection consistency, while XOCT uses a 3D multi-scale generator with cross-dimensional supervision.
  • OCT-to-OCTA synthesis: Existing methods report synthesizer outcomes mainly through pixel fidelity, including MAE, PSNR, and SSIM, rather than downstream measurements.Vascular-aware evaluations use on-image proxies such as vessel-density error, biomarker trends, and density, caliber, and tortuosity indices.
  • Task-based evaluation of synthesis: Segmentation-based evaluation can distinguish synthesizers that PSNR and SSIM rank as equivalent, supporting downstream task performance as the discriminating measure.Task-based protocols anchor cross-modal benchmarks such as ultrasound-to-MR and are advocated across generative medical-imaging research.
  • OCTA vessel segmentation: A frozen segmenter reaching Dice above 0.85 for vessel and FAZ segmentation on real OCTA serves as the fixed measuring instrument rather than a contribution.This establishes the evaluation instrument for testing whether synthesized OCTA supports clinically relevant segmentation measurements.

3 The Downstream-Task-Fidelity Protocol

The protocol evaluates whether OCTA’s clinical measurements survive synthesis, using two frozen networks rather than training a new model. It standardizes projections and quantifies target-specific utility gaps between real and synthetic OCTA.

  • Protocol contribution: The contribution is a task-based measurement protocol, not a new synthesizer, and nothing is trained inside it.It tests whether the clinical measurement OCTA is acquired for survives synthesis using two frozen networks.
  • Protocol ingredients: The protocol combines a real-OCTA segmenter S, an OCT-to-OCTA synthesizer G, and held-out real OCTA volumes with labels, while S and G remain frozen.G maps a structural OCT volume o to synthetic OCTA volume ˆv = G(o).
  • Projection standardization: Both real and synthetic volumes pass through the identical projection operator P before segmentation, making the comparison a real-versus-synthetic gap test rather than a projection mismatch.For each band, P uses maximum-intensity projection over the band’s depth range, followed by fixed orientation and percentile normalization.
  • Utility gap: ∆t is the target-specific drop in Dice when synthetic rather than real OCTA is input to the same segmenter and projection pipeline.Targets include large vessels, FAZ, and capillaries; ∆t > 0 means synthesis cannot support the measurement as well as the real image.
  • Causal controls: Two fixed controls distinguish causal failure mechanisms: matched blur tests lost resolution, while decoder retraining on synthetic OCTA tests real-to-synthetic domain shift.The matched-blur control uses the Gaussian width σ⋆ whose radial power spectrum most closely matches the synthetic projection.

4 Experimental Setup

The experiments use paired OCT/OCTA data with a matched 50-eye test split, two legitimately trained synthesizers, and a frozen OCTA segmenter as the downstream probe. A separate Spectralis dataset provides cross-vendor, specialist-graded evaluation data.

  • Data: 300 eyes from OCTA-500 provide paired OCT/OCTA volumes, layer segmentations, and pixel-level vessel/FAZ labels.The dataset uses 6 mm, 400 × 400 en-face acquisitions.
  • Data: 50 test eyes form the intersection of the synthesizer’s 70-eye held-out split and the segmenter’s held-out set.The 20 test eyes included in the segmenter’s training are excluded from downstream evaluation.
  • Data: Heidelberg Spectralis volumes provide a separate 3–4 mm macular-field dataset with per-eye diagnoses graded by a retina specialist.This scanner and field differ from OCTA-500’s 6 mm acquisitions.
  • Synthesizers: XOCT and TransPro are trained state-of-the-art comparators with XOCT reproduction within 0.1 dB PSNR and 0.003 SSIM.TransPro is trained from scratch on the same split to XOCT-comparable reconstruction; MuTri is omitted because trained weights were unavailable.
  • Segmenter: The frozen OCTCube probe segments real OCTA vessels and FAZ at Dice 0.873/0.866/0.833 for large vessel/FAZ/capillary.Its encoder was pretrained on 26,685 structural-OCT volumes, and the setup reproduces the large-vessel test Dice of 0.873.

5 Synthesis Collapses the Fine Capillary Network

Synthetic OCTA preserves coarse vascular structure but collapses the fine capillary network, and matched-blur controls show that this deficit reflects structural fabrication rather than lost resolution. Retraining on synthetic images does not recover the capillary signal.

  • Downstream utility: 0.798 → 0.635: capillary Dice collapses, a 0.163 absolute drop and five times the large-vessel loss, while large-vessel Dice falls only from 0.862 → 0.831.Every target gap is significant across 50 held-out eyes (paired Wilcoxon p < 10^-3).
  • Downstream utility: 0.333 vs 0.329 mm2: XOCT keeps FAZ area within measurement repeatability, whereas TransPro inflates it to 0.426 mm2, approximately 30% higher.The coarse FAZ survives, while loss is specific to the fine capillary network.
  • Matched-blur control: 0.635 and 0.591: XOCT and TransPro capillary Dice remain 20–26% below real OCTA at matched σ = 0.4 px resolution.Real OCTA must be blurred to σ≈1.0–1.2 px, or 2.5–3× the synthetic blur, to segment as poorly; high-frequency energy is present but misplaced.
  • Adaptation test: 0.617: a decoder trained and tested on synthetic OCTA still fails to recover the capillary network, below 0.635 from the real-trained segmenter on synthetic images and 0.798 on real OCTA.This adaptation removes real-to-synthetic distribution mismatch, supporting genuine information loss in the synthesized vascular signal.

6 Discussion and Future Work

The findings are limited to one OCTA engine and a surrogate clinical metric, so synthetic OCTA cannot replace acquisition when capillary-level detail matters. In a SpectralisMNV retraining test, neither synthesizer reproduced the neovascular lesion, underscoring the need to evaluate pathological flow networks directly.

  • Limitations: Synthetic OCTA is not a substitute for acquisition when clinical interpretation depends on capillary-level detail.The capillary result used Optovue SSADA at 6 mm, and Dice served as a surrogate for clinical reading.
  • Neovascular AMD: On full-retina images, both synthesizers reproduced arcade vessels but left a dark macular void, unlike the discrete branching neovascular lesion in real OPL–BM images.This qualitative comparison used XOCT and TransPro retrained on the same SpectralisMNV data.
  • Neovascular AMD: Neither synthesizer reproduced the neovascular lesion after retraining on SpectralisMNV data.The test used 478 paired Heidelberg Spectralis volumes at 3–4 mm and held out eyes graded as showing definite MNV.
  • Neovascular AMD: MNV is pathological flow in the normally avascular outer-retinal OPL–BM slab, which OCTA detects, types, and monitors.The OPL–BM slab is the slab read for macular neovascularization, whereas exudative fluid on structural OCT triggers treatment.

7 Conclusion

Both state-of-the-art synthesizers collapse the fine capillary network, a structural failure that reconstruction metrics do not reveal. Clinical utility—not reconstruction fidelity—should be judged by whether synthesized OCTA preserves the vasculature OCTA is acquired to measure.

  • Both state-of-the-art synthesizers collapse the fine capillary network, exposing a structural failure invisible to reconstruction metrics.
  • Clinical utility, rather than reconstruction fidelity, should be the yardstick for models synthesizing OCTA.The relevant test is whether the synthesized images preserve the vasculature OCTA is acquired to measure.
Loading 2608.15626v1…